NVIDIA’s Grace Hopper GH200 chip is the company’s response to the demands of advanced artificial intelligence and high-performance data processing in enterprises and data centers. Announcing this specialized processor last year, NVIDIA CEO Jensen Huang said that the technology industry is hitting a wall with traditional architectures, so it is increasingly turning to GPUs and accelerated computing to solve complex computing problems.
Despite the fact that NVIDIA’s flagship chip is designed for data centers and AI, GPTshop.ai sells the GH200 as part of a desktop workstation. Now anyone can get access to the NVIDIA Grace processor and Hopper GPU, provided they have at least $50,000 to buy a base system on the GH200. The benchmarking website Phoronix has a review of the computer for comparative analysis.
A desktop workstation based on the GH200 platform looks completely overkill on paper. The Grace GH200 processor has 72 cores and 480GB of LPDDR5X memory, which is not as much as the 1TB that can be installed on the Threadripper 7000, but it’s still a solid amount. The real star of the show is the Hopper-based H200 graphics module, which measures 814 mm² and has 16,896 CUDA cores. Interestingly, GPTshop.ai offers both the HBM3 version of the GH200 and the new HBM3e version, which includes more VRAM.
The system is also equipped with Noctua fans and cooling system. It has additional components such as an 8TB SSD, 30TB HDD, mouse, and keyboard. There’s even an RTX 4060 graphics card, presumably for graphics output or access to ray tracing, which is one of the few features Hopper doesn’t support. The only option for an OS is an Ubuntu server; Windows is not a great operating system for the latest processor.
Based on tests, this is NVIDIA’s most powerful AI module to date, designed for giant-scale AI, or high-performance computing applications. The company claims that it delivers up to ten times faster performance for applications that work with terabytes of data, and that this power boost allows scientists and researchers to create unprecedented solutions to complex problems. From a corporate perspective, the processor serves as an alternative to x86 processors in the server segment, allowing NVIDIA to provide a competitive product against AMD and Intel offerings. NVIDIA said that the Grace processor delivers twice the performance of Intel Sapphire Rapids and AMD Genoa processors, with the same power and 3.5 times the efficiency of the latest generation of AMD Epyc Milan processors.
According to the tests, the GH200 is slightly ahead of the Xeon Platinum 8380 2P and slightly behind the performance of the Epyc 9654 Genoa 2P, which Phoronix declared “not bad at all for an initial display.” The GH200 was also found to be the fastest compared to single processor Genoa(X) 2P configurations and Intel Emerald Rapids. The GH200 had almost twice the performance of the 128-core Arm Ampere Altra Max processor. The GH200 was equal to the AMD Epyc 9684X Genoa-X 1P processor for the Rodinia HPC test in the LavaMD test case. For AMG, the superchip system almost matched the Xeon Platinum 8380 2P in terms of CPU performance. Phoronix called the CPU performance “very exceptional” in the NWChem computational chemistry benchmark. A single GH200 almost matched the AMD EPYC Genoa in the leading 2 processor configuration.
Meanwhile, we can’t be sure how well the H200 half of the device will perform, as the Phoronix review did not test the GPU. However, the chip’s official specs strongly suggest that it is very powerful, which is not surprising since there is so much demand for the H200 (and almost all AI-enabled GPUs these days, for that matter).
Although $50,000 is a lot of money for a desktop workstation, it seems like a decent deal considering that the H100, the predecessor to the H200, costs $40,000 separately. For $50,000, getting a faster GPU and a top-of-the-line server CPU inside a pre-assembled system sounds like a pretty good deal.
Loading comments …