MLPerf Training 5.0 Results are out with a few new architectures. This edition was still mostly NVIDIA dominated, which makes sense given NVIDIA’s market share. Google had a Trillium system submitted. AMD had some llama runs using the Instinct MI300X and MI325X, which is exciting.
MLPerf Training v5.0 is Out
We often call MLPerf NVIDIA’s MLPerf because they have dominated since the benchmark suite was created and heavily influence the vast majority of submissions across benchmarks. When you see a Dell, Supermicro, Gigabyte, or other submission, usually NVIDIA is doing a lot of the legwork behind the scenes to ensure the models run well on their hardware, then the OEMs tune their systems. As a result of all the work over the years, NVIDIA submits on all workloads and wins in many simply because there is no competition.
NVIDIA, as one might expect, was eager to share that its new Blackwell architecture was faster than Hopper. Most NVIDIA results were paired with Intel Xeon CPUs or NVIDIA Grace Arm CPUs.
The company also shared a brief with its roadmap to Gigawatt AI factories. This comes the day after Broadcom Tomahawk 6 was Launched for 1.6TbE Generation but NVIDIA is showing Specturm6 CPO 102.4T as being a 2026 generation product so Broadcom will lead in the Ethernet switching market for some time.
On the subject of AI Factories, we recently toured the Dell factory where it is building huge numbers of NVIDIA GB200 NVL72 systems for its US customers. Certainly check that one out here:
AMD for its part showed generational performance gains. On the Llama2 70B LoRA benchmark it showed 30% generational uplift for the AMD Instinct MI325X over the MI300X.
AMD also was able to show performance better than the NVIDIA H200 systems which are contemporaries of the MI325X. We saw the MI350/MI355 generation at Computex 2025, and AMD has an upcoming AI event, but MLPerf Training results are published twice per year. Just to note, it does not appear as these charts are on a zero Y-axis scale.
AMD also managed to help some of its partners submit as well.
Final Words
Browsing the results, there were 2-3 older model desktop GPUs and the Google Trillium, but there was not a lot here outside of NVIDIA across benchmarks and AMD focused on one benchmark. I would not say this was an overwhelming set of results, but it is a good time to take a moment and note that the hardware is getting faster while software optimization is happening at the same time.
