Conversation

comma
@comma_ai
.
@Qualcomm
proved as difficult to work with as
@nvidia
. You can't buy chips from these companies. The comma six was cancelled, and now the custom Qualcomm SOM was cancelled. Any startups making competing AI chips? Something like the Tesla FSD chip.
Quote Tweet
comma
@comma_ai
Anyone out there from @Qualcomm interested in selling us 20k SDM845 chips, with an expectation of 100k? Would much rather buy original then go through the Shenzhen markets. We have a custom SOM designed with an integrated panda. This is a key step to lowering our C3 BOM cost.
Show this thread
Image
70
301
Eren Doğan
@kurumuz
Replying to
@comma_ai
@Qualcomm
and
@nvidia
*sigh*, I almost feel like you will not be able to find many people doing this. Someone needs to rise up to the occasion for small but fast inference chips though, balanced SRAM and compute, fast matmuls, good compiler stack for fusion.
1
7
Eren Doğan
@kurumuz
Replying to
@kurumuz
@comma_ai
and 2 others
So many TFLOPS thrown away on these big cards because we don't have proper fusion stacks that just work on pytorch eager code(I guess can't expect all code to be graph-able and static). These things could be 10x faster, but they are just not.
2
6
Show replies