[coryjett]

Problem and result

A results-only report for a setup the community reports don't cover yet: two RTX 3090s on PCIe 3.0 x16, one per CPU
socket, in a dual Haswell-EP server (AVX2 only), running original Flash-Next IQ3_S with a layer split. Tuned over 26
one-change-at-a-time rounds for a coding-agent workload, plus one draft-protocol standard run on the final
configuration (custom track): prompt 1,127 / 1,610 tok/s at 4K / 16K, decode 114-129 tok/s, sanity checks 12 of 12.

Two findings that may matter beyond this box:

• STRATA_PREFILL_CPU_SHARE=auto is off by default under a layer split. Set explicitly (max 4096), it cut the small
  incremental prompt reads that coding agents mostly send (300-4,000 new tokens on a cached ~66K context) by 8-13%
  (about -0.6 s at 1K and -0.9 s at 1.5K), with the 32K read, decode and greedy output unchanged within production's
  own spread. OLDER_GPUS.md reports no gain from it on a P40, so this looks specific to a PCIe 3.0 split with ~78% of
  the experts resident, not a general default change.
• strata bench run --url on a server using two identical cards records one card and none of the engine options
  (bench_facts.hardware keeps cards[:1] when there is no run config and the server names one GPU type). The same
  run with --profile <run config> --no-start records both cards and the options. The report keeps the
  --profile run and notes the discarded --url run.
• The same runner's secret filter (key|token|secret|password|auth) also redacts --conversation-cache-min-tokens
  because its name contains "token". The value is 20000 (the README says so; the runner's files are left as written).

The standard run used the server's learned expert profile (expert_profile_save), which the runner records but does
not count as a declared custom profile; the sweep rounds used the shipped one. Round 20 found no difference between
the two. The README says so.

Related discussion

None. Results only; no engine or default change proposed. The NUMA worker-pool on/off pair (one socket against both)
is in the findings list, for anyone following the closed #3708.

Approach

Adds bench/results/2026-10-28-community-t630-2x3090-pcie3/ (report, raw per-run output from rounds 21 and 24-26,
redacted configs, engine logs, the measurement scripts), the standard run's folder as the runner wrote it, and one
entry in docs/COMMUNITY_BENCHMARKS.md. Home paths, LAN addresses, host names and a private repository's file names are
replaced with placeholders.

Validation

Passed:

• strata bench run --profile <production run config> --no-start --track custom, Strata 0.2.7 official Linux
  release (engine 0.1.52), 2026-10-28 13:43 UTC: status complete, 21 of 21 required requests, sanity 12 of 12. The
  engine log shows exactly the run's 22 requests and nothing else during it.
• The production, agent-turn and incremental-read numbers come from the raw files in the folder (sweep.jsonl,
  speed-*.txt, rounds 21 and 24-26).

Failed:

• None observed.

Not run:

• The optional concurrency stage (the server has one slot).
• No builds: results and docs only.

Limits

• Apart from the standard run, the measurements use our own scripts at the server's temperature 0.6, not greedy.
  The scripts read from a private repository (the long-prompt and agent-turn contexts are its source files), so they
  document the method rather than run as is.
• The trim, NUMA, long-prompt and opt-in-switch findings are summarized from earlier rounds (2026-10-24 to 10-28)
  whose raw files are not included; they can be added on request.
• The standard run reused a production server that had answered other requests before it, not a fresh load.
• One machine; the CPU-share result needs confirming on other PCIe 3.0 splits before anyone changes a default.

🤖 Generated with Claude Code

[Niko1221]

Thank you for measuring this and writing it up. We are building a Strata website with a page for community benchmarks, and results like these will be added there when it launches, so we are closing this pull request here; the numbers are not lost. If your numbers change with a new version, you are welcome to send them again.