AMD's April 2, 2025 MLPerf Inference v5.0 post reports the MI325X Llama 2 70B run
AMD's April 2, 2025 MLPerf Inference v5.0 post reports the MI325X Llama 2 70B run, but it leaves the November price unmeasured.
| AMD Instinct MI325X | |
|---|---|
| memory capacity | 256GB HBM3E |
| memory bandwidth | 6TB/s |
| GPUs per platform | 8 discrete GPUs |
| prefill batch size | 65K |
| decode batch size | 2048 |
| prefill compute | 1.5K TFLOPs |
| decode compute | 1.4K TFLOPs |
| frequency restriction | 1700 MHz |
The post describes an 8-GPU MI325X platform with an EPYC CPU and 256GB of HBM3E per GPU.
It reports Llama 2 70B inference at 1.5K TFLOPs prefill and 1.4K TFLOPs decode.
The 6TB/s bandwidth figure appears alongside the missing QPS, TTFT, and TPOT figures.
For the fab, the 1700 MHz frequency cap matters more than the headline memory size, because it sets the clock limit for the OAM package.
The exact MLPerf score, server power draw, EPYC model, server model, and any November price remain unmeasured.
A buyer cannot price the full 8-GPU node from this page.
The page supports the benchmark and clock cap, not a November price.
After rocm.blogs.amd.com. We did not report this. The pictures, if any, are theirs.
