June 2, 2026
Cheap chip, costly chaos
Bringing Up DeepSeek-V4-Flash on AMD MI300X
AMD’s bargain AI chip looks amazing on paper, but commenters say the software struggle is very real
TLDR: Doubleword says AMD’s MI300X could be a cheaper, easier-to-find alternative to Nvidia for AI, but getting a major model to run was far messier than it should be. In the comments, people were intrigued and optimistic, but the big mood was clear: AMD might save money on chips while costing extra pain in setup.
This is the kind of tech story the comments were born for: a company found a cheaper, roomy alternative to Nvidia’s wildly in-demand AI chips, then ran straight into the classic catch — the hardware looks great, the software throws a tantrum. Doubleword says AMD’s MI300X should be a star for running big AI models because it has loads of memory and a lower price tag, but getting DeepSeek-V4-Flash working turned into a long, painful obstacle course. In plain English: the chip is powerful, available, and cheaper than the famous option, but making it actually behave still takes serious effort.
And the community reaction? Equal parts hopeful, battle-scarred, and very "here we go again." One commenter immediately wanted to know if the fixes could scale to an even bigger setup, basically treating the blog post like the season finale before the sequel. Another dropped the patch link like a receipts queen entering the chat. The loudest vibe, though, was a grimly amused consensus that AMD can work, but only if you’re willing to suffer for it first. Doubleword itself sounded bullish, saying AMD makes sense for less chatty, high-volume AI jobs — just don’t expect a smooth ride. That got backup from a commenter training on AMD hardware who casually admitted they got a model working “but it took a lot of work on the software side,” which is about as close as engineers get to screaming into a pillow. The meme practically writes itself: cheap chip, expensive headache.
Key Points
- •Doubleword says that as of early May 2026, vLLM with DeepSeek-V4-Flash does not work on AMD MI300X.
- •The article presents MI300X as attractive hardware on paper, with 192GB of HBM3, comparable FP8 compute to H100, and lower apparent rental cost.
- •A major compatibility issue is that MI300X supports only the older fnuz FP8 dialect, while broader software support has moved toward the OCP-standard FP8 format.
- •The article says some vLLM FP8 paths account for e4m3 versus e5m2 but not fnuz versus OCP, which can cause values to be off by a factor of two.
- •The article identifies missing optimized attention kernels in AMD’s AITER library as another major limitation, forcing slower fallbacks to generic Triton implementations.