Ampere Computing announced its latest generation cloud data center processor “AmpereOne”, which is the industry’s first 192-core general-purpose CPU processor. It is manufactured by TSMC 5nm process, based on Armv8.6+ instruction set self-developed architecture, with 136, 144, 160, 172, 192 cores (previous generation Ampere Altra 32-128 cores) and stable frequency up to 3.0GHz.
Two integrated 128-bit vector units per core, supporting FP16, BF16, INT16, INT8 data formats.
Each core has 16KB Level 1 instruction cache, 64KB Level 1 data cache, 2MB Level 2 cache (1MB in the previous generation), while all cores share 64MB Level 3 cache.
Memory supports eight channels of DDR5 ECC (previous generation eight channels of DDR4), up to 16 strips in a single system, with a maximum capacity of 8TB.
Expansion support 128 PCIe 5.0 (previous generation 128 PCIe 4.0) from 32 controllers.
Power consumption range of 200-350W, compared to the previous generation of 10-180W increased a lot.
Performance-wise, Ampere compares AMD and Intel x86 platforms, but is a bit of an alternative.
One is to compare the number of virtual machines supported, 42U 16.5 kW rack space, AmpereOne can be up to 7296, AMD Skyline 9654 Genoa, Intel Xeon 8480 + Sapphire Rapids are only 2496, 1680.
But in fact, 42U rack power space has opened up a lot, many have increased to 20-50 kW, which is certainly more favorable for AMD, Intel, but was deliberately avoided by Ampere.
The second is to compare AI performance, Stable Diffusion, DLRM two projects, claimed to be 2.3 times and 2 times ahead compared to AMD Skylar 9654.
But the fact that Ampere uses FP16 precision and AMD is FP32 precision is not fair, not to mention that most of the FP16 precision load has been handed over to the GPU for execution.
