@rajaxg.bsky.social
o 1 ExaFlop of raw FP8/INT8 compute performance
o 5 PB/s of "HBM" bandwidth at 138 TB Capacity
o 2.5 PB/s of GPU-GPU bandwidth
o All while maintaining a 132 KW power envelope
o At $3M price
x.com/RajaXg/statu...
#HPC #AI
@rajaxg.bsky.social
o 1 ExaFlop of raw FP8/INT8 compute performance
o 5 PB/s of "HBM" bandwidth at 138 TB Capacity
o 2.5 PB/s of GPU-GPU bandwidth
o All while maintaining a 132 KW power envelope
o At $3M price
x.com/RajaXg/statu...
#HPC #AI
github.com/electronicar...
github.com/electronicar...
Have 3 VS Code Windows open each SSH connected to a remote system
First system is 8 Intel PVC 1550 GPU
Second one is 8 MI300X
Third one is 8 H100s
All three running my torchure tests in parallel:)
Have 3 VS Code Windows open each SSH connected to a remote system
First system is 8 Intel PVC 1550 GPU
Second one is 8 MI300X
Third one is 8 H100s
All three running my torchure tests in parallel:)
Discussion: huggingface.co/papers/2411....
Project page: research.nvidia.com/labs/dir/edi...
Arxiv: arxiv.org/abs/2411.07126
Discussion: huggingface.co/papers/2411....
Project page: research.nvidia.com/labs/dir/edi...
Arxiv: arxiv.org/abs/2411.07126