As soon as you scale out to multiple machines, you're already thinking about non-uniform access. If you just push that down one layer further, I don't see all that much of a difference if you're already having to do the work in your software to deal with it.
Interesting seeing Jalapeno and now Mia moving away from UMA to maximize NUMA
As soon as you scale out to multiple machines, you're already thinking about non-uniform access. If you just push that down one layer further, I don't see all that much of a difference if you're already having to do the work in your software to deal with it.
they have access to OpenAI IP on Jalapeno so not sure why they are not killing it yet on these.
Since Microsoft doesn’t have to pay NVIDIA’s margins for these, they’re probably getting them 2-3x cheaper.
That five trillion dollar market cap has an expiry.