My read is that a number of Desktop units shipped to reviewers over the last weeks. I’ve read the primarily AI performance review provided, helpfully, by Ihl here and elsewhere but haven’t seen anything else.
Have there been other reviews published? Perhaps this thread could collect links to them?
Thanks very much!
I think there is a review embargo until 7th of August
Oh, that’s interesting. Thanks @inffy
Glad it’s been useful. Just for a bit more context, as a rare AMD RDNA AI/ML guy (see my general AMD GPU doc) I was approached a while back to help give the machines a workout on the AI/ML enablement/profiling front. Bugs and reports were were originally published just as “Strix Halo” (that was released in Feb and with lots of mobile and now desktop chips running the APU), but I was given the go ahead to mention that this was on pre-prod Framework Desktop hardware last month (but also requested to leave all non-LLM stuff to the reviewers, which is fine by me, I already have an actual job 
).
BTW, a while back I started (but never finished) this doc LLM Inference Benchmarking Cheat‑Sheet for Hardware Reviewers to potentially help HW reviewers since very them (including most of the people that focus on AI/LLMs, surprisingly, seem to have much idea about how inference works (eg, the relationship between compute, memory bandwidth, memory size, quants, and model architectures have on inference), or the best ways to get repeatable and reliable results, etc.
In general, testing methodology is still pretty rough (for home hardware, I’m not expecting anyone to report TTFT or ITL, or do concurrency sweeps, but most aren’t even running anything very repeatable or relevant to local AI enthusiasts). That’s where hopefully this repo will hopefully come in handy: GitHub - lhl/strix-halo-testing - the llm-bench provides a simple script for bs=1 (the most common local LLM inference use case) that will let any generate arbitrary pp/tg sweeps across different backends and configs for different models.
I’m actually re-running some tests now atm btw - the latest Vulkan (amdvlk 2025.Q2) and ROCm (TheRock 7.0 nightly, rocWMMA HEAD) have shown some pretty decent perf improvements just in the past month or so. I do wonder if any media reviewers will get this testing right, as there are easily 2-4X differences based on backend, compile/runtime flags, and with kernel/driver configs. (Of course, all my testing is for Linux, I have no idea how things are on the Windows side).
Sneaky 
Well, tbf, sort of impossible to submit any bug reports or do profiling without mentioning the chipset! 
Part of the ‘fun’ is when you agree to an embargo it’s a little harder to know how strict that is in any public info/comms about the device under test 
I think I can reveal I’ve been testing more than one board, and as someone who gets the base fundamentals (but AI is far from my day-to-day), I can say testing AI anything is about the polar opposite of testing anything else that you’d see in pop reviews of hardware.
It’s more like enterprise side, where groups like ChipsAndCheese, STH, etc. have their customized (and often proprietary) applications and benchmarking tools, to hit specific targets like memory bandwidth, IO for huge busses, microbenchmarks for specific CPU features, etc.
Most reviewers like to have a tool like Cinebench or Geekbench: “I can run it on anything. I click button, it gives me number. I like that.”
But AI models are not only evolving constantly, the way you run the model on hardware can give insanely varying performance (not to mention there are multiple flavors of models tweaked for different purposes).
Add in the fact that ROCm and CUDA are two different beasts, Vulkan also exists, Intel’s off doing their own thing, Apple’s in la-la land with Metal… I don’t begrudge reviewers who shrug off AI entirely 
All that to say I’ve been hitting some walls with my clustered testing, and I’ve run into many of your posts in my research, so thank you! (I’m upgrading my nodes to Rawhide now and switching gears from Vulkan to ROCm).
“testing more than one board… lah-dee-dah… clustered testing…”
lol
Currently torn between sticking an RTX 4000 Ada SSF into this little <4.5litre PC I already own, and ordering a Framework desktop… so looking forward to some comparison testing.
And now I’m learning that the RTX 2000 Ada is a thing… and is only £600-ish… looking at benchmarks for Stable Diffusion, it seems to have better performance than my RX6800 (or RX 9070XT which I still struggle to get working reliably) so will likely outperform the 8060S …
Now yes, it’s only 16Gb VRAM, so might be less of an option if you’re after large-LLMs (LLLMs?) … but might be an easier pill to swallow my use case.
Not the Framework, but the HP Z2 G1a Mini - https://www.youtube.com/watch?v=wCBLMXgk3No
Let the review fun begin!
The Framework Desktop may be the easiest PC you’ve ever built | PCWorld
It’s just a basic overview of the build process, but it doesn’t have any performance data yet - sounds like that stuff will be coming soon in a separate article. Hopefully we can start seeing the review floodgates open soon! 
Come on Media. I want some reviews. I’m so excited.
Looks like the NDA has been lifted now:
David Heinemeier Hansson (also known as his initials: DHH)'s blog about Framework Desktop is also available now.
I had a glance at the reviews. As there seem to be some of the reviewers here in this forum, who actually have reviewed the Desktop, may I ask something regarding the Bios? Only LevelV1 tech showed the bios a little bit but did not go into detail (not that it looks like there are too many options anyway) so it is still not clear to me if the bios allows for different power settings of the APU. Is it always set to those 120W/(peak 140W) or is there any sort of “silent” or “power saver” mode or anything that reduces that power limit, for those who’d like to tune for higher efficiency or even more silent operation?
And if there are no bios settings to that end, is it possible to do something on OS level in Linux, for example reducing max frequency of the APU (and thereby indirectly also reducing max power draw)?
If someone can give me some insights on that I’d be very much obliged. Otherwise I have to figure it out myself as soon as I can get my hands on the mainboard.
ryzenadj has support for Strix Halo so you should be able to have pretty good control over all the various power limits. I’ve done some past writeups on how Ryzen APU power management works that may be of interest.
Cool. Sorry for those noob questions but should this work independently from the question how locked down the bios/uefi of the Desktop is or isn’t, by merit of having a Strix Halo APU? If so that would be great.
