2 ms·
The AMD MI300 is nowhere near Hopper. A real threat is Cerberas. Their approach avoids the memory bottleneck of loading the whole model DRAM->SRAM for each tok
by alecco 2y ago
The AMD MI300 is nowhere near Hopper.
A real threat is Cerberas. Their approach avoids the memory bottleneck of loading the whole model DRAM->SRAM for each token batch. I hope their IPO goes well before the AI bubble pops.
- jiggawatts 2y agoI can see variants of the Cerebras approach taking slices of the VC investment pie. For example, arrays of identical and simple chips with in-chip memory could have similar performance as the monolithic Cerebras wafers-scale chips. Not for all workloads, but some, such as inference.