Hacker News new | ask | show | jobs
by huragok 19 days ago
If I had the capital I’d make an household inference appliance.

No peripherals except Ethernet, integrated compute (cpu+gpu+mem) and secondary storage (+mobo, psu). No accoutrements, just the minimum amount of hardware to run a model as a utility.

Even the appliance faceplate would be a display showing stats like an old HiFi stereo.

Edit: something like a series of modules consisting of a RISC-V CPU + Vortex GPGPU + memory

11 comments

You're describing the mac mini/studio with some facelift.
Yeah but like running linux hopefully
so you have invent unified memory for linux first because that’s the limitation today
Fairly sure most iGPUs these days are zero-copy and can dynamically allocate memory so what does "unified memory" mean to you exactly? A wider bus would be nice but it's not exactly a groundbreaking new invention.
I was actually pretty far off:

> Unified memory in Linux creates a single address space accessible to both the CPU and GPU, eliminating the need to manually copy data between system RAM and video memory. It is enabled via NVIDIA's CUDA, AMD's ROCm/HIP, or generic kernel-level Heterogeneous Memory Management (HMM).

So it does exist and is available for platforms that matter.

It is interesting how apple claimed that "unified memory" is something special, and ppl believed them.

Intel and AMD had been doing this for years already, and had linux support for it from day 1.

"Just". And then GPUs, and RAM? And cooling? Will you really appreciate it when sitting right next to it?
Raspberry Pi and other SBCs, Android phones and practically all of the embedded devices with a display and microprocessor.

All have unified memory. Linux runs just fine on all of those.

Ah ok. I replied to ~45 minutes stale page.
Absolutely, but not under the control of Apple.
Isn't that what what George Hotz is doing over at tiny? https://tinycorp.myshopify.com/
Yes, but for inference. 45k is so far out of the budget of a professional unless you earn ridiculous money and have no dependents.
A professional AI engineer? Earning hundreds of thoundands of dollars a year?
Who lives on one of the coasts where that job likely requires them to be, where rent is $3-5k and mortgages within spitting distance of that, sure.
It could heat your home in the winter and your pool in the summer.
Is warming a pool in the summer real where you live?
Yes. Solar thermal heaters on the roof are common in Florida and other parts of the south. Some people also use heat recovery devices attached to the AC condenser. Further north I've only seen natural gas heating (e.g. in very rich NYC exurbs). The amount of shade over the pool has a big effect.
I think the closest to that in existence is the LLM ASIC designed by Taalas:

https://taalas.com/products/

Unfortunately their chatbot, while amazingly fast, doesn't know anything about the company running it.

Anyway I wouldn't mind an ASIC running a diffusion language model locally. Even if eventually it would become dated. Beats outsourcing all that to a company that's running on VC money which in the future might either perish or worse - dominate the market and charge whatever they wish.

Is that the nvidia spark?
Yes, and a lot of others.

A bit too expensive for a home appliance though, isn't it?

I'm keeping an eye on Tenstorrent for this. Pricing seems like its going to end up being in between a super memory dense unified memory platform, and a purpose built GPU.

Definitely on the edge of what would make sense at home, but its interesting.

I lasted about 25 seconds on that site. Way too much friction for me to endure just trying to figure out what it is
Yeah, I don't know who thought that website was a good idea.
"Login to order"

That's a new one.

I feel like this is some sort of satire? There's no actual information or substance to anything on any page of that site.
the pheriphels support, or the appliance faceplate is tens of dollars, that not where you make the saving

95% of the price is going to be in GPU+CPU+RAM

Sounds like reinventing the home server.
build a Xeon / epyc 4u server. 12 channel ram.
Yes, just a big cool Cerebras wafer for the closet please.
A single wafer comes with 44GB RAM, the reason why Cerebras is so interesting is because the architecture scales up to 1.6PB RAM.
Central heating / thinking.