AMD-powered startup aims high, claiming 20,000 accelerators by 2024 that outperform Nvidia's latest
-
AMD-powered startup TensorWave plans to have 20,000 MI300X accelerators deployed by 2024, claiming they beat Nvidia's H100 chips in raw performance specs.
-
TensorWave touts the MI300X's higher floating point performance, larger 192GB HBM3 memory, and lower cost compared to Nvidia alternatives.
-
TensorWave is currently installing systems with 8 MI300X GPUs per node, targeting 40kW power draw per rack with rear door heat exchangers for cooling.
-
While interest is high in AMD as an alternative to Nvidia, some customers question if performance will match what they currently get from Nvidia.
-
Long-term plans include direct-to-chip liquid cooling and using GigaIO's FabreX to connect up to 5,750 GPUs together in a single domain.