Small form-factor PCs and gaming systems have emerged as bright spots in the mature PC market over the last decade or so. Intel's NUC form-factor introduction was the turning point in the small form-factor (SFF) market. Over several generation of products, the NUC family has expanded to address different market segments. But even with those developments, DIY enthusiasts have traditionally given the NUCs a cold shoulder by pointing to the lack of flexibility and limitations such as the inability to use a discrete GPU. So to to address the requirements of the performance enthusiasts, Intel introduced the Skull Canyon NUC in 2016 and followed it up with 2018's Hades Canyon NUC.

Following that same strategy of developing enthusiast-tier NUCs, at CES 2020, Intel officially announced the Ghost Canyon NUC series of products (NUC 9 Extreme Kits), which are based on the Intel NUC 9 Extreme Compute Elements. Among other notable changes, the series adds the ability for end-users to add a standard PCIe video card to the system system. The products have been made available to OEMs and ecosystem-enablers since late 2019, and they are finally reaching retail availability this month.

With their latest NUC finally available at retail, Intel has sent us an engineering sample of the top-end Ghost Canyon NUC, the NUC9i9QNX. Intel has equipped the NUC with matching build components to create a premium SFF gaming / workstation system. Read on for a review of the NUC9i9QNX and an analysis of the Intel Compute Element and its associated ecosystem.

Introduction and Product Impressions

The Ghost Canyon NUC9i9QNX is Intel's current top-end NUC with a NUC 9 Extreme Compute Element (NUC9i9QNB) housed in a 238mm x 216mm x 96mm chassis. NUCs have traditionally been associated with the ultra-compact form factor (100mm x 100mm boards in a 0.63-litre or 0.42-litre volume chassis). The Skull Canyon and Hades Canyon products with their higher TDP processors had to be accommodated in 0.69L and 1.2L chassis, but the Ghost Canyon NUC kits with their support for PCIe video cards takes it to a whole new level with a 4.94L chassis.

Though make no mistake: while biggest than the smallest NUCs, this is still well within the realm of SFF PCs. In fact, the smallest SFF PC with support for PCIe video cards that we happen to be aware of is the ZOTAC ZBOX MAGNUS series, with a 5.8L chassis. The Ghost Canyon NUC chassis includes a built-in power supply while the ZOTAC mini-PC uses an external adapter, which if counted would add further to its volume. As a result, the Ghost Canyon NUCs take the title of the smallest ever PC lineup to support user-replaceable discrete GPUs by a comfortable margin.

Intel's NUC lineup has traditionally included board and kit variants, allowing its partners to provide value additions (such as a passive chassis or additional I/O ports in the end system). Kits (other than the ones that come with a pre-installed OS) require the end-user to add storage, DRAM, and install an OS to complete the system. Some kits support a 2.5" drive in their chassis configuration, while others are M.2-only from a storage perspective. The Ghost Canyon series of products also follows a similar philosophy, while greatly increasing the flexibility for the end-user. Towards that, a Ghost Canyon is system comprised of multiple parts, which can be picked and chosen by OEMs / end-users to create a system for specific use-cases:

• The Compute Element
• Baseboard (or backplane)
• Chassis
• PSU
• DRAM (up to 2x DDR4-2666 SODIMMs)
• Non-volatile storage (HDD or SSD)
• Discrete GPU (optional)

An off-the-shelf Ghost Canyon NUC9i9QNX kit leaves only the DRAM, non-volatile storage, and discrete GPU to the choice of the end-user. Prior to the platform analysis and overview of our review configuration, let's take a look at the pre-decided components in the above list.

The Compute Element

The Compute Element is the board component used in the previous NUC generations. The NUC9i9QNX we are reviewing today comes with the NUC9i9QNB (NUC 9 Extreme Compute Element). The Compute Element comes with a soldered processor - the Core i9-9980HK. This belongs to the Coffee Lake Refresh-H family, and has a 8C/16T configuration with a 45W TDP. It can turbo up to 5 GHz. And, as we've previously covered in our look at  Intel's Compute Element prototype, the NUC 9 Compute Element re-imagines the traditional motherboard in a discrete PCIe x16 card form-factor.

The Intel NUC 9 Extreme Compute Element (NUC9i9QNB)

The Compute Element comes with a cooling shroud containing a single fan and two M.2 heat-sinks with thermal pads pre-attached. These align with the two M.2 slots (1x M.2 2280, and 1x M.2 22110) in the Compute Element to the left of the processor. On the right, we have the two SODIMM slots that can operate at speeds of up to DDR-2666 for DIMMs up to 16GB, while 32GB DIMMs are rated for up to DDR-2400. The gallery below provides additional photographs of the Compute Element and the cooling shroud.

The NUC9i9QNB comes with two Thunderbolt 3 ports, two gigabit Ethernet ports, a HDMI 2.0a display output, and four USB 3.1 Gen 2 Type-A ports in the rear. Headers include front panel audio and I/O connectors, a CEC connection, a SATA connector (FPC style with power), 2x USB 2.0 headers, and 2x USB 3.1 Gen 2 Type-C headers internally. The Compute Element has its own power connection to the PSU.

The Baseboard

Intel's backplane/baseboard design for the NUC 9 Compute Elements is comprised of two PCIe x16 slots and a sole PCIe x4 slot, as well as a PCIe x4-backed M.2 22110 slot. One of the x16 slots is dedicated to the Compute Element. The other x16 slot operates in x16 mode as long as the PCIe x4 slot and the M.2 slot remain unoccupied. If either slot is occupied, then x16 slot operates in x8 mode. Essentially, the x16 attachment to the Compute Element can electrically operate as x16 or (x8 + x4 + x4). Further down in the platform analysis section, we will see that these 16 lanes come directly from the processor (and not the CM246 PCH). The baseboard has an explicit power connection to the PSU for supplying power to the PCIe slots.

Baseboard for the NUC 9 Extreme Compute Element in the NUC9i9QNX (Photo Credit: KoolShare.cn Forums)

The baseboard photograph reveals two ASMedia ASM1480 mux/de-mux chips to enable the splitting of the PCIe lanes. Note that these are true muxes and not PCIe switches, so physical lanes are being reallocated when additional cards are connected, and the GPU is hard limited to x8 bandwidth (~8GB/sec) in that case. And while unlikely, thanks to Intel's bifurcation rules, it should be noted that anything that tries to split the x16 slot (e.g. a M.2 riser with multiple M.2 slots) would disable the PCIe x4 slot as well (though the M.2 slot remains usable).

On that note, for the purposes of our review sample, Intel equipped the system with a 380GB Optane SSD 905p, connected to the M.2 slot in the baseboard, as well as an ASUS Dual GeForce RTX 2070 MINI 8GB GDDR6 card in the x16 slot. Due to the baseboard architecture, the card operates in x8 mode in the supplied review configuration.

Chassis

The NUC9i9QNX chassis has been designed with the intent of retaining the ease of installation that the NUCs are famous for. The dimensions have also been kept as small as possible while providing the ability to install a dual-slot PCIe card (up to 202mm in length and 131mm in height) in the baseboard along with the Compute Element.

The top panel of the chassis slides off after the two screws in the rear are loosened. This panel also includes two 80mm x 15mm fans below the perforated top portion. The 12V supply for the fans are delivered through a snap-on connector, as shown in the above gallery. The perforated side panels can be removed by sliding up after the removal of the top panel. On the side opposite to the Compute Element slot, the top bar on the chassis can be taken out (to aid in the installation of a PCIe card) by removing two screws.

The I/O ports in the rear panel are directly off the Compute Element. However, the front I/O ports require the routing of on-board headers in an appropriate manner. The front panel of the NUC9i9QNX has a power button and an audio jack (connected to the front panel audio header on the board), a SDXC slot, and two USB 3.1 Gen 2 Type-A ports. The last three are enabled by a daughterboard containing a Genesys Logic GL3227 USB 3.1 Gen 2 hub chip and a Genesys Logic GL3590 SDXC to USB bridge chip. The hub chip in the daughterboard is perhaps unavoidable due to the inability of the headers to be individually routed to distinct ports in the front panel (given the volume constraints of the small chassis).

Overall, despite the plastic framing in the front panel, the metal-based construction in other areas gives the chassis a premium look and sturdy feel. Space management in the absence of a discrete GPU or PCIe cards is simple, but things get really cramped when the chassis is stuffed to the gills (as in the case of our review configuration).

PSU

An attractive aspect of the NUC9i9QNX is the built-in 500W 80 PLUS Platinum power supply. Most SFF PCs (including similar powerful ones like the ZOTAC ZBOX MAGNUS series) use bulky external adapters, so the ability to use just a simple power cord with the system is a welcome one from the perspective of a compact installation.

The 500W FSP PSU in the NUC9i9QNX

Intel has used the FSP500-30AS, a new PSU from FSP in a Flex-ATX form-factor. The cables enable up to 150W to be supplied to the add-in card directly. Along with the 75W budget from the PCIe connection, the NUC9i9QNX supports cards that consume up to 225W.

Our review sample of the NUC9i9QNX came with the following components pre-installed:

• 2x Kingston HyperX KHX3200C20S4/8G for 16GB of DRAM
• 1x Intel Optane SSD 905p 380GB (SSDPEL1D380GA) M.2 22110 SSD
• 1x Kingston KC2000 1TB (SKC2000M81000G) M.2 2280 SSD
• ASUS Dual GeForce RTX 2070 MINI 8GB GDDR6

A configuration with a high-end PCIe 3.0 x4 NVMe SSD relegated to being a dedicated secondary drive leaves no doubt that the system represents Intel's vision of a maxed-out premium SFF PC build.​

In the next section, we take a look at the full specifications of our review sample, followed by a detailed platform analysis along with some notes on our setup experience.

Setup Notes and Platform Analysis

The NUC 9 Extreme Kit is packaged in a fancy hard casing, signifying its premium nature. Since we had a system configuration essentially ready to benchmark, the package contents only included a few manuals and regulatory information notes along with a US power cord.

We noted the re-design of the Visual BIOS in our Frost Canyon NUC review. The NUC9i9QNX also uses the same re-designed interface. Unlike the mainstream NUC10i7FNH, the NUC9i9QNX has a few enthusiast options including the ability to fine-tune the DRAM timing parameters from a performance perspective. The entire gamut of options available in the latest BIOS (v0034) of the NUC9i9QNX is brought out in the gallery below.

The specifications of our Intel NUC9i9QNX review configuration are summarized in the table below.​

Intel NUC9i9QNX (Ghost Canyon) Specifications
Processor	Intel Core i9-9980HK Coffee Lake-H, 8C/16T, 2.4 (5.0) GHz 16MB L2+L3, 14nm (optimized), 45W TDP
Memory	Kingston HyperX KHX3200C20S4/8G DDR4 SODIMM 17-19-19-35 @ 2666 MHz 2x8 GB
Graphics	ASUS Dual GeForce RTX 2070 MINI 8GB GDDR6 Intel UHD Graphics 630
Disk Drive(s)	Intel SSD 905p Optane SSDPEL1D380GA (380 GB; M.2 Type 22110 PCIe 3.0 x4 NVMe; Optane / 3D XPoint) Kingston KC2000 SKC2000M81000G (1TB; M.2 Type 2280 PCIe 3.0 x4 NVMe; Toshiba 96L 3D TLC; Silicon Motion SM2262EN Controller)
Networking	Intel Wi-Fi 6 AX200 (2x2 802.11ax - 2400 Mbps) 1x Intel I219-LM Gigabit Ethernet Adapter 1x Intel I210 Gigabit Ethernet Adapter
Audio	3.5mm Audio Jack (Front) Optical TOSLINK output (Rear) Capable of 5.1/7.1 digital output with HD audio bitstreaming (HDMI)
Miscellaneous I/O Ports	1x UHS-II SDXC Slot (Front) 2x USB 3.2 Gen 2 (10 Gbps) Type-A (Front) 4x USB 3.2 Gen 2 (10 Gbps) Type-A (Rear) 2x Thunderbolt 3 (40 Gbps) Type-C (Rear)
Operating System	Retail unit is barebones, but we installed Windows 10 Enterprise x64
Pricing (As configured)	$2810
Full Specifications	Intel NUC9i9QNX Specifications

The block diagram of the components of the NUC9i9QNX are presented in the diagram below.

One of the aspects that needs to be pointed out here is that the 2x Front USB 3.1 ports specified in the diagram above are physically the ones on the Compute Element, and not on the front panel of the chassis. The diagram does bring out the PCIe lanes bifurcation in the baseboard, though.

Our review sample came with Windows 10 Home x64 pre-installed, but, we wiped the drive and installed Windows 10 Enterprise x64 1909 along with the March 2020 cumulative updates prior to benchmarking. Our initial benchmarking and reports collection was done without opening up the system. The AIDA64 system report for the hardware configuration supplied by Intel provided the following information:

• [ North Bridge: Intel Comet Lake-H IMC ]:
  • PCIe 3.0 x8 port #2 In Use @ x8 (nVIDIA GeForce RTX 2070 Video Adapter, nVIDIA TU106 - High Definition Audio Controller)
  • PCIe 3.0 x4 port #4 In Use @ x4 (Intel Optane SSD 900p NVMe Controller)
• [ South Bridge: Intel Cannon Point CM246 ]:
  • PCIe 3.0 x1 port #1 In Use @ x1 (Intel Wi-Fi 6 AX200 160MHz Wireless Network Adapter)
  • PCIe 3.0 x4 port #5 In Use @ x4 (Intel Titan Ridge Thunderbolt 3 Controller)
  • PCIe 3.0 x4 port #9 In Use @ x4 (Silicon Motion SM2262 PCIe 3.0 x4 NVMe 1.3 SSD Controller - Kingston KC2000)
  • PCIe 3.0 x1 port #14 In Use @ x1 (Intel I210 Gigabit Network Connection)

We were puzzled by the discrete GPU operating in x8 mode instead of x16 (as it could affect gaming performance). After reaching out to Intel, we got additional context for the review configuration. As noted briefly a bit earlier, the M.2 slot on the baseboard is directly attached to the CPU, using PCIe lanes from it rather than off of the chipset. So the power-hungry, but high-performance 905p Optane drive in the baseboard slot makes sense in that context, as Intel is giving it the best possible (and least contested) connection to the CPU in order to highlight the capabilities of CPU-attached storage.

As evident from the block diagram, the Thunderbolt ports, USB ports, and the M.2 slots on the Compute Element are attached to the chipset, which means sharing the DMI link and its maximum bandwidth of PCIe 3.0 x4 (4GB/s). Which compared to typical systems, is actually a lighter load than usual; by hanging the Optane drive off of the CPU, there's less contention for the rest of the DMI link's resources, enabling additional testing areas such as elimination of noise from Thunderbolt 3 testing and allowing for RAID-0/1 on the Compute Element. Overall this opens up a few different configuration options, so in order to highlight the pros and cons of splitting the x16 slot with a SSD, we also tested the NUC with the Optane drive hanging off of the PCH, giving the video card a full x16 slot's worth of bandwidth. This is noted as x16 in the rest of the review to signify the operation of the RTX 2070 in x16 mode.

The AIDA64 system report for the x16 configuration has a slightly tweaked version of the PCIe lane usage specified earlier:

• [ North Bridge: Intel Comet Lake-H IMC ]:
  • PCIe 3.0 x16 port #2 In Use @ x8 (nVIDIA GeForce RTX 2070 Video Adapter, nVIDIA TU106 - High Definition Audio Controller)
• [ South Bridge: Intel Cannon Point CM246 ]:
  • PCIe 3.0 x1 port #1 In Use @ x1 (Intel Wi-Fi 6 AX200 160MHz Wireless Network Adapter)
  • PCIe 3.0 x4 port #5 In Use @ x4 (Intel Titan Ridge Thunderbolt 3 Controller)
  • PCIe 3.0 x4 port #9 In Use @ x4 (Intel Optane SSD 900p NVMe Controller)
  • PCIe 3.0 x1 port #14 In Use @ x1 (Intel I210 Gigabit Network Connection)
  • PCIe 3.0 x4 port #21 In Use @ x4 (Silicon Motion SM2262 PCIe 3.0 x4 NVMe 1.3 SSD Controller - Kingston KC2000)

The NUC9i9QNX is a relatively unique system. We have evaluated SFF systems with discrete GPUs in the last few years. Systems such as the Hades Canyon NUC came with a co-packaged discrete GPU leaving no scope for the end-user to upgrade the graphics performance without buying an entirely new computer. We've also had systems with a real discrete GPU such as the ASRock DeskMini Z370 GTX1060 and the Zotac ZBOX MAGNUS EN1080K. Theoretically, users could upgrade the installed MXM card to get better graphics performance, but practically speaking, MXM card upgrades are incredibly rare and almost never officially supported by system manufacturers.

This leaves us with one true precursor to the Ghost Canyon NUC - the Zotac ZBOX MAGNUS EK71080. This 5.7L compact gaming powerhouse internally sports a single-fan version of the Zotac GTX 1080 Mini complete with the external PCIe power connector, and nothing preventing the end-user from replacing it with a similar-sized GPU. Obviously, Zotac would not officially support this, but it only serves to show that accommodation of user-replaceable discrete GPUs in compact SFF systems has been possible before. The Compute Element initiative from Intel has been accompanied by the creation of an ecosystem where add-in card vendors now have an incentive to create discrete GPU cards within the 202mm x 131mm form-factor. Along with chassis designs like the one in the NUC9i9QNX, this has now enabled sub-5L systems capable of sporting powerful user-replaceable discrete GPUs.

In the table below, we have an overview of the various systems that we are comparing the Intel NUC9i9QNX against. Note that they may not belong to the same market segment. The relevant configuration details of the machines are provided so that readers have an understanding of why some benchmark numbers are skewed for or against the Intel NUC9i9QNX when we come to those sections.

Comparative PC Configurations
Aspect	Intel NUC9i9QNX (Ghost Canyon)	Intel NUC9i9QNX (Ghost Canyon)Intel NUC9i9QNX x16 (Ghost Canyon)Zotac ZBOX MAGNUS EN1080KZotac ZBOX MAGNUS EK71080Shuttle XPC Gaming Cube SZ270R9Intel NUC8i7HVK (Hades Canyon)ASRock DeskMini Z370 GTX1060
CPU	Intel Core i9-9980HK	Intel Core i9-9980HK
GPU	ASUS Dual GeForce RTX 2070 MINI 8GB GDDR6 Intel UHD Graphics 630	ASUS Dual GeForce RTX 2070 MINI 8GB GDDR6 Intel UHD Graphics 630
RAM	Kingston HyperX KHX3200C20S4/8G DDR4 SODIMM 17-19-19-35 @ 2666 MHz 2x8 GB	Kingston HyperX KHX3200C20S4/8G DDR4 SODIMM 17-19-19-35 @ 2666 MHz 2x8 GB
Storage	Intel SSD 905p Optane SSDPEL1D380GA (380 GB; M.2 Type 22110 PCIe 3.0 x4 NVMe; Optane / 3D XPoint) Kingston KC2000 SKC2000M81000G (1TB; M.2 Type 2280 PCIe 3.0 x4 NVMe; Toshiba 96L 3D TLC; Silicon Motion SM2262EN Controller)	Intel SSD 905p Optane SSDPEL1D380GA (380 GB; M.2 Type 22110 PCIe 3.0 x4 NVMe; Optane / 3D XPoint) Kingston KC2000 SKC2000M81000G (1TB; M.2 Type 2280 PCIe 3.0 x4 NVMe; Toshiba 96L 3D TLC; Silicon Motion SM2262EN Controller)
Wi-Fi	Intel Wi-Fi 6 AX200 (2x2 802.11ax - 2400 Mbps)	Intel Wi-Fi 6 AX200 (2x2 802.11ax - 2400 Mbps)
Price (in USD, when built)	$1553 (barebones) $2810 (as configured)	$1553 (barebones) $2810 (as configured)

BAPCo SYSmark 2018

The Intel NUC9i9QNX (Ghost Canyon) was evaluated using our Fall 2018 test suite for small-form factor PCs. In the first section, we will be looking at SYSmark 2018.

BAPCo's SYSmark 2018 is an application-based benchmark that uses real-world applications to replay usage patterns of business users in the areas of productivity, creativity, and responsiveness. The 'Productivity Scenario' covers office-centric activities including word processing, spreadsheet usage, financial analysis, software development, application installation, file compression, and e-mail management. The 'Creativity Scenario' represents media-centric activities such as digital photo processing, AI and ML for face recognition in photos and videos for the purpose of content creation, etc. The 'Responsiveness Scenario' evaluates the ability of the system to react in a quick manner to user inputs in areas such as application and file launches, web browsing, and multi-tasking.

Scores are meant to be compared against a reference desktop (the SYSmark 2018 calibration system, a Dell Optiplex 5050 tower with a Core i3-7100 and 4GB of DDR4-2133 memory to go with a 128GB M.2 SATA III SSD). The calibration system scores 1000 in each of the scenarios. A score of, say, 2000, would imply that the system under test is twice as fast as the reference system.

Systems equipped with 65W+ TDP desktop processors get higher scores in most workloads, though only the DeskMini Z370 manages an higher overall rating compared to the NUC9i9QNX. The surprising result is the responsiveness score for the two Ghost Canyon configurations - having the Optane drive talk directly to the CPU without the DMI bottleneck makes the system significantly more responsive.

SYSmark 2018 also adds energy measurement to the mix. A high score in the SYSmark benchmarks might be nice to have, but, potential customers also need to determine the balance between power consumption and the efficiency of the system. For example, in the average office scenario, it might not be worth purchasing a noisy and power-hungry PC just because it ends up with a 2000 score in the SYSmark 2014 SE benchmarks. In order to provide a balanced perspective, SYSmark 2018 also allows vendors and decision makers to track the energy consumption during each workload. In the graphs below, we find the total energy consumed by the PC under test for a single iteration of each SYSmark 2018 workload. For reference, the calibration system consumes 5.36 Wh for productivity, 7.71 Wh for creativity, 5.61 Wh for responsiveness, and 18.68 Wh overall.

The NUC9i9QNX is hobbled slightly by the power-hungry Optane drive and high-TDP discrete GPU, making it approach the other desktop CPU-based systems in the list when the overall energy consumption is considered. Compared to a x16 configuration, operating the GPU at x8 results in lowered energy consumption for the SYSmark 2018 workloads.

UL Benchmarks - PCMark, 3DMark, and VRMark

This section deals with a selection of the UL Futuremark benchmarks - PCMark 10, PCMark 8, and 3DMark. While the first two evaluate the system as a whole, 3DMark focuses on the graphics capabilities.

PCMark 10

UL's PCMark 10 evaluates computing systems for various usage scenarios (generic / essential tasks such as web browsing and starting up applications, productivity tasks such as editing spreadsheets and documents, gaming, and digital content creation). We benchmarked select PCs with the PCMark 10 Extended profile and recorded the scores for various scenarios. These scores are heavily influenced by the CPU and GPU in the system, though the RAM and storage device also play a part. The power plan was set to Balanced for all the PCs while processing the PCMark 10 benchmark.

PCMark 8

We continue to present PCMark 8 benchmark results (as those have more comparison points) while our PCMark 10 scores database for systems grows in size. PCMark 8 provides various usage scenarios (home, creative and work) and offers ways to benchmark both baseline (CPU-only) as well as OpenCL accelerated (CPU + GPU) performance. We benchmarked select PCs for the OpenCL accelerated performance in all three usage scenarios.

UL's 3DMark comes with a diverse set of graphics workloads that target different Direct3D feature levels. Correspondingly, the rendering resolutions are also different. The VRMark benchmark targets virtual reality specifically. Its workloads are termed as 'rooms', with each one being a piece of VR content designed to require a specific level of VR performance. We used 3DMark 2.4.4264 and VRMark 1.2.1701 to get an idea of the graphics capabilities of various systems. In this section, we take a look at the performance of the Intel NUC9i9QNX (Ghost Canyon) on a comparative basis across the different workloads.

3DMark Ice Storm

This workload has three levels of varying complexity - the vanilla Ice Storm, Ice Storm Unlimited, and Ice Storm Extreme. It is a cross-platform benchmark (which means that the scores can be compared across different tablets and smartphones as well). All three use DirectX 11 (feature level 9) / OpenGL ES 2.0. While the Extreme renders at 1920 x 1080, the other two render at 1280 x 720. The graphs below present the various Ice Storm worloads' numbers for different systems that we have evaluated.

UL 3DMark - Ice Storm Workloads

3DMark Cloud Gate

The Cloud Gate workload is meant for notebooks and typical home PCs, and uses DirectX 11 (feature level 10) to render frames at 1280 x 720. The graph below presents the overall score for the workload across all the systems that are being compared.

3DMark Sky Diver

The Sky Diver workload is meant for gaming notebooks and mid-range PCs, and uses DirectX 11 (feature level 11) to render frames at 1920 x 1080. The graph below presents the overall score for the workload across all the systems that are being compared.

3DMark Fire Strike Extreme

The Fire Strike benchmark has three workloads. The base version is meant for high-performance gaming PCs. Similar to Sky Diver, it uses DirectX 11 (feature level 11) to render frames at 1920 x 1080. The Ultra version targets 4K gaming system, and renders at 3840 x 2160. However, we only deal with the Extreme version in our benchmarking - It renders at 2560 x 1440, and targets multi-GPU systems and overclocked PCs. The graph below presents the overall score for the Fire Strike Extreme benchmark across all the systems that are being compared.

3DMark Time Spy

The Time Spy workload has two levels with different complexities. Both use DirectX 12 (feature level 11). However, the plain version targets high-performance gaming PCs with a 2560 x 1440 render resolution, while the Extreme version renders at 3840 x 2160 resolution. The graphs below present both numbers for all the systems that are being compared in this review.

UL 3DMark - Time Spy Workloads

The VRMark Professional Edition comes with three rooms. Each room can be run either in desktop or HMD mode, with varying minimum requirements for the same workload. The benchmark results include the average FPS achieved, and a score based on the FPS. A pass or fail indicator is also provided based on whether the average FPS exceeds the required FPS. In this section, we take a look at the performance of the Intel NUC9i9QNX (Ghost Canyon) on a comparative basis across the three workloads in desktop mode.

VRMark Orange Room

The Orange Room is meant to test the effectiveness of a system for handling the requirements of the HTC Vive and the Oculus Rift. The recommended hardware for both VR HMDs should be able to easily achieve the desired target FPS (88.9 fps). However, in the desktop mode, the target performance is 109 fps without any frame drops. Systems benching with an average FPS lesser than that are deemed to have failed the VRMark Orange Room benchmark. The graphs below present the average FPS and score for the different systems being considered today.

UL VRMark - Orange Room

VRMark Cyan Room

The Cyan Room sits between the Orange and Blue rooms in complexity. It is a DirectX 12 benchmark. Similar to the Orange room, the target metrics are 88.9 fps on HMDs and 109 fps on the desktop monitor. The graphs below present the average FPS and score for the different systems being considered today.

UL VRMark - Cyan Room

VRMark Blue Room

The Blue Room is the most demanding of the three workloads. At the time of introduction of VRMark in October 2016, no publicly available system running as sold was able to pass the test. The Ghost Canyon NUC configuration with the RTX 2070 performs better than any previously benchmarked system, but it still doesn't meet the requirements to pass the VRMark Blue Room test.

The performance of a system in this benchmark is an indicator of its VR-readiness for future generation of HMDs. Similar to the other workloads, the passing performance metrics are 88.9 fps on HMDs and 109 fps on desktop monitors. The complexity of the workload is due to the higher resolution (5012 x 2880) and additional geometry making it necessary to increase the number of Direct3D API calls. The graphs below present the average FPS and score for the different systems being considered today.

UL VRMark - Blue Room

SPECworkstation 3 Benchmark

SFF PCs traditionally do not lend themselves to workstation duties. However, the capabilities of the Ghost Canyon NUC encouraged us to benchmark the unit as a content creation machine. Other professional workloads were also processed using the SPECworkstation 3.0.4 benchmark from the SPEC Graphics & Workstation Performance Group.

The SPECworkstation 3 benchmark measures workstation performance based on a number of professional applications. It includes more than 140 tests based on 30 different workloads that exercise the CPU, graphics, I/O and memory hierarchy. These workloads fall into different categories.

• Media and Entertainment (3D animation, rendering)
• Product Development (CAD/CAM/CAE)
• Life Sciences (medical, molecular)
• Financial Services
• Energy (oil and gas)
• General Operations
• GPU Compute

Individual scores are generated for each test and a composite score for each category is calculated based on a reference machine (HP Z240 tower workstation using an Intel E3-1240 v5 CPU, an AMD Radeon Pro WX3100 GPU, 16GB of DDR4-2133, and a SanDisk 512GB SSD). The SPEC Ratio for the tests in each category is presented in the graphs below.

Media and Entertainment

The Media and Entertainment category comprises of workloads from five distinct applications:

• The Blender workload measures system performance for content creation using the open-source Blender application. Tests include rendering of scenes of varying complexity using the OpenGL and ray-tracing renderers.
• The Handbrake workload uses the open-source Handbrake application to transcode a 4K H.264 file into a H.265 file at 4K and 2K resolutions using the CPU capabilities alone.
• The LuxRender workload benchmarks the LuxCore physically based renderer using LuxMark.
• The Maya workload uses the SPECviewperf 13 maya-05 viewset to replay traces generated using the Autodesk Maya 2017 application for 3D animation.
• The 3ds Max workload uses the SPECviewperf 13 3dsmax-06 viewset to replay traces generated by Autodesk's 3ds Max 2016 using the default Nitrous DX11 driver. The workload represents system usage for 3D modeling tasks.

SPECworkstation 3.0.4 - Media and Entertainment Workloads

We find the Ghost Canyon NUC performing better than the reference configuration across all content creation workloads typically seen in the media and entertainment industry.

Product Development

The Product Development category comprises of eight distinct workloads:

• The Rodinia (CFD) workload benchmarks a computational fluid dynamics (CFD) algorithm.
• The WPCcfd workload benchmarks another CFD algorithm involving combustion and turbulence modeling.
• The CalculiX workload uses the Calculix finite-element analysis program to model a jet engine turbine's internal temperature.
• The Catia workload uses the catia-05 viewset from SPECviewperf 13 to replay traces generated by Dassault Systemes' CATIA V6 R2012 3D CAD application.
• The Creo workload uses the creo-02 viewset from SPECviewperf 13 to replay traces generated by PTC's Creo, a 3D CAD application.
• The NX workload uses the snx-03 viewset from SPECviewperf 13 to replay traces generated by the Siemens PLM NX 8.0 CAD/CAM/CAE application.
• The Solidworks workload uses the sw-04 viewset from SPECviewperf 13 to replay traces generated by Dassault Systemes' SolidWorks 2013 SP1 CAD/CAE application.
• The Showcase workload uses the showcase-02 viewset from SPECviewperf 13 to replay traces from Autodesk’s Showcase 2013 3D visualization and presentation application

SPECworkstation 3.0.4 - Product Development Workloads

Almost all workloads see the Ghost Canyon NUC performing significantly better than the reference configuration. The NX workload alone seems to suffer, likely on account of the trace requiring features supported in professional graphics cards.

Life Sciences

The Life Sciences category comprises of four distinct test sets:

• The LAMMPS set comprises of five tests simulating different molecular properties using the LAMMPS molecular dynamics simulator.
• The NAMD set comprises of three tests simulating different molecular interactions.
• The Rodinia (Life Sciences) set comprises of four tests - the Heartwall medical imaging algorithm, the Lavamd algorithm for calculation of particle potential and relocation in a 3D space due to mutual forces, the Hotspot algorithm to estimate processor temperature with thermal simulations, and the SRAD anisotropic diffusion algorithm for denoising.
• The Medical workload uses the medical-02 viewset from SPECviewperf 13 to determine system performance for the Tuvok rendering core in the ImageVis3D volume visualization program.

SPECworkstation 3.0.4 - Life Sciences Workloads

The trend repeats for all test sets in this category also, with the Ghost Canyon NUC acquitting itself in a creditable manner.

Financial Services

The Financial Services workload set benchmarks the system for three popular algorithms used in the financial services industry - the Monte Carlo probability simulation for risk assessment and forecast modeling, the Black-Scholes pricing model, and the Binomial Options pricing model.

The large core count and ability to turbo to speeds of up to 5 GHz enable the NUC9i9QNX to process these algorithms in a fast manner.

Energy

The Energy category comprises of workloads simulating various algorithms used in the oil and gas industry:

• The FFTW workload computes discrete Fourier transforms of large matrices.
• The Convolution workload computes the convolution of a random 100x100 filter on a 400 megapixel image.
• The SRMP workload processes the Surface-Related Multiples Prediction algorithm used in seismic data processing.
• The Kirchhoff Migration workload processes an algorithm to calculate the back propogation of a seismic wavefield.
• The Poisson workload takes advantage of the OpenMP multi-processing framework to solve the Poisson's equation.
• The Energy workload uses the energy-02 viewset from SPECviewperf 13 to determine system performance for the open-source OPendTec seismic visualization application.

SPECworkstation 3.0.4 - Energy Industry Workloads

The NUCs seem to perform very poorly in the SRMP and Poisson workloads, but are otherwise quite good in the workloads in this category.

General Operations

In the General Options category, the focus is on workloads from widely used applications in the workstation market:

• The 7zip workload represents compression and decompression operations using the open-source 7zip file archiver program.
• The Python workload benchmarks math operations using the numpy and scipy libraries along with other Python features.
• The Octave workload performs math operations using the Octave programming language used in scientific computing.
• The Storage workload evaluates the performance of the underlying storage device using transaction traces from multiple workstation applications.

SPECworkstation 3.0.4 - General Operations

The Core i9-9980HK performs much better than the reference configuration as well as the older generation CPUs for the general operations. The storage results is again particularly interesting to analyze. Directly connecting the Optane SSD to the CPU's PCIe lanes results in a 40%+ performance improvement for the storage traces of professional applications.

GPU Compute

In the GPU Compute category, the focus is on workloads taking advantage of the GPU compute capabilities using either OpenCL or CUDA, as applicable:

• The LuxRender benchmark is the same as the one seen in the media and entertainment category.
• The Caffe benchmark measures the performance of the Caffe deep-learning framework.
• The Folding@Home benchmark measures the performance of the system for distributed computing workloads focused on tasks such as protein folding and drug design.

SPECworkstation 3.0.4 - GPU Compute

Miscellaneous Performance Metrics

This section looks at some of the other commonly used benchmarks representative of the performance of specific real-world applications.

3D Rendering - CINEBENCH R15

We use CINEBENCH R15 for 3D rendering evaluation. The program provides three benchmark modes - OpenGL, single threaded and multi-threaded. Evaluation of different PC configurations in all three modes provided us the following results.

The results track what was observed in the media and entertainment category workloads in SPECworkstation 3.

x265 Benchmark

Next up, we have some video encoding benchmarks using x265 v2.8. The appropriate encoder executable is chosen based on the supported CPU features. In the first case, we encode 600 1080p YUV 4:2:0 frames into a 1080p30 HEVC Main-profile compatible video stream at 1 Mbps and record the average number of frames encoded per second.

Our second test case is 1200 4K YUV 4:2:0 frames getting encoded into a 4Kp60 HEVC Main10-profile video stream at 35 Mbps. The encoding FPS is recorded.

The Ghost Canyon NUC surprisingly performs better than systems equipped with CPUs sporting much higher TDPs.

7-Zip

7-Zip is a very effective and efficient compression program, often beating out OpenCL accelerated commercial programs in benchmarks even while using just the CPU power. 7-Zip has a benchmarking program that provides tons of details regarding the underlying CPU's efficiency. In this subsection, we are interested in the compression and decompression rates when utilizing all the available threads for the LZMA algorithm.

Given the brief nature of the benchmark workload and the large number of available threads, it is no surprise that the Ghost Canyon NUC performs significantly better than the rest of the systems in this workload.

Cryptography Benchmarks

Cryptography has become an indispensable part of our interaction with computing systems. Almost all modern systems have some sort of hardware-acceleration for making cryptographic operations faster and more power efficient. In this sub-section, we look at two different real-world applications that may make use of this acceleration.

BitLocker is a Windows features that encrypts entire disk volumes. While drives that offer encryption capabilities are dealt with using that feature, most legacy systems and external drives have to use the host system implementation. Windows has no direct benchmark for BitLocker. However, we cooked up a BitLocker operation sequence to determine the adeptness of the system at handling BitLocker operations. We start off with a 2.5GB RAM drive in which a 2GB VHD (virtual hard disk) is created. This VHD is then mounted, and BitLocker is enabled on the volume. Once the BitLocker encryption process gets done, BitLocker is disabled. This triggers a decryption process. The times taken to complete the encryption and decryption are recorded. This process is repeated 25 times, and the average of the last 20 iterations is graphed below.

The BitLocker benchmark results are a bit surprising, particularly given the clear performance benefits of the Core i9-9980HK for cryptography applications in the other applications below.

Creation of secure archives is best done through the use of AES-256 as the encryption method while password protecting ZIP files. We re-use the benchmark mode of 7-Zip to determine the AES256-CBC encryption and decryption rates using pure software as well as AES-NI. Note that the 7-Zip benchmark uses a 48KB buffer for this purpose.

As expected, the 8C/16T configuration allows for fast encryption and decryption irrespective of the use of pure software or AES-NI instructions.

Yet another cryptography application is secure network communication. OpenSSL can take advantage of the acceleration provided by the host system to make operations faster. It also has a benchmark mode that can use varying buffer sizes. We recorded the processing rate for a 8KB buffer using the hardware-accelerated AES256-CBC-HAC-SHA1 feature.

These results are not surprising given the core count and operating frequency profile of the CPU in the Ghost Canyon NUC.

Agisoft Photoscan

Agisoft PhotoScan is a commercial program that converts 2D images into 3D point maps, meshes and textures. The program designers sent us a command line version in order to evaluate the efficiency of various systems that go under our review scanner. The command line version has two benchmark modes, one using the CPU and the other using both the CPU and GPU (via OpenCL). We present the results from our evaluation using the CPU mode only. The benchmark (v1.3) takes 84 photographs and does four stages of computation:

• Stage 1: Align Photographs (capable of OpenCL acceleration)
• Stage 2: Build Point Cloud (capable of OpenCL acceleration)
• Stage 3: Build Mesh
• Stage 4: Build Textures

We record the time taken for each stage. Since various elements of the software are single threaded, and others multithreaded, it is interesting to record the effects of CPU generations, speeds, number of cores, and DRAM parameters using this software.

The Ghost Canyon NUC is better than any of the other considered systems across all Photoscan stages.

Dolphin Emulator

Wrapping up our application benchmark numbers is the new Dolphin Emulator (v5) benchmark mode results. This is again a test of the CPU capabilities.

In fact, the 249s taken by the Ghost Canyon NUC is the fastest amongst all SFF PCs we have evaluated with this benchmark.

GPU Performance - Gaming Workloads

The gaming test suite for SFF PCs was revamped in mid-2018 with six different games:

• Civlization VI (DX12)
• Dota 2
• F1 2017
• Grand Theft Auto V
• Middle Earth: Shadow of War
• Far Cry 5

Most system reviews take a handful of games and process them at one resolution / quality settings for comparison purposes. Recently, we have seen many pre-built systems coming out with varying gaming capabilities. Hence, it has become imperative to give consumers an idea of how a given system performs over a range of resolutions and quality settings for each game. Our test suite is able to address this aspect.

Civilization VI (DX12)

The Civilization series of turn-based strategy games is very popular. For such games, the frame rate is not necessarily an important factor in the gaming experience. However, with Civilization VI, Firaxis has cranked up the visual fidelity to make the game more attractive. As a result, the game can be taxing on the GPU as well as the CPU, particularly in the DirectX 12 mode.

Civilization VI (DirectX 12) Performance

We processed the built-in benchmark at two different resolutions (1080p and 2160p), and with two different quality settings (medium and ultra, with the exact differences detailed here). There are no surprises here, with the RTX 2070 performing way better than the previous-generation GPUs. Operating the RTX 2070 in x16 mode results in slightly better performance (around 3 - 4 fps better, but very much dependent on the quality and resolution) compared to a x8 configuration

Dota 2

Dota 2 has been featuring in our mini-PC and notebook reviews for a few years now, but, it still continues to be a very relevant game. Our evaluation was limited to a custom replay file at 1080p resolution with enthusiast settings ('best-looking' preset). We have now revamped our testing to include multiple resolutions - This brings out the fact that the game is CPU-limited in many configurations.

Dota 2 allows for multiple renderers - we use the DirectX 11 mode. The rendering settings are set to 'enthusiast level' (best-looking, which has all options turned on, and at Ultra level, except for the Shadow Quality set to 'High'). We cycle through different resolutions after setting the monitor resolution to match the desired resolution. The core scripts and replay files are sourced from Jonathan Liebig's original Dota 2 benchmarking instructions which used a sequence of frames from Match 3061101068.

Dota 2 - Enthusiast Quality Performance

Due to the CPU-limited nature of the game, the Ghost Canyon NUC is not able to emerge as a clear winner here. However, the performance is more than enough to provide an excellent experience across all resolutions and quality settings for Dota 2.

F1 2017

Our gaming system reviews have always had a representative racing game in it. While our previous benchmark suite for PCs featured Dirt 2, we have moved on to the more recent F1 2017 from Codemasters for our revamp.

F1 2017 - Ultra Quality Performance

The supplied example benchmark (with some minor tweaks) is processed at four different resolutions while maintaining the graphics settings at the built-in 'Ultra' level.

Grand Theft Auto V

GTA doesn’t provide graphical presets, but opens up the options to users and extends the boundaries by pushing even the hardest systems to the limit using Rockstar’s Advanced Game Engine under DirectX 11. Whether the user is flying high in the mountains with long draw distances or dealing with assorted trash in the city, when cranked up to maximum it creates stunning visuals but hard work for both the CPU and the GPU. For our test we have scripted a version of the in-game benchmark. The in-game benchmark consists of five scenarios: four short panning shots with varying lighting and weather effects, and a fifth action sequence that lasts around 90 seconds. We use only the final part of the benchmark, which combines a flight scene in a jet followed by an inner city drive-by through several intersections followed by ramming a tanker that explodes, causing other cars to explode as well. This is a mix of distance rendering followed by a detailed near-rendering action sequence.

Grand Theft Auto V Performance

We processed the benchmark across various resolutions and quality settings (detailed here). The results are presented above.

Middle Earth: Shadow of War

Middle Earth: Shadow of War is an action RPG. In our previous gaming benchmarks suite, we used its prequel - Shadow of Mordor. Produced by Monolith and using the new LithTech Firebird engine and numerous detail add-ons, Shadow of War goes for detail and complexity. The graphics settings include standard options such as Graphical Quality, Lighting, Mesh, Motion Blur, Shadow Quality, Textures, Vegetation Range, Depth of Field, Transparency and Tessellation. There are standard presets as well. The game also includes a 'Dynamic Resolution' option that automatically alters graphics quality to hit a pre-set frame rate. We benchmarked the game at four different resolutions - 4K, 1440p, 1080p, and 720p. Two standard presets - Ultra and Medium - were used at each resolution after turning off the dynamic resolution option.

Middle Earth: Shadow of War Performance

While the Ghost Canyon NUC is the best of the lot, the interesting aspect to observe here is that x8 or x16 essentially makes no difference to the game's performance.

Far Cry 5

Ubisoft's Far Cry 5 is an action-adventure first-person shooter game released in March 2018. The game comes with an in-built benchmark and has standard pre-sets for quality settings. We benchmarked the game at four different resolutions - 720p, 1080p, 1440p, and 2160p. Two preset quality settings were processed at each resolution - normal and ultra.

Far Cry 5 Performance

As expected, the Ghost Canyon NUC configuration performs much better than any of the other configurations in the list. The x16 mode delivers slightly more FPS compared to the x8 mode across all tested resolutions and quality settings, unlike what we saw in the Shadow of War gaming workload.

Overall, the NUC9i9QNX combined with the ASUS RTX 2070 Mini manages to deliver more gaming performance than any other SFF PC we have seen so far - even managing to surpass larger systems in the process.

GPU Performance for Workstations - SPECviewperf 13

The SPECviewperf benchmark from SPEC provides an idea of the capabilities of the GPU in a workstation from the perspective of different CAD, content creation, and visual data analysis tools. It makes more sense to process these benchmarks on workstations with professional GPUs, but, consumer GPUs are often the choice for machines that need to handle both gaming and professional workloads.

SPECviewperf 13 includes nine different workloads representative of graphics content and behavior of actual applications. They make use of the OpenGL 4.0 and DirectX 12 APIs under Windows. SPECviewperf 13's workloads (termed viewsets) can officially be run only at two desktop resolutions (1920 x 1080, and 3840 x 2160), and need the display scaling to be set to 100% (DPI of 96). The available viewsets are listed below.

• 3ds Max (3dsmax-06)
• CATIA (catia-05)
• Creo (creo-02)
• Energy (energy-02)
• Maya (maya-05)
• Medical (medical-02)
• Showcase (showcase-02)
• Siemens NX (snx-03)
• Solidworks (sw-04)

The 3ds Max and Showcase viewsets are available only when processing at 1920 x 1080. The rest are available at both resolutions.

We processed SPECviewperf 13 at both resolutions on the Intel NUC9i9QNX (Ghost Canyon). The benchmark measures the frame rate at which the GPU renders the scenes in a viewset. Each viewset is composed of different scenes and rendering modes, and the composite score for the viewset is a weighted geometric mean of the FPS measured for the different scenes. In this section, we take a look at how its composite scores stack up against other systems targeting this market segment.

3ds Max (3dsmax-06)

The 3dsmax-06 viewset comprises of 11 different scenes. They have been created from traces of the graphics workload generated by Autodesk 3ds Max 2016 using the default Nitrous DX11 driver. Additional details are available here.

The x16 configuration for the RTX 2070 delivers significantly better results compared to the x8 configuration.

CATIA (catia-05)

The catia-05 viewset comprises of 14 different tests created from traces of the graphics workload generated by the CATIA V6 R2012 application from Dassault Systemes. Additional details are available here.

SPECviewperf 13: CATIA Viewset Composite Scores

The RTX 2070 in the NUC9i9QNX doesn't emerge as a leader in this workload, but its performance is almost as good as the top candidate in the x16 configuration.

Creo (creo-02)

The creo-02 viewset comprises of 16 different tests created from traces of the graphics workload generated by the Creo 3 and Creo 4 applications from PTC. Additional details are available here.

SPECviewperf 13: Creo Viewset Composite Scores

The significant difference between the x16 mode and x8 mode continues here, with the former putting the NUC9i9QNX within touching distance of the top performers.

Energy (energy-02)

The energy-02 viewset comprises of 6 different tests based on techniques used by the OpendTect seismic visualization application. Additional details are available here.

SPECviewperf 13: Energy Viewset Composite Scores

The NUC9i9QNX performs well in this workload, emerging as the top candidate. Strangely, the 1080p configuration seems limited by an unknown factor, with the x8 configuration performing better than the x16. At 4K, however, normal service is resumed.

Maya (maya-05)

The maya-05 viewset comprises of 10 different tests based on traces of the graphics workload generated by Autodesk Maya 2017. Additional details are available here.

SPECviewperf 13: Maya Viewset Composite Scores

The x16 configuration of the RTX 2070 emerges as the clear topper for the Maya workload.

Medical (medical-02)

The medical-02 viewset comprises of 8 different tests derived from 4 distinct datasets. Each test uses the ImageVis3D volume visualization program's Tuvok rendering core for 2D projections of 3D volumetric grids. Additional details are available here.

SPECviewperf 13: Medical Viewset Composite Scores

The NUC9i9QNX emerges second in this workload, next to the Core i7-7700HQ / GTX 1080 combination in the Zotac EK71080.

Showcase (showcase-02)

The showcase-02 viewset comprises of 4 tests created from traces of the Autodesk Showcase 2013 application rendering a racecar model with 8 million vertices using different modes. Additional details are available here.

The x16 RTX 2070 configuration is the clear winner in this workload.

Siemens NX (snx-03)

The snx-03 viewset comprises of 10 tests created with traces from the graphics workload generated by the NX 8.0 application from Siemens PLM. Additional details are available here.

SPECviewperf 13: Siemens NX Viewset Composite Scores

The performance of various systems in this workload is poor (as evident from the SPECworkstation ratings for the corresponding workload). There is not much to separate the various configurations in this workload, as evident from the scores in the graphs above.

Solidworks (sw-04)

The sw-04 viewset comprises of 11 tests created from traces of Dassault Systemes’ SolidWorks 2013 SP1 application. Additional details are available here.

SPECviewperf 13: Solidworks Viewset Composite Scores

Overall, the RTX 2070 in the NUC9i9QNX delivers excellent performance for GPU-intensive workstation workloads. In all cases, the x16 mode provides much better performance compared to the original x8 configuration of our review sample.

HTPC Credentials - Display Outputs Capabilities

The NUC9i9QNX comes with three native display outputs from the Compute Element, and their characteristics are summarized in the table below. From a HTPC use-case perspective, the entries of interest include the ability to support UHD (3840 x 2160) or higher resolutions, along with HDCP 2.2. The latter enables the display output to be used for viewing protected content such as 4K Netflix streams and play back UltraHD Blu-rays.

NUC9i9QNB Display Outputs
	HDMI	2x Thunderbolt 3
Version	2.0a	DisplayPort 1.2
Max. Video Output	3840x2160 @ 60Hz	4096x2160 @ 60Hz
HDCP	Yes (2.2)
HDR	Yes	No
HD Audio Bitstreaming	Yes

The BIOS of the NUC9i9QNX also allows for switchable graphics. The ASUS Dual GeForce RTX 2070 MINI 8GB GDDR6 in our review configuration supports a maximum of four displays using three additional display outputs (the DisplayPort output supports multi-stream transport and can drive additional displays down the chain). The end implication is the ability of the system to simultaneously drive a total of 7 independent displays. The table below lists the display outputs of the RTX 2070 card in our review sample.

ASUS Dual GeForce RTX 2070 MINI Display Outputs
	DVI-D	HDMI	DisplayPort
Version	Dual-Link	2.0b	1.4
Max. Video Output	2560x1600 @ 60Hz	3840x2160 @ 60Hz	7680x4320 @ 60Hz
HDCP	Yes (2.2)
HDR	No	Yes
HD Audio Bitstreaming	No	Yes

Supporting the display of high-resolution protected video content is a requirement for even a casual HTPC user. In addition, HTPC enthusiasts also want their systems to support refresh rates that either match or be an integral multiple of the frame rate of the video being displayed. Most displays / AVRs are able to transmit the supported refresh rates to the PC using the EDID metadata. In some cases, the desired refresh rate might be missing in the list of supported modes.

Display Refresh Rates - NUC9i9QNB

Our evaluation of the NUC9i9QNX as a HTPC was first done using the native HDMI output of the Compute Element (NUC9i9QNB) connected to a TCL 55P607 4K HDR TV via a Denon AVR-X3400H AV receiver. We tested out various display refresh rates ranging from 23.976 Hz to 59.94 Hz. Of particular interest is the 23.976 Hz (23p) setting, which Intel used to have trouble with in the pre-Broadwell days.

The gallery below presents screenshots from the other refresh rates that were tested. The system has no trouble maintaining a fairly accurate refresh rate throughout the duration of the video playback.

Display Refresh Rates - ASUS Dual GeForce RTX 2070 MINI

Our initial HTPC evaluation was followed up by using the native HDMI output of the RTX 2070 connected to a TCL 55P607 4K HDR TV via a Denon AVR-X3400H AV receiver. We tested out various display refresh rates ranging from 23.976 Hz to 59.94 Hz.

The gallery below presents screenshots from the other refresh rates that were tested. Similar to the Intel HDMI output case, the system has no trouble maintaining a fairly accurate refresh rate throughout the duration of the video playback.

UHD Blu-ray Playback Support

UHD Blu-ray playback is currently supported when using the HDMI port driven by select Intel GPUs. It also needs SGX support. The NUC9i9QNX ticks all required items, as shown by the CyberLink Ultra HD Blu-ray Advisor tool in the screenshot below.

Using CyberLink's latest PowerDVD 20, we were able to successfully play back a UHD Blu-ray, as shown above.

HTPC Credentials - YouTube and Netflix Streaming

Our HTPC testing with respect to YouTube had been restricted to playback of a 1080p music video using the native HTML5 player in Firefox. The move to 4K, and the need to evaluate HDR support have made us choose Mystery Box's Peru 8K HDR 60FPS video as our test sample moving forward. On PCs running Windows, it is recommended that HDR streaming videos be viewed using the Microsoft Edge browser after putting the desktop in HDR mode.

YouTube Streaming - Intel Display Output


YouTube Streaming - NVIDIA Display Output

Both GPUs support VP9 Profile 2 decoding in hardware and the system has no trouble playing back the 4K HDR stream in either of the above scenarios.

Various metrics of interest such as GPU usage and at-wall power consumption were recorded for the first four minutes of the playback of the above video. The numbers are graphed below.

The media engine usage stays well short of saturation despite the need to process a 4Kp60 stream. In the steady state, the power consumption at the wall for the playback averaged around 62W for the Compute Elemen's HDMI port as well as the RTX 2070's HDMI port.

The Netflix 4K HDR capability works with native Windows Store app as well as the Microsoft Edge browser. We used the Windows Store app to evaluate the playback of Season 4 Episode 4 of the Netflix Test Patterns title. The OS screenshot facilities obviously can't capture the video being played back. However, the debug OSD (reachable by Ctrl-Alt-Shift-D) can be recorded.

Netflix Streaming - Intel Display Output

Netflix Streaming - NVIDIA Display Output

The (hevc,hdr,prk) entry corresponding to the Video Track in the debug OSD, along with the A/V bitrate details (192 kbps / 16 Mbps) indicate that the HDR stream is indeed being played back. Playing the title using the RTX 2070's HDMI port consumed around 68W at the wall in the steady state. The consumption for playback using the Compute Element's HDMI port in the same scenario was 56W.

HTPC Credentials - Local Media Playback and Video Processing

Evaluation of local media playback and video processing is done by playing back files encompassing a range of relevant codecs, containers, resolutions, and frame rates. A note of the efficiency is also made by tracking GPU usage and power consumption of the system at the wall. Users have their own preference for the playback software / decoder / renderer, and our aim is to have numbers representative of commonly encountered scenarios. Towards this, we played back the test streams using the following combinations:

• MPC-HC x64 1.8.5 + LAV Video Decoder (DXVA2 Native) + Enhanced Video Renderer - Custom Presenter (EVR-CP)
• MPC-HC x64 1.8.5 + LAV Video Decoder (D3D11) + madVR 0.92.17 (DXVA-Focused)
• MPC-HC x64 1.8.5 + LAV Video Decoder (D3D11) + madVR 0.92.17 (Lanczos-Focused)
• VLC 3.0.8
• Kodi 18.6

The thirteen test streams (each of 90s duration) were played back from the local disk with an interval of 30 seconds in-between. Various metrics including GPU usage and at-wall power consumption were recorded during the course of this playback. Prior to looking at the metrics, a quick summary of the decoding capabilities of the two available GPUs is useful to have for context.

The NVIDIA GeForce RTX 2070 and Intel UHD Graphics 630 both have similar video decoding support. Based on paper specifications, we can surmise that the power efficiency while decoding using Intel's GPU should be better, while more powerful video post-processing steps could be used with the RTX 2070's GPU compute capabilities without dropping frames in the playback process.

All our playback tests were done with the desktop HDR setting turned on. It is possible for certain system configurations to have madVR automatically turn on/off the HDR capabilities prior to the playback of a HDR video, but, we didn't take advantage of that in our testing.

VLC and Kodi

VLC is the playback software of choice for the average PC user who doesn't need a ten-foot UI. Its install-and-play simplicity has made it extremely popular. Over the years, the software has gained the ability to take advantage of various hardware acceleration options. Kodi, on the other hand, has a ten-foot UI making it the perfect open-source software for dedicated HTPCs. Support for add-ons make it very extensible and capable of customization. We played back our test files using the default VLC and Kodi configurations, and recorded the following metrics.

Video Playback Efficiency - VLC and Kodi

VLC and Kodi have no trouble playing back all our test streams using either GPU, given that video post-processing steps making use of GPU compute resources are minimal to non-existent. There are no dropped frames. In terms of power efficiency, VLC (Intel) averages around 49W for the complete playback benchmarking process. The corresponding numbers for VLC (NVIDIA), Kodi (Intel), and Kodi (NVIDIA) are 58W, 43W, and 52W respectively. Given these numbers, it might make sense to utilize the HDMI port from the Compute Element for media playback duties using vanilla Kodi / VLC, and use the NVIDIA display outputs only for gaming purposes in a HTPC environment.

MPC-HC

MPC-HC offers an easy way to test out different combinations of decoders and renderers. The first configuration we evaluated is the default post-install scenario, with only the in-built LAV Video Decoder forced to DXVA2 Native mode. Two additional passes were done with different madVR configurations. In the first one (DXVA-focused), we configured madVR to make use of the DXVA-accelerated video processing capabilities as much as possible. In the second (Lanczos-focused), the image scaling algorithms were set to 'Lanczos 3-tap, with anti-ringing checked'. Chroma upscaling was configured to be 'BiCubic 75 with anti-ringing checked' in both cases. The metrics collected during the playback of the test files using the above three configurations are presented below.

Video Playback Efficiency - MPC-HC with EVR-CP and madVR

The UHD Graphics 630's struggles with madVR's GPU compute-intensive rendering algorithms (in the Lanczos case) are well known from our previous HTPC reviews. The key here is the performance of the RTX 2070 in those scenarios. The GPU handles the madVR configuration with aplomb, rarely breaking a sweat and keeping the GPU loading factor below 40% for the most part. This means that we have GPU compute resources to spare and HTPC enthusiasts can tweak the madVR algorithm settings further to suit their personal tastes without the risk of choking the GPU. From a power efficiency viewpoint, the Intel EVR-CP configuration is the best at around 50W, while the madVR (Lanczos) NVIDIA combination comes in around 63W. These numbers are the average power consumption numbers over the length of the playback benchmarking process, and are along expected lines.

Power Consumption and Thermal Performance

The power consumption at the wall was measured with a 4K display being driven through the HDMI port of the discrete GPU. In the graphs below, we compare the idle and load power of the Intel NUC9i9QNX (Ghost Canyon) with other SFF PCs evaluated before. For load power consumption, we ran the AIDA64 System Stability Test with various stress components, as well as a combination of Prime95 and Furmark, and noted the maximum sustained power consumption at the wall.

The power consumption numbers are along expected lines, matching what we had observed for a similar mobile CPU / desktop GPU combination in the Kaby Lake / Pascal generation (Zotac ZBOX MAGNUS EK71080). The additional factors leading to the slightly higher numbers (295W vs 272W) include a dedicated secondary drive, a power-hungry 905p series Optane SSD, and a relaxation in the permissible sustained package power consumption that we analyze further down.

Our thermal stress routine starts with the system at idle, followed by four stages of different system loading profiles using the AIDA64 System Stability Test (each of 30 minutes duration). In the first stage, we stress the CPU, caches and RAM. In the second stage, we add the GPU to the above list. In the third stage, we stress the GPU standalone. In the final stage, we stress all the system components (including the disks). Beyond this, we leave the unit idle in order to determine how quickly the various temperatures in the system can come back to normal idling range. The various clocks, temperatures and power consumption numbers for the system during the above routine are presented in the graphs below.

Intel NUC9i9QNX (Ghost Canyon) System Loading with the AIDA64 System Stability Test

The AIDA64 system stability test does not have many surprises in store. The core frequencies stay well above the advertised base speeds of 2.4 GHz throughout the loading period. The package and core temperature stay around 90C for the most part, though they do momentarily reach the junction temperature (100C) as the transition from a full-stress to a standalone GPU stress scenario takes place. Things do get to a stable state quite soon. Meanwhile the RTX 2070 never crosses 70C, pointing to a well-designed thermal solution by ASUS.

On the power consumption side, we have the first surprise - the CPU package is allowed its PL2 limits for a short duration (close to 100W) in the beginning. In the steady loading stage, the number drops down to 65W (compared to the 45W we saw for the Core i7-7700HQ in the Zotac ZBOX MAGNUS EK71080). This accounts for the additional load power consumption identified in the beginning of this section. The GPU power consumption tops out around 165W in the AIDA64 stress test. The at-wall numbers show spikes of up to 325W, but the steady loading state numbers are around 280W with all stress components enabled.

Intel NUC9i9QNX (Ghost Canyon) System Loading with Prime95 and Furmark

The frequency-related observations made in the AIDA64 system stability test hold true for our artificial power virus test involving Prime95 and Furmark also. The clocks stay well above the advertised base numbers. The temperature numbers tell a similar story too, though the absolute numbers are a bit lower, coming in around 80C instead of 90C for the package. The GPU temperature stays south of 70C throughout. On the power consumption side, we again see the CPU package allowed to dissipate up to 65W in the steady state. The at-wall numbers are slightly north of 290W in the steady state, as graphed in the beginning of this section.

Overall, our power consumption and thermal solution testing revealed a few surprises - the CPU package power budget is 65W (instead of the 45W we saw with mobile CPUs used in SFF PCs in the previous generation). The system as a whole is able to handle this without any undue cause for alarm with respect to the temperatures. ASUS does deserve plaudits for a well-designed thermal solution that doesn't allow the GPU to go beyond 70C even in high-stress scenarios. The system as a whole is surprisingly quiet at idle given the dimensions of the chassis and the number of fans in the system.

Miscellaneous Aspects: Storage Performance

The preceding sections analyzed the performance of two Intel NUC9i9QNX configurations for a variety of workloads, ranging from office usage to professional and industrial workstation applications. We also looked at gaming performance and the suitability of the PC for home-theater use. One of the aspects that we touched on and off across all sections was the evaluation of the storage subsystem. A little bit of additional analysis is in order. In particular, we operated the Intel SSD 905p Optane drive in two modes - directly attached to the CPU's PCIe lanes, and attachment through the PCH.

Intel NUC9i9QNX Storage-Specific Benchmarks

SPECworkstation 3.0.4 wpcStorage SPEC RatioPCMark 10 Storage Bandwidth (Primary Drive)PCMark 10 Storage Access Time (Primary Drive)PCMark 10 Storage Score (Primary Drive)PCMark 10 Storage Bandwidth (Secondary Drive)PCMark 10 Storage Access Time (Secondary Drive)PCMark 10 Storage Score (Secondary Drive)

Directly attaching the Optane drive to the CPU's PCIe lanes yields 40%+ benefit for workstations based on SPECworkstation 3's wpcStorage workload. PCMark 10's storage bench shows a 50%+ increase in storage bandwidth and a 35%+ decrease in average access time for consumer workloads. The storage bandwidth for the secondary drive attached to the PCH also suffers when the primary drive contends with it for access to the CPU through the DMI link, as shown in the PCMark 10 storage bandwidth graph for the secondary drive above.

On the networking side, we are yet to set up our 802.11 ax / Wi-Fi 6 testbed for small form-factor PCs, and hence, there are no bandwidth numbers to report yet. However, it must be noted that the NUC 9 Extreme Kits, like the Frost Canyon NUC we recently reviewed, come with 802.11ax / Wi-Fi 6 support, and its theoretical maximum bandwidth of 2400 Mbps betters the 867 Mbps offered by the Wireless-AC 8265 in the Hades Canyon NUC and the 1733 Mbps offered by the Wireless-AC 9560 in the Bean Canyon NUC. The AX200 WLAN component takes advantage of the MAC built into the CM246 chipset, but uses a dedicated PCIe x1 link to interface (unlike the AX201 / CNVi combination in the Frost Canyon NUC). The AX200 has a 2x2 simultaneous dual-operation in 2.4 GHz and 5 GHz bands and also comes with support for 160 MHz-wide channels.

Concluding Remarks

Following on the trail of Skull Canyon and Hades Canyon, Intel's Ghost Canyon NUC (NUC9i9QNX) is the latest and greatest performance-oriented mini-PC to come out of the company. A love letter of sorts to enthusiasts who want it all in a mini-PC, it's improved on Intel's earlier designs in a number of ways, making it perhaps the best high-performance NUC yet. In particular, by putting in enough room for a standard PCIe video card – but without making the NUC itself so large that it ceases to be a small form factor PC – Intel has resolved the one issue that has always dogged these NUCs: these days the GPU will go out of date long before the CPU will. All of which has made for one of the most interesting SFF PCs we've looked at in some time.

Overall, Ghost Canyon has given us with the opportunity to evaluate a sub-5L SFF PC sporting a user-replaceable discrete GPU. Intel's specific review sample configuration has also allowed us to explore the effect of the PCIe link width on various GPU workloads and get an idea of the increased responsiveness of the system when the primary storage drive is directly attached to the CPU and not bottlenecked by the DMI link. The pros and cons of the NUC9i9QNX are summarized below:

Pros:

• The NUC 9 Extreme Kits are completely unique systems unparalleled in terms of performance potential in the sub-5L chassis volume category
• The system design of the NUC9i9QNX allows for a sustained 65W processor package power dissipation, ensuring that workloads can take full advantage of all eight physical cores at good clock rates
• The current NUC 9 Extreme Kits cover all bases in terms of I/O for almost every market segment, with the Compute Element alone providing more I/O capabilities compared to any other off-the-shelf SFF PC currently in the market
• The Wi-Fi antenna pigtail connectors on the Compute Element are vastly improved compared to the previous NUC boards. The shifting of the coaxial receptacle connector to the MMCX Micromate style makes it easier to affix and have a more secure connection to the board as well as the chassis.
• The components of the NUC 9 Extreme Kits can be upgraded independent of each other. The discrete GPU / PCIe add-in cards make up a significant chunk of the upgrades made to a desktop PC over its lifetime. The Compute Element initiative makes that a simple affair
• The NUC 9 Extreme Kits belongs to the rare breed of SFF PCs that support switchable graphics. For example, all six display outputs from the Hades Canyon NUC are driven by the Radeon GPU, and the case with the ZBOX MAGNUS SFF PCs is similar. The NUC9i9QNX allows simultaneous usage of display outputs from both the integrated GPU as well as the discrete one

Cons:

• The cramming of a large number of components in a tight space throws challenges in cable management
• The front I/O ports of the NUC 9 Extreme Kits are sourced off a hub chip on a separate daughterboard. The cable linking the Compute Element to the daughterboard may easily get unseated during dGPU installation, leading the hub to operate at USB 2.0 speeds (this was the case in our review sample until a full disassembly and reassembly was done. The process restored the advertised USB 3.1 Gen 2 operation)
• Ease of installation has been a hallmark of all the NUC kits from Intel so far. The need to keep things as small as possible means that the NUC 9 Extreme Kits end up making some compromises in this aspect. In particular, scenarios where a discrete GPU needs to be installed in the cramped space are a bit of challenge
• Long-term thermals in dusty environments may be of concern. Quick cleaning access to the fan in the Compute Element is not available when a discrete GPU is installed
• The 35W+ idle power consumption of the NUC9i9QNX review configuration is a tad too high for traditional NUC enthusiasts used to sub-10W idling numbers (even accounting for a discrete GPU in the mix)
• The x8 vs. x16 PCIe link width tradeoff for the discrete GPU is a tough choice to make. With the current configuration of the NUC 9 Extreme Kits, we either get increased system responsiveness or better performance for GPU compute workloads, but not both at the same time

The Ghost Canyon NUC9i9QNX is a SFF enthusiast's dream come true. The NUC 9 Extreme Kits completely re-define the standard for other SFF PCs in the market. Beyond the product itself, the ecosystem that Intel is slowly developing around the Compute Element initiative holds importance in the longer term. Getting add-in card vendors to design for a compact chassis with well-defined requirements is a great first step. Moving forward, we would like to see some innovation around power delivery from the PSU to various build components. If we were to be given a choice of one thing that could be fixed in the NUC 9 Extreme kits, it is the elimination of the sea of PSU cables and the associated volume requirements / management headache.

The NUC9i9QNX review sample configuration allowed us to explore the benefits of attaching Optane storage directly to the CPU without the DMI limitations. However, we also saw that the operation of the GPU in x8 mode instead of x16 resulted in noticable penalties for GPU-intensive workstation workloads. Fortunately, gaming workloads were much milder, and only saw a difference of a few FPS. These two sets of observations make us yearn for Thunderbolt 3 and M.2 PCIe x4 ports directly attached to the CPU in addition to dedicating a x16 link for the GPU. An upgrade of the gigabit ports to NBASE-T would also be welcome. Some items in this wish-list are already in Intel's future roadmap. Hopefully, we will be seeing all these in future Compute Elements. The initiative replaces the socketed CPU currently identified as the core of a DIY desktop upgrade with an 'add-in card' form-factor Compute Element.

Discrete GPUs in SFF PCs - The NUC9i9QNX (L) and the Zotac ZBOX MAGNUS EK71080 (R)

The OEM perspective is also an interesting aspect. Prior to the launch of the first NUC UCFF PC, vendors like Zotac had been playing around with slightly larger mini-PCs such as the MAG-ND01 (a 7in x 7in board compared to the NUC's 4in x 4in size). The launch of the Ghost Canyon NUC kits reminds us of the same. Zotac was one of the first vendors to put a discrete user-replaceable GPU in a ~5L chassis in the ZBOX MAGNUS EK71080 (though they didn't advertise the user-replaceable part to end-users). Intel has now managed to integrate similar capabilities in a more compact chassis.

In both the original NUC and the current Compute Element initiative, it has come down to Intel to take a proof-of-concept from one of its OEM partners and develop it along with the ecosystem necessary to make the product take off in the market. The emergence of the NUC enabled vendors like ASUS, MSI, ASRock, and Zotac to create and widely market their own UCFF systems.

But if we're to repeat that here, then in the context of the Compute Element initiative, what role would such vendors have? We have already seen ASUS create an add-in card specifically catering to the NUC 9 Extreme Kits. GIGABYTE and MSI apparently have similar GPU cards in the pipeline. Many chassis vendors have also signed up to create Compute Element-compatible cases. However, it remains to be seen whether board and system vendors like ASRock and Zotac plan to create their own Compute Element-like products and whether they would be able to take advantage of the ecosystem that Intel is developing. As an example, the current Compute Elements don't have a NBASE-T port. It could be interesting if Intel allows its partners to create their own Compute Element with a NBASE-T port, or, say, a USB 3.1 Gen 2x2 port. When Intel shifts to NBASE-T in their own Compute Element lineup, Intel's partners could offer 10GBASE-T or additional Thunderbolt ports. Or, to dream boldy, perhaps an AMD Renoir-based Compute Element in the near term from these vendors?

Overall, the great performance profile of the NUC9i9QNX is only a small part of the equation. The NUC demonstrates Intel's vision for the bulk of the desktop PC market moving forward, albeit in a product that's premium in everything from performance to build to pricing. The latter of those suits Intel for now, but it is almost certainly leaving a much larger market unserved.

Currently, the lowest-priced Ghost Canyon board is the $664 NUC9i5QNB, while the $1553 NUC9i9QNX we looked at today uses the $1274 NUC9i9QNB board. We can totally imagine a user buying a Ghost Canyon kit chassis with a lower performance Compute Element (at, say, $300 to $500) and moving to a higher performance Compute Element a year or two down the line. In that context, we believe Intel (or its partners) should start catering to a wider range of price points. Assuming that Intel can build upon upon its initial success with the Compute Element initiative, the future of the desktop PC market does look bright.