Comment by serbuvlad
1 day ago
You can burn anything* into an ASIC to make it cheaper per-call.
non-backreferencing grep is not very difficult to implement in an ASIC either. But it's probably not worth it because of how relatively rarely you use it and of the data transfer costs.
LLMs are great candidates for ASIC-burning because they're slow compared even to network speeds and run all the time. The issue is that you don't want to burn a specific model or architecture that then becomes obsolete.
So you've got two possible futures, and both guarantee large price drops: (a) LLMs keep getting better and better and better, so ability/$ keeps rising; or (b) LLMs plateau in ability, in which they will start getting ASIC'd.
Grep (or ripgrep at least) is i/o bottlenecked at this point. It's impossible to process data at faster than i/o speeds, since you have to get the data to the processor somehow. That doesn't change whether that processing is grep on a CPU, or LLM on an ASIC.
ripgrep may be.
I did a toy project once implementing a limited version of grep on an FPGA and was able to get some speedup over GNU grep at the time, though marginal.
In any case, LLMs aren't IO bound :))
I am now imagining a future where developers buy fancy "grep cards" for their machines. I don't hate it.
22 replies →
I've evaluated some of these and in almost all cases they failed to outperform Intel's ridiculously optimized CPU regex library, hyperscan.
Right, but I read the comment's point as: being i/o bound is the eventual limit of the asymptote and it's the same one grep has.
GPUs and inference ASICS also have large amounts of high bandwidth memory, plus lots of high speed storage cache, and dedicated very high bandwidth scale-out and scale-up networks. Because they are also often bound by I/O bandwidth.
If your problem is grepping crazy amounts of data, the infrastructure for LLMs isn't a bad place to look for an example.
> plus lots of high speed storage cache
I wouldn't be surprised if that's what Apple is focused on for their next generation platforms – I wonder if more layers of caching between their SSDs and unified memory are on the cards.
i/o could still be sped up though? And I dunno, maybe this actually is an argument for how the llm version could end up being faster, because there is lots of investment in crazy fast i/o hardware and protocols to get the data into the chips.
It's almost a certainty that LLMs have a "core" that will essentially never become obsolete, possibly even 80% to 90% of their parameters. The rules of English and other languages, core ideas in math and science, all of history, nearly all literature, etc. We don't really understand what's going on inside LLMs enough yet to make good use of this, but one day we will have "core logic" neural networks with stable weights burned into ASIC that are doing the heavy lifting, with more dynamic continually-tuned models manipulating the inputs and outputs into those core models. There are also likely stable expert models on topics that don't change much that we could already do this with.
Inference costs cannot keep falling forever, but they do still have a long way to go.
it doesn't work that way though
> LLMs plateau in ability, in which they will start getting ASIC'd.
They don't need to plateu for that to happen. There are companies already building AI on ASIC, and IIRC they were approach 12 months lead time. A 12 months old frontier model (Sonnet 4.5, GPT-5, Kimi K2) for 1% of the price is still a rather good value proposition.
In order for the ASIC to achieve 1% of the price, there'd have to be 99% overhead in GPU implementations which for some reason you'd have to be able to eliminate in ASICs but not in GPUs. That seems rather implausible.
It will be a better value proposition now than it was 12 months ago. It's likely to be better yet in another 12 months. There may be room for a parallel to Moore's Law here.
This I agree with.
Right now, I do actually use OpenAI's gpt-oss-safeguard-20b for somethings, was released 11 months ago, and is $0.075/M input / $0.30/M output now. I could see this model being in fairly widespread use at 10x speed and 1/10th cost if it was introduced today. Meaning, that for some usecases (moderation) i think dedicated chips can pan out today.
But for more general models, its tougher. Gemini 3 pro was launched in November, if ASICs brought it down 1/10th in cost, it would be $0.20/$1.2. GPT 6 Luna is $0.1/$0.50. Luna is better at a lot of things, but not everything. So 1/10th doesn't really make the ASICS investment worth it in my opinion, but if it brought it down to 1% ($0.02 / $0.12) it would be a really compelling model with a lot of use.
BUT, do i think something like Luna is probably generally capable of doing a huge amount of knowledge work. So if Luna came out at 1/10th the cost a year from now, it would probably be compelling for a while.
It all depends on the rate of improvement in cost/capability.
1 reply →
To be clear, I agree with the overall premise of the article!
But I would probably take a long horizon bet that the grep implementation on my machine will remain cheaper than an equivalent ai task, even though I think those ai tasks will become far cheaper over time.
I just think the original comment's model of asymptotic approach is probably more likely to be accurate than the model of the line blowing through this grep-like cost level.
This whole story really reminds me of crypto coins. Like.. going from mining one coin, or lets say token, to millions of fractions like 0.00000000001 bitcoin a week.
There is a reason it took 4 and a half billion years for human intelligence to develop.
Around half of that was going from single celled life to multi celled life.
The complexity gets faster as you get on with it.
2 replies →
???
Yes, because it was a largely random undirected process.
2 replies →
postgres in asic however...
Do you mean most of uses of grep are backreferencing? In my 30-yr career I did not use regex backrefs once except for learning them.
No, it just gets much more complicated to implement grep both in general, and especially in hardware if you support backreferences, since those make it impossible to compile the regular expression into a state machine.