Hacker News Reader

Comment by johnvanommen | Hacker News Reader

Possibly the best deal there is

I really need to shut up, or bite the bullet and by one.

If you graph the tokens per second on the 5090, your jaw will hit the floor at how cheap it is

With only 32gb of vram, you can only run small/quantized models, in which case what's the point? At $4000, that gets you 20 months of 10x claude or chagpt subscriptions, which provide far better models. You'd need some use case where you can tolerate worse models, and use a steady supply of them. That doesn't match most people's usage patterns.

Or you want to process private data or don’t have reliable connectivity. There are a few more reasons for local models I think.

Also, electricity isn't free.

With enough solar panels it is!

Not quite.

Free for approximately 8 hours (assuming perfect weather conditions) and excluding unit cost and maintenance cost.

It has a cost.

My area has a net-metering plan available, so you can send any surplus out to the grid to offset energy pulled from the grid, essentially treating the grid like a large battery. That can extend the 8 hours into full 24-hour coverage with enough panels.

If you can do what you need with qwen3.6-27b, it starts to look really interesting. That model is crazy good for the size, but it's a pain tweaking the params to run it on a 4090 with decent context and decent token speed. A 5090 looks tasty from that point of view, and only more so if you think in terms of the probability of that model being roflstomped by something in the same weight class in the next couple of years. I reckon that probability is significantly non-zero, but fundamentally it's a guess.

>If you can do what you need with qwen3.6-27b, it starts to look really interesting.

What's the use case here? Churning out massive amounts of slop code through autonomous agents? Running openclaw 24/7? I think the proliferation of codex and claude code, compared to any of the cheaper open models suggests that at least for most software development, the 50-75% discount of open models isn't worth the hassle of the decreased intelligence.

I think there is a reasonable basis for taking a gamble that small models capable of fitting on a 32GB card will continue to advance over the next 5 years and eventually approach Gemini Flash 3.5 / Sonnet 4.6 levels of capabilities, which I would consider to be past the threshold of “probably worth the cost and hassle of running 24/7” if the upfront cost of the hardware was palatable.

My use case would primarily be in search, integration, and indexing other software projects with my own, as well as transcription/indexing of interesting video and audio content (eg Dwarkesh interviews) that I don’t have time to watch but want to easily search and apply to my projects, and search/indexing for useful information from things like Linux kernel and security mailing lists. Basically there is a lot of stuff that, if the cost were low enough, I would point a reasonably intelligent AI at to distill out useful information and apply it to my projects, or just cherry pick the interesting things out and surface them to me so I don’t have to wade through all the mundane stuff and man-made slop getting in the way.

>My use case would primarily be in search, integration, and indexing other software projects with my own, as well as transcription/indexing of interesting video and audio content (eg Dwarkesh interviews) that I don’t have time to watch but want to easily search and apply to my projects, and search/indexing for useful information from things like Linux kernel and security mailing lists. Basically there is a lot of stuff that, if the cost were low enough, I would point a reasonably intelligent AI at to distill out useful information and apply it to my projects, or just cherry pick the interesting things out and surface them to me so I don’t have to wade through all the mundane stuff and man-made slop getting in the way.

All of that feels like something that a $20 chatgpt pro subscription is for, maybe with slightly better tool use capabilities. There's no way that a $4000 purchase on a GPU would ever be worth it if all you're doing is running a handful of queries per day.

It would require much more than a couple of queries per day, I want to basically do bulk ingestion and search/evaluation/integration across tens of thousands of videos and software projects (if it were cheap enough and smart enough). It would basically be setting up and operating a pretty large data ingestion and coding agent pipeline, which I would want to itself be mostly automated.

It’s ok if you don’t want to do the same kind of thing but I find it weird how dismissive so many people get about wanting to use LLMs for large projects, or how anybody who says they’re using them for these kinds of things (I’m doing similar for other stuff) gets challenged on what they’re doing it for.

I don't have 5090, I have 395+ and I use for gpu assisted OCR, embeddings vector, speach to text and etc. I have a freedom of using a large library of various models and I can fit a lot in 128gb.

I don't use it for coding, I have $20 Gemini, $20 codex, etc.

But then I got the framework board for $1700, now it's $2700

The 5090 is crap for inference. Unless you like dummy models, sure they will run at light speed. All the rage is MoE with 500B-1T weights nowadays.

MoE is fine. You can put the shared weights on the 5090 (will fit handily even for the largest models) and expert weights on CPU, possibly with weights offload from storage.

Even if you could fit a 500B model's expert weights in very fast system RAM, it would run so slow as to be useless.

That's really only "useless" if the only thing you care about is a quick real-time response. Contrary to common perception, MoE models do benefit from batching requests together even when run on a single node, you just have to ensure you have at least ~5 parallel requests in flight (and that's for the very sparsest models) to really see the aggregate benefit.

(Intuitively, that's because the issue of whether any active weights are being shared among requests - thus, any memory throughput is being reused - is a generalized birthday problem. That's why even having a few parallel requests is quite effective. Especially since the "random" choice of experts happens anew at any single layer, so there's a lot of independent samples.)

This is just wishful thinking.

For prefill, it's really easy to batch MoE and get really good tk/s, even on a single stream.

For decode, you will run into the problem that:

1) you need more parallel requests which means more memory for context

2) 5 requests will not give you very much expert overlap on parallel requests

You don't need "very much" expert overlap to see aggregate gains at scale, you just need some of it; that's where the "birthday" framing becomes relevant. Memory for context is an issue, but recent models like DeepSeek V4 use very little of it even at relatively large contexts.

>You don't need "very much" expert overlap to see aggregate gains at scale, you just need some of it

I'm not sure what you are claiming. Decode is bottle-necked by memory bandwidth. To see a speed up of 2x, you have to ensure each expert weight memory fetch can be used by 2 parallel streams. What exactly is the average factor you are claiming for 5x parallel streams (due to "birthday paradox" factors)? The Birthday paradox isn't really relevant here. It's about coverage, not parallelism.

> Memory for context is an issue, but recent models like DeepSeek V4 use very little of it even at relatively large contexts.

This is not true.

An aggregate speedup of 2x is a lot, we don't need that in a local context. Local hardware is heavily constrained by power and thermals, not just bandwidth; so all we really care about is raising compute intensity for decode a little bit to relax the memory bandwidth constraint. The average factor will depend on just how sparse the model is and how far you can push parallelism, there isn't just one single answer.

But you won't see 2x expert re-use, the speedup with 5 streams will be tiny.

Reply on news.ycombinator.com