Worker · RPC
Corsa
AMD Ryzen AI Max+ 395Radeon 8060S · ROCmx86Windows 11
- MachineCorsair AI Workstation
- Memory128 GB unified
- Holdsits share of the model
- Enginellama.cpp b1341 · ggml-rpc-server
peak during run 140 °F
Hive Benchmarks · Oct 11, 2026
Same speed helper, two boxes. Now it makes answers 21–37% faster.
DSpark is DeepSeek's official draft model for V4-Flash-0731: a small model guesses words ahead and the big model checks them. On three boxes it made answers 24–32% slower. On just the two AMD boxes, joined only by the USB4 cable, the same drafter made answers 21–37% faster. That is our fastest DeepSeek yet, at about 20 tokens per second.
USB4 cabledirect box-to-box link
Head · llama-server :8081 · DSpark on
CC
AMD Ryzen AI Max+ 395Radeon 8060S · ROCmx86Windows 11
- MachineROG Flow Z13 GZ302EA
- Memory128 GB unified
- Holdsits share + the 10.9 GB DSpark drafter + the bench
- Enginellama.cpp b1341 · llama-server
peak during run 185 °F
20.1tok/s
Best answer speed (512 prompt after 4K history). Plain 2-box did 14.7.
+37%best
Biggest gain, on the 512 prompt after 4K history
48%kept
Draft guesses the big model accepted (1,232 of 2,554)
4/4clean
Every test finished, zero errors, zero heat alerts
+30%
512 prompt · no history
14.89 → 19.42 tok/s
+33%
2,048 prompt · no history
14.97 → 19.92 tok/s
+37%
512 prompt · +4K history
14.72 → 20.14 tok/s
+21%
2,048 prompt · +4K history
14.46 → 17.56 tok/s
Plain vs. DSpark, same two boxes, same test
The Hive Benchmarking Stack (llama-benchy 0.4.0): prompts of 512 and 2,048 tokens, 128-token answers, with no history and with 4,096 tokens of history. Each test runs 3 times. Both runs used the same model file, engine, boxes and flags; the only change is the DSpark drafter on the head.
Answer speed
tokens per second · higher is better
21–37% faster at every size. Same guess rate as on three boxes; the fast link is what made it pay.
Prompt reading
tokens per second · higher is better
4–26% slower. The drafter has to read the prompt too.
Wait for the first word
seconds · lower is better
Up to 8 seconds longer on the long tests, from the slower prompt reading.
Plain 0731, 2 boxes0731 + DSpark, 2 boxes
Temperatures
All boxes in the run were logged. CC carries the drafter as the head and peaked at 185 °F, under the 203 °F automatic stop. Corsa peaked at 140 °F.
red line = 203 °F automatic stop
What this tells us
On three boxes DSpark kept 51% of its guesses and lost; on two boxes it kept 48% and won. The difference is the link: every guess-and-check round crosses the boxes, and the 1 Gb network to the third box made each round too slow. Over USB4 alone the rounds are cheap enough to pay off.
Draft acceptance is summed over every logged request, including warm-ups. DSpark drafts up to 3 tokens per step (the recommended setting for this drafter).