Why Hardware-Software Co-Design Is AI's Real 100x: Dylan Patel of SemiAnalysis
- Source
- youtube.com
- Author
- Sequoia Capital
- Date
Context
In a Sequoia interview, Dylan Patel of SemiAnalysis, a semiconductor research firm, argues that AI efficiency gains can come from three layers: chip hardware, low-level software such as kernels (the small programs that run the math on a chip), and model design. Patel says improving each layer might give about 2x, which would multiply to 8x, but designing all three together can give about 100x. That figure is Patel's illustration, not a measurement.
Patel's main example is DeepSeek. Patel says the shapes of its expert sub-networks were optimized for Nvidia's Hopper chip in V3 and for Blackwell and Huawei's chip in V4. Google's TPUs, Patel adds, are an amazing chip yet run DeepSeek poorly while running other kinds of models very well. The conclusion is that no chip is better in isolation: whether it is the right choice depends on where a lab's models are headed.
videoBringing Continual Learning to Every Company: Trajectory’s Ronak Malde
videoCrusoe CEO Chase Lochmiller: Turning Stranded Energy Into Intelligence
videoDavid Silver and Lasse Espeholt of Ineffable Intelligence: A Superlearner for the Age of Experience
videoCracking the Code on Offensive Security With AI ft XBOW CEO and GitHub Copilot Creator Oege de Moor
Checking sign-in…
Loading comments…


