Tenstorrent Galaxy Blackhole

A 6U air-cooled chassis holding 32 Blackhole ASICs and rated by Tenstorrent at 23 PFLOPS of Block FP8, with 6.2 GB of on-die SRAM at 2.9 PB/s, 1 TB of GDDR6 at 16 TB/s, and 32 TB/s of aggregate Ethernet from ten 400 GbE links per ASIC plus up to 56 external 800 GbE QSFP-DD ports carrying 11.2 TB/s. It draws 8 to 10 kW on average and 12 kW at maximum, and is hosted by a single AMD EPYC 9004 with up to 32 cores and up to 576 GB of DDR5. Note that the Galaxy does not use the retail p150 bin: its figures work out at 718.75 TFLOPS and 193.75 MB of SRAM per ASIC, against 664 TFLOPS and 180 MB on the p150 card, so the ASICs here are a higher bin of the same silicon and Tenstorrent does not publish their Tensix core count. The compute figure recorded is Tenstorrent's own 23 PFLOPS for the machine, not 32 times the card, which would understate it by eight percent. Tenstorrent labels nothing dense or sparse and publishes no sparse rate.