"Each Galaxy system features 1 TB of GDDR6, 16 TB/s of memory bandwidth, and 23 petaFLOPS of dense FP8 performance, all in a system that'll set you back only $110,000."
"Tenstorrent's mesh network isn't limited to a single node. It can be extended to support larger models, higher throughput, or more interactive user experiences by adding more systems."
"At least for DeepSeek V3, Tenstorrent claims its four-node Blackhole Galaxy Superclusters can process a 100,000 token prompt in less than four seconds."
Tenstorrent's Galaxy Blackhole AI compute platform is now available, featuring 32 Blackhole accelerators per 6U system with 100 Tbps bandwidth. Each system offers 1 TB of GDDR6, 16 TB/s memory bandwidth, and 23 petaFLOPS performance for $110,000. The architecture supports expansion for larger models and improved user experiences. The Galaxy Supercluster, priced at $440,000, can accommodate up to 32 nodes. Software improvements have enhanced model support and performance scaling, allowing for rapid processing of large prompts and high token output per user.
Read at Theregister
Unable to calculate read time
Collection
[
|
...
]