It looks like AI has followed Crypto chip wise in going CPU > GPU > ASIC
GPUs, while dominant in training large models, are often too power-hungry and costly for efficient inference at scale. This is opening new opportunities for specialized inference hardware, a market where startups like Untether AI were early pioneers.
In April, then-CEO Chris Walker had highlighted rising demand for Untether’s chips as enterprises sought alternatives to high-power GPUs. “There’s a strong appetite for processors that don’t consume as much energy as Nvidia’s energy-hungry GPUs that are pushing racks to 120 kilowatts,” Walker told CRN. Walker left Untether AI in May.
Hopefully the training part of AI goes to ASIC’s to reduce costs and energy use but GPU’s continue to improve inference and increase VRAM sizes to the point that AI requires nothing special to run it locally
The ASICs will come as soon as the architecture stabilizes. The problem with building specialized hardware today is it won’t necessarily be capable of running models that come out tomorrow.
With AMD’s IP, they could make a hybrid chip, eg a (for example) bitnet ASIC hanging off a GPU for flexible, cuda-compatible compute where needed.
Nvidia sorta does this now (with tensor cores being a separate part of the die), but with their history of MCM designs, AMD could take it to an extreme.
GPUs are basically halfway to ASICs already. Probably about half of a lot of modern GPU die area is dedicated to or needed to support specialized AI hardware.
I think the next steps are integrated memory and compute, and ternary operations. Integrated memory and compute would probably need specialized hardware, I doubt it would make sense to include anything other than matrix operations and some common AI functions in that sort of processor.
Also, isn’t that just an NPU?
They aren’t specialized though!
There are a lot of directions “AI” could go:
- Is autoregressive bitnet going to take off? In that case, the compute becomes extremely light, and the thing to optimize for is memory bandwidth and cache.
- Or diffusion or something with fewer passes like that? In that case, we go the opposite direction, throw bandwidth out the window, and optimize for matmul compute.
- What if it’s both? In that case, one wants a truckload of ternary adders and not too much else.
- Or what if some other form of sparsity takes over, as (given the effectiveness of quantization and MoE), there’s clearly a ton of sparsity to take advantage of. Nvidia already bet on this, but it hasn’t taken off yet.
There’s all sorts of wild directions the sector could go. Having the flexibility of an ASIC die would be a huge benefit for AMD, as they don’t have to ‘commit’ to any particular direction like Nvidia’s monolithic dies. If a new trend takes off, they can take an existing die and swap out the ASIC relatively quickly, without taping out a whole new GPU.
What about an asic makes it more flexible or easy to swap out? I thought it was just a general term for any application specific chip.
I think NVIDIA’s strategy is to have high throughout and compute for fp4,8,16,32,bf8,sparsity,etc and not really commit to much. I guess if the application is very well known there’s a bit of overhead in that.
I’m not in industry or anything so I’m not very confident in this but I don’t really see how an ASIC would be that different overall
Because it’s a separate physical die.
Taping out, aka simply designing a large GPU chip for production is at least a 9 figure cost. Hence Nvidia/AMD offer a relatively small selection of physical dies in products, as each die has a huge fixed cost. But AMD has specifically taken the approach of splitting up chips into smaller sections, and linking them together by placing them right next to each other, stacking them, and so on.
Hence, if AMD, say, acquire a niche ASIC company, theoretically they can slap a variant of their design next to existing GPUs, or even next to existing CPUs, and have it share the memory bus, general compute, and other functions, without paying the full 9 figures for a massive new chip. There’s still testing costs, but it’s not so prohibitive.
Isn’t this the idea behind tinygrad?
-
Tinygrad is (so far) software only, ostensibly sort of a lightweight PyTorch replacement.
-
Tinygrad is (so far) not really used for much, not even research or tinkering.
Between that and the lead dev’s YouTube antics, it kinda seems like hot air to me.
Ready for pickup in 2-4 days. https://tinycorp.myshopify.com/products/tinybox-red
You’re right to be skeptical, but on the other hand, Geohot has done some incredibly impressive things in the past. Like that time he livestreamed buying a new model of Corolla and updating the software on for the comma.ai self-driving hardware to work for it and immediately took it onto the highway.
That’s a premade 8x 7900 XTX PC. All standard and off the shelf.
I dunno anything about Geohot, all I know is people have been telling me how cool Tinygrad is for years with seemingly nothing to show for it other than social media hype, while other, sometimes newer, PyTorch alternatives like TVM, GGML, the MLIR efforts and such are running real workloads.
-
It’s like bitcoin mining all over again. And providing roughly an equal amount of value.
If you think AI provides the same value as bitcoin you are brain dead
Consider the perspective of the average person: who is gaining value from AI in its current form? It sure as hell isn’t us.
It could be if it’s run locally.
If you agents run on your hardware, navigate crappy apps and websites and such for you, what do you need the corporate cloud for? How can they show ads or monetize you through that?
That’s the war raging right now, open-weights vs closed weights.
Consider the perspective of the average person: who is gaining value from AI in its current form? It sure as hell isn’t us.
Ah yeah, the millions of people using it to generate images, generate summaries, translate documents, clean up photos etcetcetc are getting no value from it at all 🙄
Perplexity received 780 million queries in May, CEO Aravind Srinivas shared onstage at Bloomberg’s Tech Summit on Thursday. Srinivas said that the AI search engine is seeing more than 20% growth month-over-month.
edit: why am I having this discussion, if you don’t like AI feel free to go literally anywhere else on Lemmy
