Cerebras AI Day - Hardware Keynote - Sean Lie
HTML-код
- Опубликовано: 7 сен 2024
- CTO and Co-Founder Sean Lie takes a deep dive into the CS-3's architecture. Sean describes how the Cerebras CS-3 delivers 2x performance at the same power and price as the CS-2. Sean also describes how Cerebras builds clusters that can train trillion parameter models at record speeds while operating with the simplicity of a single device. Finally, Sean describes how Cerebras can do even better to accelerate AI model development through innovations like sparsity.
The performance comparison with H100 isnt apples to apples. CS3 has a much larger area. Actually the ratio of CS3 area and H100 area is approx 57. So, 57 H100s with perfect scaling would give the exact same performance. It would make more sense to compare against an Nvidia cluster of GPUs with approx equal area as CS3.
Impressive technological achievement. Can't wait to try these 24T-parameter models!
if you can mix this with northpole from ibm
looks like I can't play solitaire on this hardware
going to war with nvidia