Tenstorrent recently unveiled its Blackhole AI accelerators, claiming they can surpass the Nvidia A100 in raw compute and scalability, making significant strides in chip performance.
The Blackhole chip's core boasts 745 teraFLOPS of FP8 performance and an impressive 32GB of GDDR6 memory, yet it falls short in memory capacity compared to the A100.
The Blackhole Galaxy, with its 32 interconnected chips, promises to deliver 23.8 petaFLOPS of FP8 performance, showcasing a dramatic speed advantage over traditional Nvidia setups.
Davor Capalija highlighted the modularity of the Blackhole architecture, stating, 'You can make an entire training cluster just using this as a Lego,' emphasizing flexibility in system design.
Collection
[
|
...
]