To be fair, having seen his software evolve, and having seen ROCm evolve, I'm more optimistic for his software in a year than yours.
He picked his problem better. The whole reason that tinygrad is, well, tiny, is that it limits the amount of overhead to onboard people and perform maintenance and rewrites. My strong impression is that the ROCm codebase is simply much too large for AMD's dev resources. You're trying to race NVidia on their turf with less resources. It's brave, but foolish.
I can see how Tinygrad could succeed. The story makes sense. AMD's doesn't, neither logically nor empirically. NVidia would have to seriously fumble.
That said I'm deeply worried about anyone whose based their company on amd gpus. The only reason why they do well in hpc is because there's an army of dreadfully underpaid and over performing grand students to pick up the slack from AMD. Trying to do that in a corporate environment is company suicide.
> That said I'm deeply worried about anyone whose based their company on amd gpus
Sony Interactive and Microsoft XBox seem to be doing great without an army of underpaid students. AMD does great at the top and bottom: the corporates in the middle that are unwilling or unable to pay people to author/tweak their software for AMD GPUs will do better going with Nvidia, which has great OOTB software, and a premium to go with it.
I suppose if AMD had infinite resources, it'd fix this post-haste.
Sony and Microsoft don't need to interact with AMD's software, they just use the hardware. Just see that Sony made its own upscaling software via ML and even created hardware with several pieces of RDNA architectures, so much so that they called it RDNA 2X (RDNA 2 (compute) + WMMA from RDNA 3 and RT from RDNA 4).
This is false. 3D VCache is enabled by TSMC's 3DFabric packaging. It also didn't really play a role in AMD passing Intel. Chiplets are also enabled by TSMC technology, CoWoS.
When AMD passed Intel, they hadn't even decided to use TSMC at all yet. Of course now Intel is behind in leveraging TSMC technology. They started late.
He picked his problem better. The whole reason that tinygrad is, well, tiny, is that it limits the amount of overhead to onboard people and perform maintenance and rewrites. My strong impression is that the ROCm codebase is simply much too large for AMD's dev resources. You're trying to race NVidia on their turf with less resources. It's brave, but foolish.
I can see how Tinygrad could succeed. The story makes sense. AMD's doesn't, neither logically nor empirically. NVidia would have to seriously fumble.