Six weeks with Intel's Arc Pro B70 in a production inference fleet
2 days ago
- Intel Arc Pro B70 offers 32GB VRAM for ~$1,000, but fails under sustained inference load.
- Under light bursty loads the card works fine for weeks, but sustained high concurrency causes engine memory CAT errors and device loss.
- Crash damage accumulates after several resets, requiring a full host reboot to restore stability.
- Extensive troubleshooting ruled out thermals, power, PCIe, and driver updates; the likely culprit is the closed GuC firmware blob.
- Intel's software ecosystem is in poor shape: its LLM libraries are archived, and alternative paths like llama.cpp and Vulkan also crash.
- A used RTX 3080 Ti at half the price runs same sustained workloads without any device losses.
- The card is removed from the fleet; not recommended for sustained serving until Intel ships a fix.