Qwen3.8-27B UD-Q4_K_M on H200 NVL
GGUF IQ4_XS+Q5_K
The tested file was 16.46 GB. The repo's current file is 16.56 GB: it was re-uploaded after the run.
Engines
llama.cpp
b11571 · 097cf33d9 ↗measured 2026-11-09 21:00 UTC
78.6 tok/s
27% of the roofline286.7 tok/s possible
- now, a token
- 12.72 ms
- floor
- 3.49 ms
- read, a token
- 15.74 GB
More
- memory, measured
- 4513 GB/s
What was run
llama-bench -m Qwen3.8-27B-UD-Q4_K_M.gguf -p 0 -n 33 -d 512 -ngl 99 -r 1 --no-warmup -o json -ub 512 -b 2048 -ctk f16 -ctv f16 -fa -1Speed at context
128 tokens
512 tokens
Where the time goes
- Now, a token
- 12.72 ms
- Roofline floor, a token
- 3.49 ms
- Gap to the floor
- 9.23 ms
- Kernels a token
- 497
Launch heavy: 1914 kernel launches per token
Per role ms a token. floor (bytes ÷ measured memory speed) above the floor
Per kernel (8): ms now / ms at the roofline, launches a token
No floor = the kernel reads no weights (norms, conversions, copies): all of its time is gap.
Run h200nvl-in-model
Tested on
- GPU
- H200 NVL · 140 GB
- System RAM
- 1008 GB
- CPU threads
- 224
- Memory, measured
- 4513 GB/s
- Temp, max
- 59 °C
- Power
- 126 W mean of 500 W
- SM clock
- 1780 MHz
- Driver
- 615.78.08
- Measured
- 2026-11-09 21:00 UTC · context 512 tokens · measured, in-model (nsys): 33 of 39 weight roles timed
Qwen3.8-27B on other GPUs
RTX 5090vLLM + MTP, 4-bit draft layer · W4A16-AWQ · 4-bit head202.1 tok/s H200vLLM · W4A16-AWQ100.8 tok/s · 34%2× Radeon R9700vLLM + MTP k=2 · Quark W4A1692.3 tok/s H100 80GBvLLM · W4A16-AWQ87.8 tok/s · 41%H200 NVLllama.cpp · UD-Q4_K_M78.6 tok/s · 27%2× RTX 3090vLLM · W4A16-AWQ68.6 tok/s · 76%RTX 4090vLLM · W4A16-AWQ54.4 tok/s · 88%RTX 3090vLLM · W4A16-AWQ46.2 tok/s · 80%L40SvLLM · W4A16-AWQ43.6 tok/s · 89%
The fastest setup measured on each GPU, one user. The percentage is that GPU's own roofline.