logoalt Hacker News

Frostbite: A new CPU architecture discovered, designed and verified by AI

562 points • by shioyama • today at 10:01 AM • 298 comments • view on HN

Comments

shioyama • today at 1:43 PM

Wow, I posted this and come back a few hours later, and the comments are... a bit depressing honestly.

Do people here know what interaction nets are? They are a fundamentally different programming paradigm, different from the most basic harvard architecture of the 1940s. Whether you think this particular implementation will genuinely support them, and whether that will bring the performance they claim it does, it another question (I'm not sure). But at least can we focus on the topic?

https://en.wikipedia.org/wiki/Interaction_nets

➕ show 15 replies
throwaway_new_1 • today at 11:03 AM

I am interested to have end products with really good UX driven by peak performance and optimizations which blows my mind because it’s so well perfected as humans weren’t part of its process.

As of now, I don’t see any of that being the goals even let alone the result. Instead what I see is that everyone is in a race to have as much product out as possible by the same standards that applied to products made by humans, but even lesser quality.

I the last two decades we have gotten so used to average or less than average product quality that we are willing to pay 2000$ for a device running an operating system like macOS or windows that ‘works most of the time’.

I hope we have products thanks to AI that compete on reliability of software and hardware to the point that the big tech will have to reinvent their processes just to be able to justify prices.

➕ show 22 replies
jkhdigital • today at 1:50 PM

The title is a bit misleading—humans came up with all the insights and innovations here. What AI has done, which it is exceptionally good at, is take decades of brilliant ideas buried in research literature and synthesize them into a coherently engineered product. It would seem that AI is now doing an end-run around the innovator’s dilemma, or at least part of it, by dramatically lowering the R&D cost of productizing laboratory ideas.

➕ show 3 replies
pizza234 • today at 10:54 AM

I'm quite skeptical; their baseline is a in-order CPU, instead of a modern out-of-order design, and they're testing a small amount of programs.

Also, I'm not an expert in CPUs, but the underlying idea doesn't seem entirely novel - WaveScalar explored dataflow-based architecture 20+ years ago.

➕ show 8 replies
aurareturn • today at 10:09 AM

My prediction:

In the AI/AGI/ASI era, the 4 most valuable things will be:

- Model

- Compute (chips/energy)

- Manufacturers

- Materials

You can already see it beginning at the chip design stage. If AI makes it so that it takes 2 months to design a chip instead of 2 years, then there will way more chips designed to accelerate applications that previously weren't viable economically.

Therefore, it's manufacturers like TSMC that will become more valuable. There will be more chip companies, more designs, more demand.

And in order to make things, you need materials.

In order to design the chip, you need the model and AI hardware.

➕ show 7 replies
burpingtree • today at 7:26 PM

Isn’t this similar to every other alternative architecture that it relies on a “sufficiently smart” compiler to handle the all the dependencies and get increased efficiencies. It isn’t clear to me what exactly is different this time and why it will work better in practice than all the prior attempts.

yvdriess • today at 11:08 AM

Did they independently rediscover 80's style fine-grained Dynamic Dataflow architectures?

(RIP Arvind)

➕ show 1 reply
pona-a • today at 1:52 PM

Genetic algorithms for circuit synthesis aren't a new concept. It found its uses in niche areas like antenna design but wasn't particularly useful for much else. Specifying good utility functions is difficult, and hill climbing them takes impractical resources and time.

The technology may be different but the conclusion is similar.

nerdralph • today at 1:40 PM

Not all CPU/GPU architectures have a single global memory space. AMD GCN GPUs have global VRAM, local SRAM for each compute unit(CU), and a global SRAM that all CUs could access. Neither GCC nor LLVM were great at optimizing OpenCL to make effective use of those different memories, so getting the highest performance sometimes required the use of AMD GCN assembler.

https://nerdralph.blogspot.com/2017/02/inside-amd-gcn-code-e...

https://gpuopen.com/learn/amdgcn-assembly/

nickalaso • today at 4:27 PM

Yep. In 1 year from now, imagine what we will be able to do to improve existing computational power per watt/unit of input material.

Oh boy, its going to be a wild ride, hopefully at the end of it we are all merged with the digital Overmind God and not in a 'I have no mouth and must scream' kind of way. But nothing we can do about it now.

notorandit • today at 2:47 PM

I would like to know about Mr. Torvalds' ideas about this architecture.

He's been quite critical because hw architects don't talk a lot to software developers.

dwroberts • today at 11:26 AM

Feels like the visualization is doing a lot of heavy lifting, because the explanation sounds quite mundane (randomly rewriting interactions between two points in a program graph?)

It is also written as if this is a new idea for programming languages, but modelling things as a graph of nodes is like 101 stuff?

amluto • today at 3:04 PM

I note that the kernels they tested are computation-heavy tasks on essentially fixed- and small-sized working sets, possibly even tasks that are efficiently representable without state that gets used an a priori unknown number of times. (LZ4 decode is a bit of an exception - you don’t know how many times a given position will be copied in the usual streaming decoder, but there may well be a different decoder in which you figure out what future tokens look at an output byte before you emit and discard that byte.)

I wonder how well their design works on the kinds of programs that build large data structures and slowly mutate them, or whether their current design even works at all for such problems.

(Don’t get me wrong: lots of useful problems clearly fit their model. And even if there are problems that don’t, I imagine that much of the compute part of a memory-heavy problem would still work well on something much like their design.)

Roark66 • today at 3:04 PM

That is a very nice use of AI. An exploratory project that would take a team many months and would never get funding in such scope is being done much more efficiently.

As to the architecture itself, it is promising, but it is hard to say if it will not suffer from difficulties in writing software for it.

It reminds me of one British computer that had physical hardware nodes, each connected to few adjacent ones. With no time synchronisation, each running independently, communicating through ports to adjacent nodes only, and so on. It died outside of academia, because of the difficulties in programming it and progress on classic architectures.

As we are close to the limit of miniaturisation and at the same time experimenting becomes cheaper I wonder if novel architectures will have their another heyday.

wcoenen • today at 12:47 PM

The description of interaction nets reminds me of Wolfram's attempts to model physics as "hypergraph rewriting". Is there a link between the two?

metalsiliconYT • today at 6:41 PM

Reminds me of CGRAs/dataflow processors which are what most AI inference accelerators are at some simple level.

adastra22 • today at 10:31 AM

Why would you hold off evaluation criteria? This isn’t LLM training.

➕ show 1 reply
mtimpe • today at 4:06 PM

This is interesting but as far as I can tell there isn't a call stack or program counter. A source level debugger for Vine is a requirement and if it doesn't already exist would represent a substantial engineering project. Also, values can leak if a wire isn't connected properly or a net could get stuck if nodes were waiting on each other in a loop or for a value which never arrives. It's the equivalent of deadlock or a memory leak, and they should reproduce reliably, but they are a new kind of bug class to learn and deal with.

titzer • today at 2:15 PM

Hmm, interesting. Interaction nets are basically dataflow graphs for pure programs. But AFAICT they do not model stateful computations or control dependence for termination. I don't see how they addressed those problems in the IR, but I only had a few minutes to read the first part and skim the second parts.

yubblegum • today at 6:23 PM

Wonder who designed Vine. Is that an AI creation as well?

0o777 • today at 2:17 PM

How is this different from letting coding agents write SystemVerilog (or Chisel/XLS), cocotb, and sby for verification + some formal verification using Lean or TLA+?

dapperdrake • today at 10:58 AM

The frostbite tiles and their work-stealing is reminiscent of a 2d-grid of basic blocks from compiler design. Register renaming or something similar to it also seems to help building intuition about this.

➕ show 1 reply
pshc • today at 10:57 AM

Instruction nets sound intriguing, but it looks like execution is based on rewrite rules. Does it need a lot of copying and marshalling to get work done?

WithinReason • today at 11:00 AM

Wouldn't a SIMD/SIMT architecture achieve similar speedups? There are many ways to parallelise programs. What am I missing?

➕ show 1 reply
atorni • today at 11:47 AM

the title is a bit misleading. humans built the checker and picked the workloads. the swarm did the grinding.

➕ show 1 reply
bhouston • today at 11:26 AM

> Moreover, single-core CPU performance has largely stalled since the early 2000s, pushing people towards large multicore CPUs such as Intel Xeon, AMD Zen, and NVIDIA Vera.

Huh? Have you tried the Apple's A + M-series of processors? It has dramatically improved performance and tends to increase single core around 15% per year and it compounds.

The drive towards multi-core architectures is real but it isn't because single core performance has stalled, it is that we want to scale faster than single core performance increases.

These are basic facts that this article is getting wrong. How many more basic errors are in this?

➕ show 2 replies
K0balt • today at 3:16 PM

Impressive results. In memory compute?

brcmthrowaway • today at 11:12 AM

Is this a legit company? Looks like it was founded by 2 interns...

cwoolfe • today at 4:11 PM

We are living in the golden age of Computer Science.

talon8635 • today at 12:01 PM

Maybe it did, maybe it didn’t. But you can almost guarantee it will when it iterated and tests and changes and iterates and test and changes and…

How many times has a bold and maybe even exaggerated claim come out and people respond “this is false hype”, or “this is misleading”, or “ehh, only kind of in some technical sense, but don’t worry”. Only for continual iteration and breakthrough to have all of us agreeing one week later that it is indeed doing whatever thing and even that this is old news now.

If the frequency increases, that’s the intelligence explosion…

Truly the last invention (for humans)

api • today at 12:52 PM

AI is already superhuman. I have seen this coding.

The big superhuman areas are speed, motivation (infinite), context size, and associative memory.

We are in the “singularity,” which will be a sigmoid, but we don’t know where it will reach its next level.

What does it not seem to have? True volition. Creativity too, maybe, depending on how you define that, which is a can of worms.

➕ show 2 replies
enricozb • today at 10:36 AM

In my opinion, the benchmarks chosen look small (e.g. json parsing with 4096 chars, mergesort with 1024 elements, fib(20)).

sylware • today at 10:38 AM

Is the core RISC-V ISA not too alien to this µarchitecture?

cupofjoakim • today at 1:20 PM

The scepticism among the comments is likely warranted - but i for one really think this area needs more research. x86, ARM and RISC-V are for sure "good enough", but surely those won't be the "final" design of cpu architecture.

➕ show 1 reply
zzzeek • today at 2:00 PM

> We believe AI fundamentally changes these economics. In the search for higher performance and lower power, extreme co-design of workloads, ISA, and hardware will continue to be demanded.

yeah this is why the "AI datacenter" debate frustrates me. People take it at face value that the enormous power demands of AI are immutable - they're thinking in a 20th century way where we're powering like, air conditioners or something that aren't changing. It's not at all a given that AI is going to demand power in a year or two the way it does now, these are problems that should be solved, not used as a reason to abandon the whole technology.

CMay • today at 1:41 PM

I want this to be interesting, but instinct says despite the investment and namedropping on their about page, this is nothing until they show something. The language is vague in the wrong ways. It's either truly great, or they're afraid it sucks, or they need to waffle for some competitive reasons that aren't clear to a reader. It doesn't seem like there's enough fundamental information here to care.

Is it that they don't want to overstate their claims to investors who have to decide whether to actually build these, in case the results they're predicting don't add up?

Either way, we need to be working on EDAI, so if they've got something, hopefully it's something.

sigbottle • today at 2:09 PM

Yo what the heck. That's kind of funny, the only other guy I know working on this were the HVM guys. They were validated I guess?

moron4hire • today at 7:01 PM

Am I the only one who gets really bothered by the use of "by AI" instead of "with AI," in these kinds of statements? Like, the LLM didn't just decide "oh, this looks like a neat thing to work on" all by itself.

ytxm • today at 10:32 AM

Absolutely breathtaking number of claims, and impossible to find the seam between human endeavour and spambot output. Would love a sober overview of this from a suitably competent human, history is littered with promising sounding architectural approaches (anyone remember the Mill?) that went nowhere.

➕ show 2 replies
Iolaum • today at 11:18 AM

Before even reading anything the word discovered on the title triggered me. It's was not there waiting for someone to find it - like electrostatic attraction/repulsion. It was either designed or "statistically and agentically arrived at it" if my high level mediocre knowledge about how people use AI to research those things is even a little relevant.

➕ show 1 reply
barthold123 • today at 12:06 PM

They claim to be backed by $24M seed funding [1]. Interestingly, they seem to have copied the entire approach from Tendrils Compute (even using their Vine & Ivy languages), who recently got $30M seed funding [2]. Are the companies related, or did architect labs just steal their idea and dump it into an LLM?

[1]: https://architectlabs.com/blog/seed

[2]: https://tendrils.co/blog/2026-11-10/30m-raised

worldsavior • today at 11:50 AM

Why does it say "discovered"? What is this, a finding? This is an invention.

smj-edison • today at 3:39 PM

Reading through this it reminds me of all the exciting computer hardware research happening in the 80s, with the rise of VLSI and systematizing chip design. Then Intel ate everyone's lunch, lol.

I think for HPC there's still a lot of untapped potential for chips like these where you're doing primarily "computation" in the sense of crunching numbers. We're seeing that with GPUs, TPUs and with Tenstorrent. For general purpose applications though Amdahl's law strikes again, and the longest serial task is what ends up slowing everything down. It seems modern CPUs are mainly designed around this bottleneck, which is why imo out of order execution will almost always beat in order execution designs for most day-to-day needs.

One other thing I've been wondering about, what would realtime programming look like on Frostbite? Could you make guarantees in execution time? To be fair realtime is quite tricky on out of order cores, but I'd be curious what mental model you'd use for realtime on something like frostbite, or if serial logic still rules.

rossng • today at 10:29 AM

We will continue to invent Transputer-shaped things forever, it seems. Maybe AI actually makes it tractable/ergonomic though.

➕ show 1 reply
sorrywhat • today at 12:56 PM

Cool, the first thing on this page should be a link to an emulator and a compiler. I can generate the pretty animated SVGs myself after that.

mentos • today at 11:31 AM

I'd love to see them run the experiment "Invent a general-purpose computer that exists entirely within a GPU." see if that turns up any potential.

utopiah • today at 2:02 PM

Is it the hardware equivalent to Ginux, the so called not-Linux clone done in 11 days, that's actually not it?

einpoklum • today at 10:39 AM

A new architecture discovered, designed, verified, tested, used, enjoyed, glazed, and eventually forgotten solely by their LLM.

When they actually make a CPU and offer it for use, then we'll talk about their magic LLM-designed CPU.

➕ show 1 reply
mcjenkisns • today at 1:53 PM

Ideas are plenty, let’s see a product

🔗 View 25 more comments