← All reports

Report · 2026-11-13

Prefix caching on a hybrid model, one H100: Qwen3.8-27B FP8, vLLM 0.31.0, chat-like traffic

1× NVIDIA H100 80GB HBM3 Qwen3.8-27B-FP8 FP8 weights, BF16 KV cache vllm 0.31.0 TTFT p99 ≤ 2.0 s · TPOT p99 ≤ 50 ms

Quick-mode measurement on one H100 SXM on RunPod Secure Cloud (November 5, 2026): Qwen3.8-27B FP8, a hybrid model (48 Gated DeltaNet and 16 full-attention layers), on vLLM 0.31.0 with default automatic prefix caching (mamba cache mode 'align', 784-token cache blocks with the default BF16 KV cache), --max-num-seqs 512, text-only, chat-like workload (median 1,024 input / 256 output tokens). The same SLO search ran four times with 0%, 50%, 75% and 90% of every prompt drawn from eight shared prefixes; cache hit rates are vLLM's own prefix-cache counters (0.0%, 21.1%, 40.8%, 55.6%). Capacity within p99 TTFT ≤ 2 s and p99 TPOT ≤ 50 ms rose from 3.43 to 5.83 requests per second. No accuracy gate (random-token prompts); the model answered a smoke-test question correctly at each start.

ConclusionUnder SLO (TTFT p99 ≤ 2.0 s · TPOT p99 ≤ 50 ms), serving the same billable throughput as 64 cards on No prompt reuse (0% cache hits) takes: with 50% shared prefix (21.1% cache hits), 64 → 55 cards and 14% less GPU board energy per token; with 75% shared prefix (40.8% cache hits), 64 → 48 cards and 25% less GPU board energy per token; with 90% shared prefix (55.6% cache hits), 64 → 39 cards and 37% less GPU board energy per token.

Results

Each value is the median of 2 runs.

No prompt reuse (0% cache hits)50% shared prefix (21.1% cache hits)75% shared prefix (40.8% cache hits)90% shared prefix (55.6% cache hits)

SLO goodput

output tok/s per node · higher is better
No prompt reuse (0% cache hits)1,379No prompt reuse (0% cache hits): 1,379 output tok/s per node
50% shared prefix (21.1% cache hits)1,63150% shared prefix (21.1% cache hits): 1,631 output tok/s per node
75% shared prefix (40.8% cache hits)1,87275% shared prefix (40.8% cache hits): 1,872 output tok/s per node
90% shared prefix (55.6% cache hits)2,30990% shared prefix (55.6% cache hits): 2,309 output tok/s per node

Tokens per joule

tok/J, GPU board power · higher is better
No prompt reuse (0% cache hits)2.27No prompt reuse (0% cache hits): 2.27 tok/J, GPU board power
50% shared prefix (21.1% cache hits)2.6550% shared prefix (21.1% cache hits): 2.65 tok/J, GPU board power
75% shared prefix (40.8% cache hits)3.0575% shared prefix (40.8% cache hits): 3.05 tok/J, GPU board power
90% shared prefix (55.6% cache hits)3.5990% shared prefix (55.6% cache hits): 3.59 tok/J, GPU board power

Decode bandwidth utilization

% of rated HBM bandwidth

Not measured in this report.

GPU board power

kW, steady state
No prompt reuse (0% cache hits)0.61No prompt reuse (0% cache hits): 0.61 kW, steady state
50% shared prefix (21.1% cache hits)0.6150% shared prefix (21.1% cache hits): 0.61 kW, steady state
75% shared prefix (40.8% cache hits)0.6175% shared prefix (40.8% cache hits): 0.61 kW, steady state
90% shared prefix (55.6% cache hits)0.6490% shared prefix (55.6% cache hits): 0.64 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.

MetricNo prompt reuse (0% cache hits)50% shared prefix (21.1% cache hits)75% shared prefix (40.8% cache hits)90% shared prefix (55.6% cache hits)
Engine buildvllm 0.31.0 · KV BF16vllm 0.31.0 · KV BF16vllm 0.31.0 · KV BF16vllm 0.31.0 · KV BF16
Decode bandwidth utilization (MBU)————
GPU board power0.61 kW0.61 kW0.61 kW0.64 kW
SLO goodput1,379 tok/s1,631 tok/s1,872 tok/s2,309 tok/s
Tokens per joule2.272.653.053.59
TTFT p99946 ms699 ms530 ms558 ms
TPOT p9945.9 ms47.0 ms45.8 ms48.2 ms
SLO metyesyesyesyes
Accuracy (none)————
Accuracy Δ vs. reference————
Cards for equal goodput64554839

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
595.91.07
Host
RunPod Secure Cloud 1x H100 SXM 80GB (US-MO-1, pod h4dvtjc8j7mu7e)
Engine
vllm 0.31.0
Kernel libraries
vllm/vllm-openai:v0.31.0 --max-model-len 32768 --max-num-seqs 512 --language-model-only --enable-prompt-tokens-details
Model
Qwen3.8-27B-FP8 (?)
Quantization
FP8 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 ±1.2%)
Accuracy gate
none · max drop 0.5 pts

Not covered

Do not extrapolate this report to the following.

  • Whole-node power (no BMC access on this host)
  • Quick mode: short windows, few repeats
  • Synthetic shared prefixes (8-prefix pool, random tokens)
  • Accuracy: no accuracy gate was run

Reproduce

Run on the same hardware and software versions. Results should fall within ±1.2%.

shell
tokenwatt run --spec h100-hyb-bf16-chat-prefix-cache.spec.yaml --label 'hyb-bf16-chat-0' --role other --out runs/hyb-bf16-chat-0.json
tokenwatt run --spec h100-hyb-bf16-chat-prefix-cache.spec.yaml --label 'hyb-bf16-chat-0.5' --role other --out runs/hyb-bf16-chat-0-5.json
tokenwatt run --spec h100-hyb-bf16-chat-prefix-cache.spec.yaml --label 'hyb-bf16-chat-0.75' --role other --out runs/hyb-bf16-chat-0-75.json
tokenwatt run --spec h100-hyb-bf16-chat-prefix-cache.spec.yaml --label 'hyb-bf16-chat-0.9' --role other --out runs/hyb-bf16-chat-0-9.json

tokenwatt report runs/hyb-bf16-chat-0.json runs/hyb-bf16-chat-0-5.json runs/hyb-bf16-chat-0-75.json runs/hyb-bf16-chat-0-9.json --title 'Prefix caching on a hybrid model, one H100: Qwen3.8-27B FP8, vLLM 0.31.0, chat-like traffic'

Download raw JSON