DeepSeek V4 Flash on a Single AMD MI300X

374 points · 100 comments on HN · read original →

Points and comments are a snapshot, not live.

Article body wasn't reachable. The HN discussion summary is below.

What commenters are saying

Commenters note the MI300X is an OAM module, not a PCIe card, making single-unit purchase difficult. The cheapest access is AMD Developer Cloud at $1.99/hour. At 830 tokens/second, generating $0.54 worth of tokens per hour, self-hosting doesn't beat DeepSeek's API price. Privacy, stable throughput, and cache hits for agentic workloads justify the cost. The MI350P PCIe card with 144GB memory may fit the model but needs two cards. Demand for high-end GPUs is real, driven by various compute bubbles.