Nvidia believes that its Grace superchip will annihilate its competitors’ efforts. Up to two times the performance of Intel’s Ice Lake Xeon processors are possible.
Recently, NVIDIA unveiled its Grace CPU at the GPU Technology Conference in San Francisco. The Grace CPU incorporates 144 Arm cores into a platform that NVIDIA believes is powerful enough to completely disrupt the traditional server market.
The Nvidia Grace CPU, according to the company’s initial presentations, could deliver 50 percent more performance in the SPEC benchmark than two 64-core AMD EPYC processors while consuming half the amount of power. Despite the fact that the provided result was likely a highly optimized benchmark, it demonstrated that Nvidia is not playing around, and it provides a great deal of context to Nvidia’s attempts to purchase Arm outright.
Given the failure of that attempt, it has opened the door for others to make an attempt to purchase Arm. Nvidia claims that when the Grace CPU ships in early 2023, it will be the fastest server processor on the market, according to the company. Make sweeping and audacious claims! In any case, given Nvidia’s bullish momentum, who are we to disagree? It won’t happen, at least not until the ships have been independently tested in a wide range of scenarios.
Considering the recent announcement of the Grace processor, Tom’s Hardware did some digging and discovered yet another result, which simulates the performance of the Grace CPU compared to Intel’s Ice Lake Xeon platform. The likelihood that the outcome is accurate, or even mostly accurate, is that Intel and AMD will face significant competition.
When Ian Buck, vice president of Nvidia’s Accelerated Computing business unit, gave a presentation, Tom’s discovered a benchmark that compares Grace to Intel’s Ice Lake, which he used to make his discovery. According to the Weather Research and Forecasting (WRF) model, Grace is twice as fast and 2.3 times more energy-efficient than Intel’s Ice Lake processors.
With two Xeon 8360Y processing units, the Intel platform could handle up to 72 cores and 144 threads at the same time, according to Intel. According to Nvidia, the platform was drawing 572W in a single-node configuration.
Of course, we must exercise extreme caution when relying on vendor-provided benchmarks, which are frequently cherry-picked to highlight the best features of a particular platform. However, if these benchmarks are accurate and can be replicated, Nvidia appears to be in a strong position to capture a significant portion of the lucrative high-performance computing market.
Nvidia’s Grace CPU Superchip is an Arm v9-based processor with 144 cores spread across two dies that are connected by a 900 GB/s NVlink. The processor is based on Nvidia’s CUDA architecture. The system is capable of providing memory bandwidth of up to 1TB/s. That’s a significant amount of data, which explains why Nvidia chose to demonstrate benchmarks that were bandwidth constrained in order to save time.
The bottom line is that Nvidia is well positioned to compete with Grace in the enterprise market in the years to come. This holds true for Arm in general. The real competition, on the other hand, is not with the current generation, but with the next. Grace’s primary competitors will be Intel’s Sapphire Rapids and AMD’s Genoa.
The real question is whether or not future Arm processors will be capable of competing as gaming processors. If Arm processors make significant inroads into the high-performance computing market, it’s possible that the gaming market will follow suit