In my view, AMD will eventually offer a full AI development platform. Chinese software firm DeepSeek has already shown that skilled engineers can write low-level optimizations to make traditionally “weaker” NVIDIA GPUs train AI models far more efficiently. I expect those same techniques to be applied to AMD hardware. In fact, George Hotz’s Tinygrad project is already hard at work commoditizing petaflop-scale training on GPUs—proof that open, high-performance AI stacks on AMD are within reach.
But is that performance truly unlocked with AMD’s current software stack? I wonder how technically feasible it is to be able to toe the line between custom and generalized performance for all types of models
Hard to know, as AMD hasn’t gone into specifics. However, the software stack wouldn’t be particularly problematic in this case. It’s more a matter of optimizing the chip for specific workloads. If Meta and Amazon can do it with the help of Broadcom and Marvell, AMD can too. Their benchmarks are already very impressive, and their success in optimizing for ChatGPT and LLaMA workloads proves their capabilities. Truly, the only thing missing is a solid rack-scale system. They've already signaled their intent to diversify their AI offerings, and I’m sure FPGAs will play a role in that.
From Gemini: "This remains the biggest hurdle. Even if AMD's hardware matches or surpasses Nvidia's in raw performance, the established developer ecosystem around CUDA means many AI developers are "locked in" to Nvidia. Switching requires effort and risk, which many aren't willing to take unless the performance/cost advantage is overwhelmingly in AMD's favor. Nvidia isn't standing still. They are continuously innovating and releasing new, more powerful architectures (like the upcoming Blackwell) that maintain their performance lead."
One of the only really good articles on AMD that I have read on Substack. Good work. Also note that AMD got TSMC fab capacity for the first 2nm node thanks to AMD’s chiplet strategy. Launching a monolithic chip at 2nm is a huge risk for TSMC. Don’t under estimate the value of AMD’s good relationship with TSMC. Chiplets are far less risky for TSMC to manufacture…this will become increasingly important for the company.
Nvidia won’t be beaten but AMD looks like a great catch up trade
I remember reading your deep dive on AMD a while back which was crystal clear and so visionary. Congrats Daniel, I tip my hat to you!
In my view, AMD will eventually offer a full AI development platform. Chinese software firm DeepSeek has already shown that skilled engineers can write low-level optimizations to make traditionally “weaker” NVIDIA GPUs train AI models far more efficiently. I expect those same techniques to be applied to AMD hardware. In fact, George Hotz’s Tinygrad project is already hard at work commoditizing petaflop-scale training on GPUs—proof that open, high-performance AI stacks on AMD are within reach.
How will FPGAs improve inference performance?
They allow customization beyond what standard GPUs can offer
But is that performance truly unlocked with AMD’s current software stack? I wonder how technically feasible it is to be able to toe the line between custom and generalized performance for all types of models
Hard to know, as AMD hasn’t gone into specifics. However, the software stack wouldn’t be particularly problematic in this case. It’s more a matter of optimizing the chip for specific workloads. If Meta and Amazon can do it with the help of Broadcom and Marvell, AMD can too. Their benchmarks are already very impressive, and their success in optimizing for ChatGPT and LLaMA workloads proves their capabilities. Truly, the only thing missing is a solid rack-scale system. They've already signaled their intent to diversify their AI offerings, and I’m sure FPGAs will play a role in that.
From Gemini: "This remains the biggest hurdle. Even if AMD's hardware matches or surpasses Nvidia's in raw performance, the established developer ecosystem around CUDA means many AI developers are "locked in" to Nvidia. Switching requires effort and risk, which many aren't willing to take unless the performance/cost advantage is overwhelmingly in AMD's favor. Nvidia isn't standing still. They are continuously innovating and releasing new, more powerful architectures (like the upcoming Blackwell) that maintain their performance lead."
One of the only really good articles on AMD that I have read on Substack. Good work. Also note that AMD got TSMC fab capacity for the first 2nm node thanks to AMD’s chiplet strategy. Launching a monolithic chip at 2nm is a huge risk for TSMC. Don’t under estimate the value of AMD’s good relationship with TSMC. Chiplets are far less risky for TSMC to manufacture…this will become increasingly important for the company.
Hi I enjoyed the article but where did you see the insider buys? I couldn’t find any source on that.
It's referring to Exec Phil Guido, who started buying more shares this year for the first time in 12 years.
Thanks for the reply! Found it now.
Great Analysis.
Thanks!