← All reports

Report · 2026-11-18

Qwen3-8B BF16 on 1× H100 SXM (Lium spot): vLLM defaults vs vLLM's recommended setting

1× NVIDIA H100 80GB HBM3 Qwen3-8B BF16 weights, BF16 KV cache vLLM 0.33.0 TTFT p99 ≤ 2.0 s · TPOT p99 ≤ 50 ms

Quick-mode measurement on one Lium spot-tier H100 SXM (Atlanta), 2026-11-13/14, $1.50 per GPU-hour at rent time, vllm/vllm-openai:v0.33.0. The node passed Bench's node check first (700 W limit, 688 TFLOPS BF16, 3,032 GB/s copy bandwidth). The larger prefill batch from vLLM's optimization guide didn't move the latency-bound capacity: 3,577 vs 3,576 tok/s; the limit is time per output token. $0.12 per million output tokens at $1.50/h. Lium has given TokenWatt credits for earlier node evaluations and this run was paid from that balance; Lium had no input on it. Write-up: /blog/mi350x-vs-h100-software-and-price.html

ConclusionUnder SLO (TTFT p99 ≤ 2.0 s · TPOT p99 ≤ 50 ms), serving the same billable throughput as 64 cards on vLLM defaults takes: --max-num-batched-tokens 16384 failed the accuracy gate (-4.0 pts) and is excluded from capacity math, despite 3,577 tok/s goodput.

Results

Each value is the median of 2 runs.

vLLM defaults--max-num-batched-tokens 16384

SLO goodput

output tok/s per node · higher is better
vLLM defaults3,576vLLM defaults: 3,576 output tok/s per node
--max-num-batched-tokens 163843,577--max-num-batched-tokens 16384: 3,577 output tok/s per node

Tokens per joule

tok/J, GPU board power · higher is better
vLLM defaults5.33vLLM defaults: 5.33 tok/J, GPU board power
--max-num-batched-tokens 163845.33--max-num-batched-tokens 16384: 5.33 tok/J, GPU board power

Decode bandwidth utilization

% of rated HBM bandwidth
vLLM defaults51%vLLM defaults: 51 % of rated HBM bandwidth
--max-num-batched-tokens 1638451%--max-num-batched-tokens 16384: 51 % of rated HBM bandwidth

Axis runs to 100% of rated bandwidth.

GPU board power

kW, steady state
vLLM defaults0.67vLLM defaults: 0.67 kW, steady state
--max-num-batched-tokens 163840.67--max-num-batched-tokens 16384: 0.67 kW, steady state

GPU board power only — excludes CPUs, memory, NICs, fans and PSU losses.

All numbers

Cards for equal goodput = 64 × reference goodput ÷ configuration goodput, rounded up.

MetricvLLM defaults--max-num-batched-tokens 16384
Engine buildvLLM 0.33.0 · KV BF16vLLM 0.33.0 · KV BF16
Decode bandwidth utilization (MBU)51%51%
GPU board power0.67 kW0.67 kW
SLO goodput3,576 tok/s3,577 tok/s
Tokens per joule5.335.33
TTFT p99406 ms407 ms
TPOT p9938.6 ms38.3 ms
SLO metyesyes
Accuracy (gsm8k-5shot)94.090.0
Accuracy Δ vs. referencereference-4.0 pts
Cards for equal goodput64—

Setup

Any change to these fields makes it a different report.

Accelerator
1× NVIDIA H100 80GB HBM3
Rated HBM BW
3.35 TB/s per GPU
Driver / runtime
580.105.08
Host
Lium spot-tier pod (interruptible marketplace host), node 00750dfa, Atlanta, US; $1.50 per GPU-hour on 2026-11-13
Engine
vLLM 0.33.0
Kernel libraries
vllm/vllm-openai:v0.33.0
Model
Qwen3-8B (Dense)
Quantization
BF16 weights, BF16 KV cache
Parallelism
1 replica × TP1
Workload trace
chat-mix-v1
Input tokens
median 1,024 · p90 4,096
Output tokens
median 256 · p90 1,024
Arrival
Poisson, rate swept to SLO boundary
Power source
GPU board power sum (NVIDIA) · 1 Hz · 120 s steady state
Runs per config
2 (variance band ±0.4%)
Accuracy gate
gsm8k-5shot · max drop 1.0 pts

Not covered

Do not extrapolate this report to the following.

  • Whole-node power (no BMC access on this host)
  • Context lengths beyond the chat-mix-v1 trace
  • Quick mode: short windows, few repeats — validation only, not for publishing
  • Accuracy: quick mode scores GSM8K on 50 questions with one seed, so one question is 2 points and the 1-point gate can't tell a real drop from sampling noise. A −4 pt flag is two questions
  • Spot (interruptible) node; it wasn't interrupted during the runs

Reproduce

Run on the same hardware and software versions. Results should fall within ±0.4%.

shell
Raw run files, logs and the exact launch commands: bench/results/h100-tuned-2026-11/ in the TokenWatt repository (README.md).

tokenwatt run --spec lambda-qwen3-8b-tp1-dp1.spec.yaml --label 'vLLM defaults' --role baseline --out runs/vllm-defaults.json
tokenwatt run --spec lambda-qwen3-8b-tp1-dp1.spec.yaml --label '--max-num-batched-tokens 16384' --role tuned --out runs/max-num-batched-tokens-16384.json

tokenwatt report runs/vllm-defaults.json runs/max-num-batched-tokens-16384.json --title 'Qwen3-8B BF16 on 1× H100 SXM (Lium spot): vLLM defaults vs vLLM's recommended setting'

Download raw JSON