3 ms·
It's super interesting that we've built the biggest possible chip, a wafer-scale beast, and then, immediately in v2.0, the question is: how do we interconnect t
by rektide 5y ago
It's super interesting that we've built the biggest possible chip, a wafer-scale beast, and then, immediately in v2.0, the question is: how do we interconnect these? How do we scale them out?
There's so many interesting interconnection technologies growing, but there's so little mainstreaming, so little general availability. AMD tried so hard to make HyperTransport a viable, interesting ecosystem almost two decades ago (as I understand it by borrowing a lot from DEC). But they gave up & now InfinityFabric is for internal use & communication only.
Meanwhile we have other connects like CXL, CCIX, GenZ, OpenCAPI, TileLink (on RISC-V). NVLink and other commercial offerings are there. Few are targeting quite the colossal width we see here, that we see in Tesla Dojo chips. And the presence on the market is unbelievably slim. Some CCIX and some CXL has been used, primarily as attached FPGA coprocessing, not a lot more it feels like. NVLink is well used inside NVidia ultra-dense boxes. But a lot of this tech is still in very formative stages.