Tenstorrent Blackhole p150, tt-metal: Llama-3.1-8B decode tok/s per-user + aggregate (vendor-measured b32)
33.1 tok/s per user at batch-32; 1059 tok/s aggregate
Fields
id
benchmarks/comm-tt-blackhole-p150-llama31-8b-decodeMetric
tok_s_per_user
Value
33.1
Unit
tok/s per user at batch-32; 1059 tok/s aggregate
Model
Llama-3.1-8B
Methodology
VENDOR-MEASURED in the tenstorrent/tt-metal model matrix. Batch-32 decode, ISL=128. Llama-3.1-8B on Blackhole p150: 33.1 t/s/u and 1059 t/s AGGREGATE; the row is marked '*' = Blackhole software under active development. 2x p150 DP: 1190 t/s. 32 GB GDDR6 at 512 GB/s lets it hold bigger models than the n150's 12 GB.
Measured by
vendor
Reproducible
yes
notes
The strongest single-card answer in the Tenstorrent lineup. p150a/b $1,399 / p100a $999 NEW. Llama-3.1-8B is 'Experimental' on the single p150 card but 'Complete' on multi-card/QuietBox configs - the software is still maturing. Buy Wormhole or Blackhole for the open stack; do NOT buy the legacy Grayskull for LLM (no current LLM demos in tt-metal).
Notes
The strongest single-card answer in the Tenstorrent lineup. p150a/b $1,399 / p100a $999 NEW. Llama-3.1-8B is 'Experimental' on the single p150 card but 'Complete' on multi-card/QuietBox configs - the software is still maturing. Buy Wormhole or Blackhole for the open stack; do NOT buy the legacy Grayskull for LLM (no current LLM demos in tt-metal).