Couple of forward thinking points:
1) Infinity-Fabric as we know is AMD's glue that joins all chiplets and provides linear-scaling in performance. It's similar to the tech I believe that's behind Cerebras' mammoth chips (after all Gary Lauterbach was the founder of SeaMicro and ex-AMD that developed Infinity-Fabric). So in theory AMD can also put out huge chips that are 100X Volta-V100 performance for certain large workloads but imo, only once AMD attains the market-share and ecosystem to make it a viable play.
2) This is also why the 20x25 efficiency initiative was so important per Papermaster. Imo only once you can completely squeeze out efficiency from micro-architecture, only then are you ready for big macro-architecture products.
3) These socketed chiplet designs with pieces from Gloflo, TSMC and Samsung is just the beginning. It allows AMD to play and grow in the current ecosystems. Once AMD captures enough market-share and pays of debt, I strongly feel AMD will create it's own "dense-compute" market.
This is also why partnership with Samsung is important because Samsung has Memory, SRAM and fabs in-house, together with AMD's GPU and CPU and IF we have everything we need for dense-compute monolithic chips.
There's a reason why Dr. SU keeps saying "AMD is HPC, AMD is HPC, AMD is HPC". It will happen imo and when it does we're looking at $100+ easy on tech like this. It will start within Radeon imo.
Beast is yet to come.
Was haltet ihr von dieser Meinung?
Er ist einer von den wenigen aus dem Yahoo Board, der eigentlich Plan von der Materie hat. |
Thread abonnieren
