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
temperature not logged this run
Hive Benchmarks · Oct 11, 2026
Fewer boxes, faster. Two boxes read prompts up to 53% faster than three.
DeepSeek-V4-Flash-0731 fits on two AMD boxes, so we dropped the NVIDIA box and its 1 Gb network link and ran the same test on CC + Corsa over the USB4 cable only. Prompt reading jumped 41–53%, the wait for the first word fell by about a third, and answers got 2–4% faster.
USB4 cabledirect box-to-box link
Head · llama-server :8081
CC
AMD Ryzen AI Max+ 395Radeon 8060S · ROCmx86Windows 11
- MachineROG Flow Z13 GZ302EA
- Memory128 GB unified
- Holdsits share + the model file + the bench
- Enginellama.cpp b1341 · llama-server
peak during run 181 °F
259tok/s
Best prompt reading (512 prompt after 4K history). Three boxes did 169.
2.7sec
Wait for the first word on a 512-token prompt. Three boxes took 3.9.
15.0tok/s
Answer speed, 2,048 prompt. Three boxes did 14.4.
4/4clean
Every test finished, zero errors, zero heat alerts
+41%
512 prompt · no history
132.8 → 187.2 tok/s
+43%
2,048 prompt · no history
153.7 → 219.2 tok/s
+53%
512 prompt · +4K history
169 → 259.1 tok/s
+47%
2,048 prompt · +4K history
164.5 → 241.7 tok/s
Three boxes vs. two boxes, same model, same engine, 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. Same model file, engine and flags; the change is dropping the third box and its slower network link.
Prompt reading
tokens per second · higher is better
41–53% faster at every size. No more waiting on the slower network link.
Wait for the first word
seconds · lower is better
29–35% shorter. Up to 12 seconds less on the biggest test.
Answer speed
tokens per second · higher is better
2–4% faster. Answers still pass between the two boxes over USB4.
3 boxes (CC + Corsa + Spark)2 boxes (CC + Corsa)
Temperatures
CC as the head peaked at 181 °F, well under the 203 °F automatic stop. Corsa's temperature logger had stopped on its time limit just before this run, so Corsa was not logged; its last reading at the start was 127 °F. The logger was restarted for the next run.
red line = 203 °F automatic stop
Where this sits on the ladder
Two boxes hold this model with room to spare, and with only the USB4 cable between them it runs faster than on three. Bar widths are pooled model memory; dashed lines are each model's file size.
The boxes still take turns (layer split): each one runs its half of the model while the other waits. Spark stayed idle for this run.