Huawei Atlas 300I Duo stirs LLM doubts

Big VRAM is the selling point here, but the ecosystem looks like the real bottleneck. Until people can show reproducible local-LLM runs outside Huawei’s stack, this stays an interesting inference card rather than a mainstream homelab buy.

• –Huawei’s official product page lists 96GB or 48GB LPDDR4X, 280 TOPS INT8, 140 TFLOPS FP16, and 150W power, so the hardware spec is real and not just rumor
• –The Reddit replies point to support friction: drivers, host compatibility, and dependence on Huawei servers or Huawei’s software stack
• –I found official Huawei docs, but not convincing public firsthand token/s benchmarks for common open models like the ones buyers usually want
• –That makes the card compelling for memory-bound inference on paper, but risky for hobbyists who want CUDA-like plug-and-play support
• –If someone has it working well, the useful proof would be a repeatable benchmark on a real open model, not a spec sheet or teardown