Geekbench 5.5.1

Latest state: the totals (same-test platform comparison) + the 21 per-workload tables · Canonical: docs/test/geekbench-report.md

§1Geekbench 5 data (5.5.1, 9950X, KVM q35 smp4/4G, strong kernel; per-workload tables 2026-11-04, totals updated through the gb-opt round 2026-11-15)

ItemRuxLinux (4-CPU pin)Ratio
Multi-core total75047027105%
Single-core total24652165110%

Test-platform note: the Linux side is the host-native taskset -c 0-3 pin (topology-matched to the guest's 4 vCPUs; a same-test platform KVM Linux comparison is queued). Two native self-runs differ by ~15% (2549 vs 2165) — deltas within ±15% are noise. The per-workload tables come from the 2026-11-04 full capture (that round's SC total was 2115; the quiet best is 2274 — same noise band).

Single-core, per workload (all 21)

ItemRuxLinuxRatio
AES-XTS8140 (re-run)7643106%
Integer · Text Compression1446153694%
Integer · Image Compression19011813105%
Integer · Navigation1416175381%
Integer · HTML519911651121%
Integer · SQLite18331480124%
Integer · PDF Rendering1758207485%
Integer · Text Rendering17711456122%
Integer · Clang1891216188%
Integer · Camera20051861108%
Float · N-Body Physics22271956114%
Float · Rigid Body Physics18851737109%
Float · Gaussian Blur2136214999%
Float · Face Detection2401277087%
Float · Horizon Detection18341459126%
Float · Image Inpainting41004092100%
Float · HDR45633614126%
Float · Ray Tracing27252042133%
Float · Structure from Motion18061661109%
Float · Speech Recognition1712174198%
Float · Machine Learning25112028124%

AES-XTS single-core carries the post-fix re-run value (8047 / confirm run 8140): the earlier low score was a kernel serial-print-tax artifact (AES-NI was probe-proven healthy — guest aesenc throughput 1268 vs native 1244 M rounds/s); the high-load capture is discarded. Source: docs/test/geekbench-report.md, 2026-11-02 standing.

Multi-core, per workload (all 21)

ItemRuxLinuxRatio
AES-XTS93318972104%
Integer · Text Compression57873851150%
Integer · Image Compression7410748199%
Integer · Navigation41873546118%
Integer · HTML580877151113%
Integer · SQLite70696829104%
Integer · PDF Rendering6566763286%
Integer · Text Rendering5794592498%
Integer · Clang7359782394%
Integer · Camera7191747796%
Float · N-Body Physics7174750196%
Float · Rigid Body Physics7463758998%
Float · Gaussian Blur76827694100%
Float · Face Detection99491029697%
Float · Horizon Detection7111761693%
Float · Image Inpainting1483812349120%
Float · HDR178721896294%
Float · Ray Tracing1337411370118%
Float · Structure from Motion7155748696%
Float · Speech Recognition59394537131%
Float · Machine Learning56415635100%

The multi-core split: the 11-04 capture's 7332 vs 7027 (104%) is pulled by scheduler-sensitive workloads — text compression (150%), inpainting (120%), speech recognition (131%), ray tracing (118%) — while AES-XTS SC 64% in the 11-04 capture was the pre-fix reading (now re-run at 105–106%); after the gb-opt round the multi-core total stands at 7504 = 107% (latest two-run class 7328–7504); Clang/PDF-rendering single-core residuals await isolated re-runs.

§2Test platform and capture method

  • Test platform: the 9950X host, q35 KVM -cpu host -smp 4 -m 4G, the strong build (build-release-opt); one boot per full run, 2700 s outer / 900 s in-guest budget.
  • Score capture: a local TLS capture server (a self-signed CA injected into the guest /etc/hosts via debugfs, pinning browser.geekbench.com to 10.0.2.2; a :443 server receives the HTTPS upload POST, parsed by parse-capture.py) — scores never touch the external network and the process is replayable.
  • Kernel bugs this suite found and fixed: mprotect returning ENOMEM for mid-VMA resizes; the TCP RX checksum reading stack garbage — commercial suites as compatibility probes.
  • An accounting note: the per-workload tables above are the 11-04 full capture (that round's SC 2115); the single-core AES-XTS takes the post-fix re-measurement (8047/8140), the historical low being the serial-print-tax artifact (AES-NI probe-verified intact). The 11-15/11-16 totals are the four-boot best; a per-workload re-capture is in the queue.

§3Reproducing

# the full pipeline (image reflink + CA injection + the capture server + parsing)
cd /home/william/rux-agents/gb-opt && ./run/bench-pipeline.sh
Permanent image store: /home/william/rux-images/testimages/geekbench-x86.img. Canonical: docs/test/geekbench-report.md (reproduction steps itemized).