REPRODUCIBILITY · CROSS-SKU · 2026-10-27

5-second avatars.
15-second FastH3.
Reproduce across GPU SKUs.

Pinned models, inputs, precision, GPU counts and timing boundaries. Two workloads only: the deployed Ref2VA avatar and the published 15s FastH3 reference. Run each unchanged on both GPU SKUs.

Published: historical measurements and a reproduction package. The new launcher has passed CPU tests, source reconstruction, historical-log checks, and local HTTP / audiovisual decoding checks. Fresh GPU runs with this launcher are pending. The exact target hardware referred to as “5500” and its results have not yet been verified.

The ZIP includes source code, patches, the avatar reference image, fixed requests and launchers. Download model weights separately. SHA-256 and file sizes · Source revision __SHORT_COMMIT__

Historical RTX PRO 5000 references

Both configurations below used 8 GPUs. These are whole-configuration request latencies, not single-GPU results or isolated quantization speedups.

5s · Deployed avatar configuration

Ref2VA Turbo · 4 steps

__AVATAR_MEDIAN__ s

Median HTTP latency over 13 measured continuation segments. 960×544; 124-frame windows with 5-frame overlap. Each measured segment plays for 4.958 s.

MXFP8 DiT · Dense Sage · NVFP4 video VAE. Historical application-ready median: 3.706 s, including additional application checks.

15s · Published reference

FastH3 · 4 steps

__FASTH3_MEDIAN__ s

Median of 3 measured requests. 1280×704, 362 frames / 15.083 s; TP1 / SP8.

MXFP8 DiT · Sage/VSA · Full NVFP4 VAE · BF16 RDMA.

The original avatar quality page shows a separate 10-second / 243-frame comparison. Its 6.993 s measurement is not a 5-second latency result. This reproduction uses the saved 16-request sequence, excluding the first 3 warmup segments. Original timing and media records.

Two pinned configurations and individual samples

Every number below is historical. Recompute the medians from the downloadable samples, then rerun both workloads with the same new launcher on each machine.

Hardware attribution: the project records describe the source system as RTX PRO 5000-class Blackwell. Some raw driver inventories report only “NVIDIA Graphics Device”. New runs record PCI device and subsystem IDs, architecture and memory as well as the driver name.

__ROWS__
Pinned caseGPUsOutputStepsHistorical median / sMeasured nTarget SKU

Dots follow the original request order. The dashed line marks actual playback duration for the selected case. These 3–13 samples do not establish a production tail-latency distribution.

Original artifact audit · 15s qualification receipt. The 15s sample mean is 13.311 s; its median is 13.304 s. This selected reference is not a claim of the fastest result across every historical branch.

Download the Docker image

Docker image release ↗ Download image loader ↓

The loader downloads all archive parts, verifies SHA-256 checksums and loads h3-sm120:20261027. Model weights and AdaLN sidecars are separate. The image passed build, dependency and CPU checks plus both runtime CLI imports; fresh GPU inference is pending.

Use Linux x86-64, NVIDIA Container Toolkit, eight idle identical SM120 GPUs, sufficient GPU/host memory, and an operational RDMA path. These profiles are not qualified for no-NIC execution. Model download, cache preparation and Docker run commands.

Intermediate stages and checkpoint paths

Stage comparisons and checkpoint map · 14-arm historical evidence JSON · Formal timing samples CSV

The 15s precision trial records medians of 13.545 s for all-MXFP8, 13.934 s for BF16 output projections, and 15.255 s for a BF16 first step. Separate matched tests record 13.551 → 13.304 s for cached normalization and early AAC, and 13.334 → 13.261 s for later output fusion, with identical MP4s to each round's quantized parent.

The earlier avatar BF16 trials use different reference inputs and, for some arms, a different resolution. They are grouped separately from the deployed 3.589779 s reference. Intermediate precision modes reuse base and adapter weights; the evidence does not establish separate trained checkpoint downloads or a perceptual-quality ranking.

GeometryDeployed 5s avatar15s FastH3
Input image1672×941 PNG; resized to 960×544; uploaded twiceNone; text-to-video/audio
Requested output960×5441280×720
Decoded output960×544; 119 continuation frames at 24 fps1280×704; 362 frames at 24 fps

Run on the 5000 and target SKU

Follow the complete setup guide to prepare pinned dependencies, model weights and AdaLN caches, then edit the machine-specific asset paths. Use the same source revision, GPU count, inputs and cache files on both machines.

git clone https://github.com/lishunyang12/h3-optimizations.git
cd h3-optimizations
git checkout __COMMIT__
python3 scripts/sku_benchmark.py prepare --variant avatar5 --output /opt/h3-runtime-avatar5
python3 scripts/sku_benchmark.py prepare --variant reference --output /opt/h3-runtime-15s
cp configs/sku-assets.example.json /tmp/h3-assets.json
# After editing paths, downloading pinned weights and building caches:
python3 scripts/sku_benchmark.py suite --cases main \
  --assets /tmp/h3-assets.json --gpus 0,1,2,3,4,5,6,7 \
  --output /results/pro5000
# On the target machine, use /results/target-sku as the output directory.
python3 scripts/sku_benchmark.py compare \
  /results/pro5000/5s-avatar-ref2va-4step/result.json \
  /results/target-sku/5s-avatar-ref2va-4step/result.json \
  --output /results/avatar-comparison.json

Repeat the comparison for 15s-4step-mxfp8-sage-nvfp4. The main suite runs exactly these two cases. The launcher exits with a reason if GPU architecture, count, memory or availability does not meet the selected case requirements.

What makes the comparison reproducible

Pinned and recorded

  • Original prompts, reference images, seeds, frame counts, steps and precision settings.
  • Model and cache byte hashes, actual source and binary fingerprints, and dependency versions.
  • Driver-reported GPU name, PCI IDs, memory samples, driver, clocks, power limit, CPU and GPU interconnect topology.
  • Separate warmup records; timing from local request launch to the complete MP4 written to disk.

Checks and interpretation

  • Four-step cases audit actual DiT calls, quantization fallbacks, VAE execution and RDMA evidence.
  • Frame counts and full audiovisual decoding are checked outside the timed interval. MP4s and logs are retained.
  • Different configurations, weights, sources, dependencies or foreign GPU activity invalidate a comparison.
  • CPU, interconnects, power and clocks still affect end-to-end latency. This is a same-workload system comparison.

The avatar replay excludes chat, RTMP and live playback delay. Quantized configurations are lossy. Cross-GPU output-byte differences are reported separately and do not automatically rank perceptual quality. Original benchmark prompts remain verbatim to preserve the workload.

Package validation record · Launcher source · All pinned configurations · Verified context and evidence boundaries