24 comments

[ 3.4 ms ] story [ 45.0 ms ] thread
I had been hoping that these would be a bit faster than the 9950X because of the different memory architecture, but it appears that due to the lower power design point the AI Max+ 395 loses across the board, by large margins. So I guess these really are niche products for ML users only, and people with generic workloads that want more than the 9950X offers are shopping for a Threadripper.
Across the board, by a large margin? Phoronix ran 200 benchmarks on the 9950x vs 395x max and found a difference of less than 5%. Not bad considering the average power use was 91 watts vs 154 watts.

If you need the memory bandwidth the strix halo looks good, if you are cache friendly and don't care about using almost double the power than the 9950x is a better deal.

I've ran a comparison benchmark for the smaller models https://gist.github.com/mhitza/f5a8eeb298feb239de10f9f60f841...

Comparing it against the RTX 4000 SFF Ada (20GB) which is around $1.2k (if you believe the original price on the nvidia website https://marketplace.nvidia.com/en-us/enterprise/laptops-work...). Which I have access to on a Hetzner GEX44.

I'm going to ballpark it between 2.5-3x faster than the desktop. Except for the tg128 test, where the difference is "minimal" (but I didn't do the math).

I see Wendell of Level1Techs combines the two in his video on this system.

Theoretically you can have the best of both worlds if you don’t mind running an Occulink E-GPU enclosure

https://youtu.be/L-xgMQ-7lW0

Thanks for the excellent writeup. I'm pleasantly surprised that ROCm worked as well as it did — for the price these aren't bad for LLM workloads and some moderate gaming. (Apple is probably still the king of affordable at-home inference, but for games... Amazing these days but Linux is so much better.)
I was about to be annoyed until you said you got preprod units. I guess I'll have to build on this when my desktop shows up.
for those who are already in the field and doing these things - if I wanted to start running my own local LLM.. should I find an Nvidia 5080 GPU for my current desktop or is it worth trying one of these Framework AMD desktops?
The short answer is that the best value is a used RTX 3090 (the long answer being, naturally, it depends). Most of the time, the bottleneck for running LLMs on consumer grade equipment is memory and memory bandwidth. A 3090 has 24GB of VRAM, while a 5080 only has 16GB of VRAM. For models that can fit inside 16GB of VRAM, the 5080 will certainly be faster than the 3090, but the 3090 can run models that simply won't fit on a 5080. You can offload part of the model onto the CPU and system RAM, but running a model on a desktop CPU is an enormous drag, even when only partially offloaded.

Obviously an RTX 5090 with 32GB of VRAM is even better, but they cost around $2000, if you can find one.

What's interesting about this Strix Halo system is that it has 128GB of RAM that is accessible (or mostly accessible) to the CPU/GPU/APU. This means that you can run much larger models on this system than you possibly could on a 3090, or even a 5090. The performance tests tend to show that the Strix Halo's memory bandwidth is a significant bottleneck though. This system might be the most affordable way of running 100GB+ models, but it won't be fast.

Jeff - check out the distributed-llama project...you should be able to distribute over entire cluster
The Framework Desktop has at least two M.2 connectors for NVME. I wonder if an interconnect with higher performance than Ethernet or Thunderbolt could be established using one of the M.2 to connect to PCIe via Oculink?
Kinda bummed, I get why he used Ollama but I feel like using llama cpp directly would provide better and more consistent results
no compilation tests?
> usually resulting in one word repeating ad infinitum

I've had that using gemini (via windsurf). Doesn't seem to happen with other models. No idea if there's any correlation but it's an interesting failure mode.

I've seen that occasionally with one of the deepseek models when using the default Ollama context size of 4096, rather than whatever the model's preferred context size was.

After having that happen, I switched my stuff to check the model's preferred context size, then set the context size to match, before using any given model.

Those numbers are better than I was expecting.
Jeff! Someone needs to make Framework MBs work in a blade arrangement, and you seem to be the likely person to get it done.
Apparently the frameworks desktop's 5g bit network isn't fast enough to scale well with LLM inference workloads, even for a modest GPU. Anyone know what kind of network is required to scale well for a single modest GPU?
Mem bandwidth sucks compared to Mac Studio ultra 3. And you cant add gpus easily although as an apu it is impressive and way better than nvidias gold box. Wendell said it better. Im waiting for the Mac Studio ultra 5
> For networking, I expected more out of the Thunderbolt / USB4 ports, but could only get 10 Gbps.

I really wish we saw more testing of USB subsystems! With PCIe being so limited, there's such allure to having two USB4 ports! But will they work?

IIRC we saw similar very low bandwidth on Apple's ARM chips too. This was during M1 or so; dunno if things got better with that chip or future ones! Presumably so or I feel like we'd be hearing about it, but also, these things can just go so hidden!

It was really cool back in Ryzen 1 era seeing their CPU get some USB on the cpu itself, not have to go through the IO/peripheral Hub (southbridge?), with its limited connection to the CPU. There's a great up breakout chart here, showing both the 1800x and the various chipsets available: relishable data. https://www.techpowerup.com/cpu-specs/ryzen-7-1800x.c1879

I feel like there's been some recent improvements to USB4/thunderbolt in the kernel, to really insure all lanes get used. But I'm struggling to find a reference/link. What kernel was this tested against? If nothing else, it's be great to poke around at debugfs, to make sure it's getting all the lanes configured. https://www.phoronix.com/news/Linux-6.13-USB-Changes

Can you imagine a Beowulf cluster of these?