Worker · RPC :50053
Corsa
AMD Ryzen AI Max+ 395Radeon 8060S · ROCmx86Windows 11
- MachineCorsair AI Workstation
- Memory128 GB unified
- GPU targetgfx1151
- Enginellama.cpp b1328 · ggml-rpc-server
peak during run 136 °F
Hive Benchmarks · Oct 8, 2026
Three boxes. Two chip makers. One brain.
DeepSeek-V4-Flash, an 86.7 GB model, ran as a single model split across two AMD Strix Halo boxes and an NVIDIA Grace Blackwell box. That's x86 and ARM, ROCm and CUDA, Windows and Linux, all talking over one llama.cpp RPC pool.
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
- Holdsthe model file + the bench
- Enginellama.cpp b1328 · llama-server
peak during run 179 °F
1 Gb network911 Mbit/s · 0.71 ms
Worker · RPC :50052
Spark
NVIDIA GB10 Grace BlackwellCUDA 13ARMLinux
- MachineASUS Ascent GX10
- Memory128 GB unified
- Holds~29 GB of the model
- Enginellama.cpp 8172e65 · CUDA build
peak during run 136 °F
3boxes
2 chip makers, 2 CPU families, 2 operating systems
13.4tok/s
Answer speed. Same as two boxes.
127tok/s
Best prompt-reading speed (2,048-token prompt)
~290GB
Pooled model memory, up from 192 GB on two boxes
Two boxes vs. three, same model, 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.
Prompt reading
tokens per second · higher is better
About 20% slower with three boxes. Spark's slice travels over the 1 Gb network instead of a cable.
Answer speed
tokens per second · higher is better
Flat. Adding a third box cost nothing on the speed you feel while it talks.
Wait for the first word
seconds · lower is better
Longer prompts and more history take longer to read before the first word comes back.
2 boxes · CC + Corsa · Oct 73 boxes · CC + Corsa + Spark · Oct 8
Temperatures, logged the whole run
Every box ran its own logger with a 203 °F cutoff that stops the work on its own. No box came close.
red line = 203 °F automatic stop
Room on each box, measured with the model loaded
The model split itself almost evenly, about 29 GB on each box. Each box still has well over half its model memory free.
AMD boxes: 96.0 GB is set aside for the GPU (the rest of the 128 GB runs Windows). Spark shares one 121 GB pool with Linux and its live services, so its free number is what's free for anything. Only the head box (CC) needs the model file on disk; the workers get their share over the wire.
What the extra room unlocks
Adding boxes adds room, not speed. Room is what decides which models can live here without an API. Bar widths are pooled model memory (96 GB per AMD box, measured; about 100 GB on Spark); dashed lines are each model's file size.
Full GLM-5.3 (744B) sizes are from unsloth's GGUF listings: UD-Q2_K_XL 253.9 GB fits three boxes with room for context; IQ3_XXS at 281.7 GB is too tight. The 5-box rung assumes two more Corsairs joining over USB4.