Cincoze GP-3000 review – An expandable Xeon-based GPU computer

Cincoze GP-3000 is an expandable high-performance GPU computer. It consists of either a 9th or 8th generation Intel processor-powered embedded computer which can be expanded with Cincoze’s proprietary GPU Expansion Box (GEB) capable of housing up to dual 250W full-length graphic cards. Additional I/O expansion is also possible through the use of various modules. With a total 720W power budget the GP-3000’s additional GPU performance massively accelerates complex industrial AI and machine vision tasks.

As a rugged computer, the GP-3000 has passed a series of stringent quality assurance tests and industry standards including MIL-STD-810G military standard, E-mark for in-vehicle applications, and EN50155 (EN 50121-3-2 only) for rolling stock environments. It can withstand hot and cold temperature extremes, shock and vibration, and high electromagnetic radiation.

In this review, I’ll cover some performance metrics from both Windows and Ubuntu and also discuss the thermals.

Hardware Overview

The GP-3000 (see PDF datasheet) loaned by Cincoze for review came installed with a Xeon E-2278GE which is an eight-core 16-thread 3.30 GHz Coffee Lake processor boosting to 4.70 GHz with Intel’s UHD Graphics 630 together with 64GB of ECC RAM and a 512GB NVMe drive. It also came with the larger of the two GPU Expansion Box options, the GEB-3601 (see PDF datasheet), which provides two PCIe x16 (electrically wired as PCIe x8), one PCIe x4, and one PCIe x1 slots together with fans for cooling. Installed into the ‘GEB’ was an NVIDIA RTX 3090 Founders Edition graphics card:

The standard I/O accessed from the front of the device includes HDMI, DisplayPort, VGA, five GbE ports, two RS-232/422/485 COM ports, two USB 3.2 Gen 2 (Type A) ports, four USB 3.2 Gen 1 (Type A) ports, a headphone, and a microphone jack. The optional Cincoze’s proprietary CMI/CFM  (Combined Multiple IO  & Control Function Module) modules can provide additional I/O such as eight GbE LAN/POE ports, six USB 3.2 ports, and two 10 GbE ports. The GP-3000 also supports four 2.5″ hot-swap SATA SSDs and includes one M.2 (M-key) NVMe SSD for storage.

The CPU is cooled with a giant heat sink which runs the full width of the computer and has twin fans on either end:

The heat sink then uses a thermal pad which covers a copper plate under which the CPU is installed:

Power is provided courtesy of a Mean Well HEP-1000-24 1000W power supply:

All together this is a full-blown server measuring 305 x 195 x 370 mm (12.00 x 7.68 x 14.57 inches) and weighing in at approximately 20kg (GP-3000: 8kg, GEB-3601: 5.2Kg, RTX 3090: 2.1kg and HEP-1000-24: 4kg) and emits around 50dBA from a number of cooling fans.

Review Methodology

I decided to review using a dual-boot of Windows 10 Enterprise version 21H1 and Ubuntu 20.04 LTS point release 2 and test with a selection of commonly used Windows benchmarks and/or equivalents for Linux together with Thomas Kaiser’s ‘sbc-bench’ which is a small set of different CPU performance tests focusing on server performance when run on Ubuntu. Additionally, I used ‘Phoronix Test Suite’ to benchmark the same set of tests on both Windows and Ubuntu for comparison purposes. On Ubuntu, I also compile the v5.4 Linux kernel using the default config as a test of performance.

Prior to benchmarking, I perform all necessary installations and updates. I also capture some basic details of the device for each OS.

Windows 10 Performance on GP-3000

I first installed a Windows 10 Enterprise Evaluation version 21H1 ISO and updated it to OS build 19043.1110. A quick look at the hardware information shows:

I then set the power plan to ‘Ultimate Performance’ and ran the same benchmarking tools I ran when reviewing Cincoze’s GM-1000 model to look at performance under Windows:

I also ran some additional benchmarks to highlight the power of the GP-3000 including

(Note: the Blender ‘classroom’ benchmark is for CUDA on the RTX 3090)

As the GPU is limited to PCIe Gen 3 x8 I also measured the transfer speeds using CUDA-Z:

To ‘push’ the GPU I tried viewing an 8Kp60FPS video in YouTube on Edge full-screen on a 2K monitor and this resulted in 12 frames being initially dropped followed by the occasional frame drop as the video played:

Ubuntu Performance

After shrinking the Windows partition in half and creating a new partition I installed Ubuntu using an Ubuntu 20.04.2.0 ISO as dual boot. After installation and updates, the key hardware information is as follows:

I then set the CPU Scaling Governor to ‘performance’ and again ran the same benchmarking tools as for Cincoze’s GM-1000 to look at performance under Ubuntu:

For additional benchmarks I ran:

(Note: the Blender ‘classroom’ benchmark is for CUDA on the RTX 3090)

CUDA-Z provided similar results to Windows:

However, when pushing the GPU by viewing the same 8K @ 60FPS video in YouTube but on Firefox full-screen on a 2K monitor the playback was not as successful compared with Windows as it occasionally stalled and continuously dropped frames. Chrome was even worse with over twice the number of dropped frames compared to Firefox:

Thermals

The design of the GP-3000 and expansion box specifically addresses cooling both the CPU and GPU which results in very good thermal management.

When running a stress test on Ubuntu whilst the ambient room temperature was around 14.6°C, the CPU temperatures peaked at 55°C:

and as soon as the test finished the temperature dropped:

For GPU temperatures I ran Unigine’s Heaven both in Windows where the GPU temperature peaked at 65°C and in Ubuntu which peaked at 57°C:

Windows vs Ubuntu

Whilst a detailed comparison between the two operating systems is beyond the scope of this review, looking at the performance tools common between the two systems showed similar results and this can be visually shown by comparing the Phoronix benchmarks for each OS:

Interestingly, when running Geekbench performance in Windows the single-core score was 1264 and the multi-core score was 7204 whereas in Ubuntu the single-core score increased to 1318 but the multi-core score reduced to 7061. Monitoring the benchmark when running and logging the CPU clock speed and temperature shows that when running the multi-core benchmark in Ubuntu the CPU throttling was more significant than for Windows:

Graphical performance was also sometimes worse in Ubuntu than Windows but typically this was due to different renders being used. For example, when Blender uses CUDA the results are identical at 18 seconds for both Windows and Ubuntu. However, when running Unigine’s Heaven Ubuntu uses OpenGL which is slower (203.7 FPS) compared to Windows using DirectX 11 (344.4 FPS). As a result, the GPU is less utilized and therefore runs cooler as shown above.

Additional GPU

The GEB-3601 GPU Expansion Box has a total power budget of 500W and is capable of housing either a single 300W 350W or two 250W full-length graphic cards. The installed RTX 3090 is a 350W GPU and it works well based on Cincoze’s internal testing due to the effective thermal system design. Nevertheless, for testing purposes, I erred on the side of caution and additionally installed just a 100W GTX 1650 Super which automatically detected:

CUDA-Z showed a similar transfer speed for the GTX 1650 Super as for the RTX 3090:

and the card could be specifically chosen when running applications such as Geekbench:

GP-3000 networking performance

Network connectivity throughput was measured for one of the 1GB Ethernet ports on Ubuntu using ‘iperf’. Upload was measured at 942 Mbits/sec and download at 839 Mbits/sec.

Power consumption

Power consumption was measured as follows:

• Powered off (shutdown) – 12W (Windows and Ubuntu)
• BIOS/GRUB – 90W
• Idle – 48W (Windows and Ubuntu)
• CPU intensive – 111W (Windows ‘Cinebench’) and 108W (Ubuntu ‘stress’)
• GPU intensive – 420W (Windows ‘Blender’) and 430W (Ubuntu ‘Blender’)

The power figures fluctuate due to the fans so the values are ‘rounded’ averages.

Final Observations

The GP-3000 performance is impressive as is the configurability, the expansion capabilities, and the high-quality documentation. The provision of a dedicated thermally controlled GPU housing which isolates the heat from the CPU and GPU is both innovative and ideally suited to the target market of running industrial AI and machine vision applications. Further information is available on Cincoze’s website.

Ian W MORRISON (Linuxium)

Ian is interested in mini PCs and helps with reviews of mini PCs running Windows, Ubuntu and other Linux operating systems. You can follow him on Facebook or Twitter.

Support CNX Software! Donate via cryptocurrencies, become a Patron on Patreon, or purchase goods on Amazon or Aliexpress