this post was submitted on 08 Nov 2023
1 points (100.0% liked)

Hardware

33 readers
1 users here now

A place for quality hardware news, reviews, and intelligent discussion.

founded 11 months ago
MODERATORS
[email protected]
1
Nvidia, Intel claim new LLM training speed records in new MLPerf 3.1 benchmark (venturebeat.com)
submitted 10 months ago by [email protected] to c/[email protected]
6 comments fedilink hide all child comments
top 6 comments
sorted by: hot top controversial new old
[–] [email protected] 1 points 10 months ago

“NVIDIA was once again the only company to run all MLPerf tests. H100 GPUs demonstrated the fastest performance and the greatest scaling in each of the nine benchmarks.”

[–] [email protected] 0 points 10 months ago (2 children)

Among many new records and milestones, one in generative AI stands out: NVIDIA Eos — an AI supercomputer powered by a whopping 10,752 NVIDIA H100 Tensor Core GPUs and NVIDIA Quantum-2 InfiniBand networking — completed a training benchmark based on a GPT-3 model with 175 billion parameters trained on one billion tokens in just 3.9 minutes.

[–] [email protected] 1 points 10 months ago

Wow I wonder how long that would have taken in Turing/Volta era hardware.

[–] [email protected] 1 points 10 months ago

What does it mean to train on such-and-such number of tokens?

[–] [email protected] 0 points 10 months ago (1 children)

How much of this performance is just the usual cheap tricks like software optimizations and task specific hardware that doesn't generalize? If it was an honest test of pure hardware muscle the AMD MI300 would come out on top.

[–] [email protected] 1 points 10 months ago

The benchmarks test workloads people are actually interested in running.

Whether good performance is achieved by good hardware, a good software stack, or both doesn't matter to the vast majority of buyers.