Comment by adastra22
1 day ago
The Oxide approach, I think, would be to pick an open interconnect standard and work with one or more NVIDIA competitors to make hardware for it. E.g. AMD and a few of the specialized AI hardware companies. Essentially provide an open TPU standard.
They already picked out a interconnect standard it is called 100 Gigabit Ethernet. They can interoperate with multiple different vendors using the sfp28 ports on the switch. I bet they are even working on new programmable hardware for a next gen 200 Gigabit Ethernet fabric with oxide controlling the entire stack from hardware to software
Ethernet would add way too much latency to this application.
Not sure Tenstorrent agrees.
ever heard of RoCE? what latency are you thinking of in the Spectrum-X 800Gbit RoCE era?
Interesting. What would the incentives be for NVIDIA, for example, to opt into this? They are building their own racks now, like the NVL72 etc.
None, they won’t. But it would out Oxide into the strong negotiating position, and not at the mercy of NVIDIA.
NVIDIA has one primary weakness here: GPUs are NOT optimal hardware for training or inference. No competitor has the market reach to challenge them though. An open standard supported by high growth, breakout startup would change that though.
thanks for the chuckle! nice to see such naiveté :-)