Despite the ARM architecture, it must be said that, as always, while ARM can outperform x86 in low-power, high-efficiency scenarios, at the moment it is unable to scale it to high-performance scenarios. This is in fact one of the reasons why chips Apple A15 They have become a (relative) disappointment for now, so in a server environment where maximum power is always sought, in theory ARM has nothing to do … or does it? That is what NVIDIA thinks and defends.
A server with NVIDIA A100 and ARM CPU beats one with x86
As you can see in the graph above, the server equipped with ARM CPU is practically at the same level as the one with an x86 CPU, and in fact manages to surpass it by a lot in the 3D-Unet niche, while in the most common ones such as the low workloads ResNet 50 they are still dominated by x86… albeit by very little difference actually.
Obviously, when we talk about inference, a CPU can never exceed the performance of a GPU regardless of its architecture. For this reason, NVIDIA has not cut a hair when stating that its A100 ARM GPU is up to 104 times faster than a CPU in the MLPerf benchmarks.
«Inference is what happens when a computer runs Artificial Intelligence software to try to recognize an object or make a prediction. This process uses a Deep Learning model to filter the data and find the results that no human being would be able to perform. MLPerf inference benchmarks use today’s most popular AI workloads and scenarios (spanning niches such as medical imaging, language processing, etc.). ” – said David Lecomber, director of HPC and tools at ARM.
Of course, we are talking about a comparison that is not too fair (a GPU against a CPU) since each element has been designed with a defined purpose, but also of course NVIDIA, a staunch defender of the ARM architecture, seems to be looking for any excuse to promote what interests them most at this time. For this reason, we can see in the graphic above how the NVIDIA A100 ARM GPU reigns supreme in everything they have tested with it, from the popular ResNet-50 image classification (AI) benchmark to natural language processing.
How will you know if you have followed the NVIDIA-ARM soap opera In recent times, the company led by Jen-Hsun Huang is still facing regulatory obstacles that are preventing its purchase, so the company is beginning to press for other directions, which, as in this case, also include the ecosystem of the servers.
In any case, and although this is not something that is going to happen in the short term, what does seem to have foundation is that the reign of x86 architecture on servers is beginning to be threatened by ARM.