Re:NVIDIA introduces TensorRT-LLM for accelerating LLM infer...
by vladgur on Friday September 08, 2023 @06:14PM
(#37440010)
Is there anything special about H100/A100 GPUs that would make this acceleration work on them and not work on 30x0/40x0 GPUs
[ Parent
| on HN ]
Re:NVIDIA introduces TensorRT-LLM for accelerating LLM infer...
by minimaxir on Friday September 08, 2023 @06:30PM
(#37440123)
Normally it's architectural changes such as FP8 support as noted in the submission, but in that case newer mid-range GPUs like the A10 and the L4 (https://www.nvidia.com/en-us/data-center/l4/) should be supported but aren't.
Re:NVIDIA introduces TensorRT-LLM for accelerating LLM infer...
by nkingsy on Friday September 08, 2023 @06:22PM
(#37440065)
https://github.com/NVIDIA/TensorRT/issues/982
Looks like tensorRT does work with 30x0, but I couldn't find much.
Re:NVIDIA introduces TensorRT-LLM for accelerating LLM infer...
by imjonse on Friday September 08, 2023 @06:29PM
(#37440121)
From the article I understand it's only for H100? Both the title and the charts imply that.
Re:NVIDIA introduces TensorRT-LLM for accelerating LLM infer...
by why_only_15 on Friday September 08, 2023 @06:55PM
(#37440306)
NVLink is the most notable (>600GB/s GPU-to-GPU communication) whereas the best case scenario on 30x0/40x0 GPUs is peer-to-peer PCIe (~30GB/s) and average case is routing PCIe through the host (~15GB/s)
Re:NVIDIA introduces TensorRT-LLM for accelerating LLM infer...
by boxfire on Friday September 08, 2023 @06:59PM
(#37440337)
3090 does 4x nvlink each way, ~100 GB/s bidirectional data
Re:NVIDIA introduces TensorRT-LLM for accelerating LLM infer...
by why_only_15 on Friday September 08, 2023 @07:03PM
(#37440364)
Oh huh I didn't know about that!
Re:NVIDIA introduces TensorRT-LLM for accelerating LLM infer...
by ban-lan-gen on Saturday September 09, 2023 @09:14AM
(#37444986)
They just want to sell some expensive hardwares?