Key points
- SemiAnalysis reports an 11-fold gain for AMD's MI355X
- The change was software, not hardware
- The same firm criticized AMD's software two weeks ago
- The full numbers haven't been published yet
SemiAnalysis, a research firm that runs an open benchmark of AI chips, reported an 11-fold performance gain for AMD's (AMD) MI355X on a demanding AI-agent test using MiniMax M3 after less than 19 days of software improvements. The firm posted the claim on X on Monday, September 7. The post does not specify whether the 11-fold gain measures throughput, task completion time or something else. "This was done entirely through software optimizations, mainly by optimizing the long-context attention op, along with other optimizations," the firm wrote on X. "This is the power of the ROCm stack: buy the hardware once, get performance updates over time through software optimizations for free." ROCm is AMD's software platform for GPU computing.
Two weeks earlier, the firm had sharply criticized AMD's performance on the same model. In its August 24 AgentX report, SemiAnalysis wrote that "AMD software performance is horrible on MiniMax especially at high context length due to AMD engineering leadership incentivizing tuning only for short context single turn workloads and ignoring long context multi turn workloads." The 11x claim so far exists only in Monday's post and its video. The August report promised an AgentX update in "three to four weeks," which puts it around September 14 to 21, so the written numbers are not yet due.
What was measured
The test is AgentX, part of SemiAnalysis's InferenceX platform, which measures "multi-turn agentic coding inference" at a context length of 1 million tokens, with sub-agents and a cache hit rate above 95 percent. Context is the information a model keeps in view during a session. Sub-agents are helper agents assigned specific tasks by the main agent. A high cache hit rate means the system reuses earlier calculations instead of redoing them. The test simulates an AI coding agent working through a long session rather than answering one short question.
The model in question is MiniMax M3, a 427-billion-parameter open-weight model from the Chinese lab MiniMax that activates 26 billion of its parameters for each generated token and supports a 1-million-token context. vLLM is widely used open-source software for serving models like it.
The post does not identify the before-and-after results needed to reproduce the 11x comparison. The InferenceX site publishes its benchmark runs through public GitHub Actions, which is where the underlying data should appear.
Why the August report mattered
The August 24 report asked in its title whether Nvidia's CUDA software moat holds up in agentic inference. For AMD on MiniMax M3, the answer was the "horrible" line above. The report also said AMD's own serving software, ATOM, "still has a long way to go to support agentic workloads relative to where vLLM/SGLang are at present." AMD had published its own MiniMax M3 optimization work for the MI355X on July 21, and cited InferenceX as its benchmark source. The weak spot SemiAnalysis found was long conversations with many turns, which is exactly what AgentX tests.
The people running the benchmark also fix the software. Cam Quilici, one of the AgentX report's authors, submitted a fix to vLLM in June that corrected a data-type bug in MiniMax M3's attention on AMD hardware, taking accuracy on one test from about 1 percent to about 96 percent. That was a correctness fix, not the speed work Monday's post describes, but it shows who writes the software for AMD chips. Contributors include AMD engineers, open-source developers and the benchmark's own researchers.
What it means for AMD (AMD) and Nvidia (NVDA)
AMD must convince buyers that its software can match Nvidia's (NVDA) established ecosystem. An 11x jump in 19 days cuts both ways. It shows AMD's software can improve fast, and it shows the starting point was low enough for an 11x gain to be possible. A large gain by itself says nothing about where the MI355X now stands against Nvidia's B200 or GB200 NVL72 on the same test, and SemiAnalysis has not yet said.
US markets were closed on Monday for Labor Day. AMD last traded at $477.45 on Friday, up 4.7 percent on the day, and Nvidia at $230.35, according to Robinhood market data. The next thing to watch is the AgentX update itself, with the numbers behind the 11x.
