What the Ryzen AI Halo reviews didn't tell you
Strix Halo ships a 50 TOPS NPU and the reviews quote it like a fact. AMD's FLM runtime delivers 11 tok/s on it, ROCm delivers 41 tok/s, and we drive the same NPU through a fused-layer engine. Every throughput figure here is historical and unsourced, and this post is the record of that gap.
The paper spec
Every Ryzen AI Halo review opens the same way: a 50 TOPS NPU. That number is AMD's paper spec for the silicon, and it is true as far as it goes. What the reviews did not tell you is that on launch day the only software that could touch the NPU was AMD's own runtime. So 50 TOPS got printed in a hundred reviews, and almost nobody had ever produced a token with the thing.
The spec sheet describes the silicon. It says nothing about the path the data has to take to get there, and on this chip the path is the whole story.
Two runtimes, two numbers
AMD ships two ways to talk to the NPU. The FastFlowLM (FLM) runtime, its closed model-serving path, gets 11 tok/s on this hardware. The ROCm stack gets 41 tok/s. Same NPU, same model, four times apart, and the only difference is the software path in front of the silicon.
| Path | Throughput |
|---|---|
| FastFlowLM (FLM) runtime | 11 tok/s |
| ROCm stack | 41 tok/s |
We drive the same NPU through a fused-layer engine instead, one that collapses the dispatch round trips between attention and feed-forward into a single call per layer. The runtime you choose is the real spec sheet. The sticker number never appears in either row.
The part the reviews could not write
There is a reason the reviews stayed quiet on throughput: there was nothing to measure with outside AMD's door. The runtime is 22 .so files of closed code sitting behind 209 xclbin bitstreams, with zero documentation. Until that stack is taken apart, the honest answer to "how fast is the NPU" is "as fast as AMD's closed path allows." Nobody writing a review in week one had a token count, only a marketing sheet.
The record
The throughput figures in this post are historical and unsourced. The 11 tok/s FLM and 41 tok/s ROCm measurements were recorded early, were never attached to a reproducible harness, and are quarantined in benchmarks/latest.json _unverified. The 50 TOPS figure is AMD's published spec, not our measurement. Treat these numbers as context, not evidence.
This post exists because the reviews could not tell you that the number is software, not silicon. We are telling you the same thing, and we are also telling you our numbers are not yet ours. The difference is the part we do not print until we can source it.
The point
A 50 TOPS NPU is a spec, not a benchmark. What you get depends entirely on the software path in front of it, and the path that ships with the box is not the one we use. We will publish throughput when it is sourced and reproducible; until then the record above is the honest one we have.