Endurance report

veloGB10's Qwen 3.8 27B NVFP4 + DFlash 2 configuration, served across four NVIDIA GB10 nodes (tensor-parallel ×4).

What this is

An 8-hour stability soak. The engine was left running continuously and driven with a mixed workload. The run measured three things:

  1. Does throughput drift down over hours (thermal throttling, memory growth, a slow leak)?
  2. Does the server stay deterministic (same prompt in → same text out)?
  3. Does anything crash, hang, or wedge overnight?

Eight contiguous hours of steady serving in a single session. The engine was left unattended and driven the whole time.

The rig

Piece What it is
Target model Qwen 3.8 27B NVFP4 (FULL, NVFP4-quantized)
Drafter Qwen 3.8 27B DFlash 2 (block speculator)
Engine veloGB10
Hardware 4 × NVIDIA GB10 (Grace Blackwell Superchip), tensor-parallel TP=4, unified LPDDR5X
Decoding DFlash 2 speculative. A block is drafted, the trunk verifies it, and the longest matching prefix is accepted

The configuration is the same one described in the README's Update: Qwen 3.8 27B NVFP4 with DFlash 2 section, at a 256K-capable context headroom.

What we ran

For 8 contiguous hours, the engine served a rotating mix of five content types (a code task, two prose/technical-narrative prompts, a short factual question, and a short descriptive prompt):

Content Prompt shape Character
code Build an LRUCache, O(1) get/put dense code generation
essay History of computing, Babbage → GPUs long prose
synth1 Energy ↔ entropy across scales technical narrative
chat_t0 Stack vs. queue short factual
chat_t1_off A quiet morning in a mountain village short descriptive

The batch mixer cycles through these five content types, so the run accumulates a mix of content types rather than a single repeated workload. Every request records client throughput (tokens/sec) and time-to-first-token. Hardware telemetry (GPU temperature, SM clock, power draw) is sampled on all four nodes every 5 minutes. Once an hour, a seed-fixed, temperature-0 request is served twice; the two completions are hashed and compared. This is a determinism canary.

The results

Throughput over 8 hours

Per-content-type mean throughput (tokens/sec), first half of the run vs. last half:

Content First half Last half Drift
code 97.74 97.59 −0.2%
essay 48.88 48.80 −0.2%
synth1 49.07 49.00 −0.1%
chat_t0 73.76 73.65 −0.1%
chat_t1_off 38.13 38.10 −0.1%

The largest measured drift is 0.2%. Throughput at hour 8 matches hour 1. The content-type spread (code at ~98, prose at ~49) reflects the compression of code relative to prose.

Time-to-first-token

Mean 202.6 ms across all ~2,400 requests. The first-half mean and last-half mean differ by +0.0 ms drift; the first-hour spread is σ = 9.4 ms. No latency change across eight hours.

Determinism

Every hourly canary (a seed-fixed, greedy request served twice and hash-compared) returned byte-for-byte identical. Eight checks, eight matches, zero failures.

Thermals

Mean GPU temperature per node over the run (min..max, mean):

Node Min Max Mean
1 64 77 73.7
2 58 69 65.4
3 59 70 68.8
4 60 72 71.4

All four nodes held moderate temperatures; the head node (node 1) ran warmest. SM clocks and power draw were stable throughout. The thermal profile is a sustained GB10 load without a thermal runaway or a throttling cliff.

Reliability

Over eight hours and ~2,400 requests: no crashes, no hangs, no wedges, no kill-condition trip, no watchdog firing.

Summary

Two caveats, stated so the data is read correctly:

  1. A short warm-up transient. The drafter's per-verify acceptance (tokens produced per speculative step) rose from ~2.9 to a stable ~3.5 during roughly the first half-hour, then held flat. This is a warm-up, not a drift; the client-visible throughput curve was flat the whole time, so a first-half vs. last-half comparison is not a drift signal.
  2. nvidia-smi reported N/A for the "memory used" field on these nodes in the sampling path. This is a GB10 unified-memory reporting behavior. The memory-monotonic watch was effectively a no-op this run. The thermal / clock / power curves are the ones that carried the drift signal, and they were clean. "No memory leak" is not independently measured here; re-check if memory growth is a specific concern.

Bottom line

After eight unattended hours and ~2,400 served requests across five content types, the Qwen 3.8 27B NVFP4 + DFlash 2 stack on 4× GB10 showed no throughput drift, no determinism failures, and no crashes.


Everything above is measured on the engine's live telemetry during a single continuous 8-hour session. Figures are representative of this run and this configuration; normal run-to-run variation applies.