Server: Google is among the CPU designers

There's a new player getting involved in server processors: Google announced Axion, the first multi-core processor on the ARM architecture. Google calls the Axion “the first ARM-based CPU” — and it appears the company doesn't consider its Pixel smartphones' Tensor G chipset to be a CPU, but merely a system-on-a-chip (SoC).

advertisement

The Axion uses ARM's Neoverse V2 cores, so it's not entirely new V3 variants. Google has so far remained mum on the number of CPU cores, clock frequencies, cache sizes, and chip manufacturer, but is touting raw comparison values. These values ​​are currently unverifiable, so they can best represent the given values:

“Axion processors combine Google's silicon expertise with Arm's most powerful CPU cores to provide instances with up to 30 percent better performance than the fastest general-purpose Arm-based instances available in the cloud today and up to 50 percent better performance.” Up to 60 percent better power efficiency than current x86-based instances.”

Just wait and see

So far, Google is only comparing per cloud instance, which means performance per CPU core is mainly relevant, but not count. In a best-case scenario, your Axion instance should be up to 30 percent faster than Ampere and perhaps Amazon's Graviton4 instances. When comparing x86, Google only has to compare itself to older generations of AMD and Intel.

In Intel's case, these are Sapphire Rapids Xeons (max. 60 cores), as there are no public cloud instances with the current generation Emerald Rapids. Additionally, Intel's 288-core Sierra Forest can reallocate cards – this processor is specifically designed to keep multiple ARM cloud data center cores in check.

New 128-core Zen 5 (Torino) Epycs are coming from AMD this year. The company is also planning a Zen 5c branch with 192 cores. Perhaps the comparisons no longer look so rosy.

With Axion, Google wants to cover all types of workloads: from web and application servers to data analysis to CPU-based AI training and inference, that is, implementing already trained AI algorithms. To do this, ARM integrates appropriate data types such as Bfloat16 and INT8 as well as vector units into its Neoverse architecture. ARM instances with Axion processors should be available by the end of the year.

(MMA)