Live data from Hacker News

We got 207 tok/s with Qwen3.5-27B on an RTX 3090

github.com

51–55 of 55 posts

Re: We got 207 tok/s with Qwen3.5-27B on an RTX 3090

#51

"Local AI should be a default, not a privilege: private data, no per-token bill, no vendor lock-in. The hardware to run capable models already sits on desks. The software to run those chips well doesn't." So figure out how to run it on Vulkan instead of requiring the user to be locked into expensive CUDA cards.

You can run pretty much every model on Vulkan, including the Qwen MoE models. You can also run pretty much every model on ROCm, Apple Silicon via MLX, and Intel hardware via OpenVINO. Nvidia got there first, but they're no longer clearly dominant in the self-hosting space, simply because of the high cost. I think Apple probably has the lead there, due to unified memory allowing big models to run without multiple big…

Yes, but they're claiming massive generation speed which you won't get on Vulkan. You won't get it on ROCm on that Strix Halo, either.

It's just funny they talk about vendor lock, and they only support nvidia.

Re: We got 207 tok/s with Qwen3.5-27B on an RTX 3090

#53

3090 is a great value card right now, but in about a year when DDR6 comes out and gpu ddr7. Everyone will be upgrading to this new standard and these 3090s just wont cut it anymore.

That sounds lovely, but it also depends on if any of that memory will make it to the consumer space. Other recent posts imply the ram shortage could be for a couple years.

Re: We got 207 tok/s with Qwen3.5-27B on an RTX 3090

#54

Why did they focus on that particular graphics card and not others, and not common laptops used by developers, or something like that?

The repo is very vibecoded (Claude is co-author, READMEs are obviously AI). This is the output of someone with a 3090 who pointed Claude Code at some research papers and possibly the upstream implementations of these techniques and then posted the output as original work.

Should it matter if its vibecoded or not, if it works?

If the performance is as claimed, and does not impact the quality of model evaluation it seems like it would be worth exploring.

Disclaimer: I have a 3090 and would love to squeeze more performance out of it.

Re: We got 207 tok/s with Qwen3.5-27B on an RTX 3090

#55

Earlier quoted context omitted.

Actually budget friendly is RTX 3060 12Gb. With one you can run 9B/12B models which are fine for text tasks like chatting or summarisation. Not for precision like tool calling or code. With two of them you can run models up to Qwen 27B and 35B with a few-turn context window (8k-16k). Dense at 14t/s and MoE at 68t/s. With three of them you can run 128k context, though you'll need a large format case and the right moth…

This seems quite unlikely. What motherboard are you getting three 16x GPUs on? That alone with the associated sever processor would be more than a used 3090, before even buying the three 3060s. Give full BOM and costs.

I already had the PC. I just mean the extra purchase of the graphics cards.

The motherboard is an MSI Pro Z690-A.

The slots are physical x16. Electronically they are x16, x4, x1 which doesn't harm anything at all.

Post reply on HN