CMP 170HX SM80 RC3

CMP 170HX SM80 RC3

NInfer port for the NVIDIA CMP 170HX (sm_80), based on the NInfer RTX 3090 branch.

Qualification

• Exact source commit: 1e2e943b722d93ccdf710d14b18dfc6c1323900f
• Validated image: ninfer-cmp170hx:sm80
  • image ID: sha256:df31cb012ad265a5ca82469ad0529fbc845ef6e8a66804c7973dce63d5a4a002
  • image version: 0.1.0-rc3
  • image provenance label: 3bbd0ab-sm80-public-candidate
• Validation host GPU: NVIDIA CMP 170HX, compute capability 8.0, 64 GiB visible memory, NVIDIA Linux driver 610.43.03.
• Fresh CUDA 13.1 SM80 source build completed all 245 build/link steps; focused protocol tests passed.
• Docker Compose runtime qualification passed: the container became healthy, /health returned OK, /v1/models reported qwen3.8-27b, and a real chat-completions request returned CMP170HX_OK with HTTP 200.
• The Compose healthcheck explicitly invokes /usr/bin/bash, avoiding the /bin/sh incompatibility with Bash /dev/tcp probing.

Home Assistant / OpenAI-compatible API note

For Home Assistant's Local OpenAI LLM integration, select its Generic OpenAI-compatible backend rather than the llama.cpp-specific backend. Operator validation after that configuration change observed more than 4000 prompt tokens/s and more than 210 generated tokens/s through the API.

Assets

• ninfer-cmp170hx-sm80-rc3.cdx.json — CycloneDX SBOM for the validated image.
• ninfer-cmp170hx-sm80-rc3.cdx.json.sha256 — SHA-256 checksum for the SBOM.
• Four privacy-reviewed validation screenshots. The public Qwen 3.8 NVTOP asset is a cropped copy; the unredacted original is not included.

Scope

This is a release candidate for the unlocked CMP 170HX configuration. It does not claim generic support for stock 8/10 GiB CMP cards or a replacement for every upstream NInfer target.