Frostbite: A new CPU architecture discovered, designed and verified by AI
119 points by shioyama 2 hours ago | 83 comments

throwaway_new_1 41 minutes ago
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.

reply
sobiolite 14 minutes ago
I think we are starting to see this in some areas: For example, for years desktop apps were built on top of slow and heavy cross-platform frameworks like Electron, because writing them natively was too burdensome. Now we're seeing rewrites in rust using native APIs and toolkits, resulting in lighter and faster apps.

Software engineering as a discipline is also shifting, and will continue shifting, to a process of continuous automated improvement. In the past, test suites were mostly a way to catch regressions. In future, they will be essential to provide a verified baseline of correct behaviour, and then you will have models looped in the background to continually merge performance and reliability improvements. Similarly you will have error and performance logs fed directly into agents to resolve incidents as soon as they arise.

We are in the baby steps of these changes at the moment, but I think in at most ten years, all non-hobbyist software will be built this way.

reply
rickdeckard 21 minutes ago
> "I am interested to have end products with really good UX driven by peak performance and optimizations [..] 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"

AI is not equipped for this. Humans are part of the experience process, and AI doesn't "understand" individual humans better than other humans do.

UX relies very much on human psychology. That's why, despite all the efforts of scoring it, UX is not an objective metric:

1. AI doesn't know what "good UX" is, as this is judged by the sum of humans using it. --> Structured data of actual usage would be needed for AI to ingest, in a way that it can derive intent from every action and identify friction.

2. A "good UX" is not a transferrable metric. The best UX for a ticket vending machine is not applicable to video-editing software. --> AI would need to understand what to apply when, and at which point it needs to evolve its expectation to the user

3. A "good UX" is not an absolute metric. Photoshop/Premiere/... has a UX which was extended and refined over decades, it is considered good by experienced users but still overwhelmingly confusing for beginners. --> Someone would need to scope the AI for the correct target-group

All this is done today by doing extensive user-testing with actual humans. There are people experienced in this who can create better UX-designs from scratch than others can, and still we keep learning something when we put a group of humans in front of it and watch them use it.

"Intelligence" alone doesn't solve this.

I'm reminded of the story of Stephen Hawking's sister, telling how he invented a "better board game" as a child and always wanted to play it with her, which was apparently the most frustrating and boring game she ever played...

reply
swader999 51 seconds ago
Not so. You can time how long it takes, how many clicks to goal, how much of the viewport is shown. There are so many metrics around UX and it can be approached in an objective way. So one UI can be evaluated on another for a given goal. Sick Fable on it and see.
reply
jaapz 13 minutes ago
Then again, often when I specifically prompt my agent that "this works like shit, make it better" - it makes it better. There are for sure best practices that agents are well equipped to handle. Often the solution to make UX better is to just make the process simpler (for the user). Agents are also well equipped to do that.
reply
samiv 6 minutes ago
With "AI" coding tools the relevant metrics are (again) LOC and token burn. But in the real life the fundamental truth is still the same, code is to be minimized.
reply
atlasduo 18 minutes ago
This is why I am not worried about job prospects in software development.

Neural networks are just another abstraction layer that allows you to produce more functionality by typing fewer characters. A very smart, complicated library, if you will.

And just like all abstractions before it, all productivity gains will be gobbled up by capitalism.

Sure, you can use this productivity boost to do more. Some will, and we will praise them for it. But in a world where primary incentive is acquiring as much centeral bank bills as possible is the main KPI, productivity will be used to deliver the same amount of work (or less!), at lower cost.

Until the world in its entirety undergoes a dramatic moral and systemic shift, I sleep well at night.

reply
troupo 36 minutes ago
> 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.

Which part of the entire history of our industry (and other industries) leads you to have a hope like this?

reply
pizza234 50 minutes ago
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.

reply
kragen 14 minutes ago
They're claiming better performance per watt, not better absolute performance, and out-of-order does worse on that metric than the superscalar in-order design they say they beat by 4.59×. That's why, for example, PULPino is in-order.

A third interesting metric is performance per transistor or per area, and generally the best designs there are the long-forgotten MISC designs like the MuP21 and the GA144.

reply
rhdunn 19 minutes ago
If you are interested in understanding how CPUs work there are a number of "build your own CPU" style video series on YouTube:

- Ben Eater's series "Building an 8-bit breadboard computer!": https://www.youtube.com/playlist?list=PLowKtXNTBypGqImE405J2...

- Fabian Schuiki's series "Build a Superscalar CPU" (breadboard and custom boards): https://www.youtube.com/playlist?list=PLyR4neQXqQo5nPdEiMbaE...

- Polymath Unlimited's series "The CPU you can build (Tutorial Series)" (logisim): https://www.youtube.com/playlist?list=PLe8IZRJAn3oD3E2tDhNpI...

- mattbatwing's series "Let's Make a Redstone Computer!" (Minecraft): https://www.youtube.com/playlist?list=PL5LiOvrbVo8nPTtdXAdSm...

reply
petra 46 minutes ago
Yes, their comparisons are weird.

They also compare their architecture simulation language to RTL, instead of comparing it to existing arcitecture sim languages.

reply
chrisjj 39 minutes ago
> Yes, their comparisons are weird.

Generated by ChatGPT, probably.

reply
amelius 15 minutes ago
Isn't this more about work-per-joule than about throughput? At least that benchmark graph is in performance-per-watt.
reply
checker659 30 minutes ago
Maybe it's Itanium reincarnated
reply
motbus3 34 minutes ago
Yeah. Something smells fishy
reply
yvdriess 37 minutes ago
Did they independently rediscover 80's style fine-grained Dynamic Dataflow architectures?

(RIP Arvind)

reply
dwroberts 19 minutes ago
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?

reply
bhouston 19 minutes ago
> 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?

reply
nottorp 11 minutes ago
> It has dramatically improved performance and tends to increase single core around 15% per year and it compounds.

Does it, or is that because of the large caches and memory bandwidth? Honest question.

reply
bhouston 4 minutes ago
> Does it, or is that because of the large caches and memory bandwidth? Honest question.

The burden to prove your "honest question" claim is on you, not me.

CPUs are incredibly complex and many things contribute to sustained improved performance year over year. To simplify it to just one or two things is showing a lack of understanding of the topic.

reply
adastra22 2 hours ago
Why would you hold off evaluation criteria? This isn’t LLM training.
reply
OJFord 2 hours ago
It's similar: they don't want it designing an architecture that overfits to exactly their 25 evaluation workloads, it's supposed to be general purpose.
reply
dapperdrake 46 minutes ago
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.
reply
aurareturn 2 hours ago
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.

reply
hkalbasi 2 hours ago
Why you wouldn't ask the ASI to design a TSMC too? Then the silicon (sand) will become more valuable. Even raw materials may become less important if ASI manages to achieve space elevator or something like that. But my prediction is that the growth of current AI generation is hitting end of the S-curve, and an AI era where AI can design chips in 2 months and fabs in 2 years is not as soon as it seems to be.
reply
aurareturn 52 minutes ago
It's possible ASI can design and build a TSMC as well. But instead of that happening within 5 years, maybe that's 10 years down the line?
reply
barrenko 2 hours ago
One giant all-encompassing loop.
reply
zarify 26 minutes ago
Agent Centipede? Seems on brand.
reply
fedpost 2 hours ago
Much like with agentic PCB design, the HDL isn't the hard part.
reply
dlisboa 57 minutes ago
I’d say TSMC becomes less valuable because there will be multiple fabs trying to meet demand. There’s no way for a 10x increase in fab capability to be handled by just the very few companies that exist.

Also most likely these bespoke chips wouldn’t need the latest nanometer processes to work more efficiently, lowering the value of highly advanced EUV machines.

reply
aurareturn 52 minutes ago

  I’d say TSMC becomes less valuable because there will be multiple fabs trying to meet demand. There’s no way for a 10x increase in fab capability to be handled by just the very few companies that exist.
Fabs follow Rock's law. The cost of a fab doubles every 4 years. https://en.wikipedia.org/wiki/Moore%27s_second_law

Right now, it costs $20b - $40b to create one fab. And that's without the R&D time and cost for the actual node process such as N2.

So could AI recreate TSMC? Maybe in a more distant future. But it's a lot harder to recreate TSMC than to design a chip and send the file to TSMC.

I think it's more likely that AI recreates some chip manufacturing breakthrough than to compete against TSMC without one.

reply
edg5000 22 minutes ago
Maybe Rock's law will change now that AI makes technological development less laborious, which is where the money was going. All non-technical overhead also gets reduced; what a 10-person company can do today assuming they can run their agents 24/7 makes the situation from just 10 years ago look medieval. So we'll see more foundries and more production capacity pop up, nibbling at TSMCs feet.
reply
aurareturn 19 minutes ago
If there is a breakthrough in chip manufacturing, IE, a completely new process, then maybe.

We don't really know that.

We do know that AI is already there or close to being able to design a chip, verify it, and send it to TSMC.

reply
luckystarr 49 minutes ago
These companies operate under a strict risk modeling framework for their investments, trying to model 10 or so years ahead, because in order to stay profitable they have to shell out 10s of billions at a time. So competition may arise, but very slowly. Heavily simplified.
reply
lazystone 52 minutes ago
> there will be multiple fabs trying to meet demand

From what I know, that's easier to say, than to implement. Creating a chip is so complex, that it works for TSMC because of the scale. On lower scale, production won't be profitable.

reply
xattt 56 minutes ago
When design time is compressed to this extent, do subsequent steps that used to be insignificant become the new bottleneck?
reply
petra 2 hours ago
I would add "simulation/emulation platform" as one of those things.
reply
DeathArrow 49 minutes ago
>Model

I think China is working towards making the models commodities.

The real breakers will be access to cheap raw materials and to cheap energy.

reply
pshc 47 minutes ago
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?
reply
WithinReason 45 minutes ago
Wouldn't a SIMD/SIMT architecture achieve similar speedups? There are many ways to parallelise programs. What am I missing?
reply
Tuna-Fish 19 minutes ago
You need either the compiler or the programmer to do the parallelization, and in practice the communication costs between cores can be large enough for the parallelization to not be worth it for simple programs.

This ISA seems to be designed to: 1. make the automatic parallelization easy (with workload split between compiler and hardware) 2. make communication costs between tiles very small, so that even very simple workloads can be automatically parallelized.

reply
enricozb 2 hours ago
In my opinion, the benchmarks chosen look small (e.g. json parsing with 4096 chars, mergesort with 1024 elements, fib(20)).
reply
brcmthrowaway 32 minutes ago
Is this a legit company? Looks like it was founded by 2 interns...
reply
sylware 2 hours ago
Is the core RISC-V ISA not too alien to this µarchitecture?
reply
imtringued 51 minutes ago
Looks like AI doesn't have to fall for C hubris. Honestly, if you want to make it in this new AI age I recommend you to question every single thing that is considered set in stone, anything that is considered "the natural order of things" in software development. Think hard outside the box.

I don't know why nobody is seeing the problems with C. From my perspective it's not useful to think of C as a portable assembler or a thin layer on "top of the machine", because it makes C machines look eternal and timeless, even though we are deep into the post C machine era of multi core processors with cache coherence.

>Mainstream programming languages, when compiled for general-purpose CPUs, hide the dependencies between the parts of a program. Any piece of code may read or write any variable through a shared, mutable state, and so modern CPU hardware assumes the worst case and rediscovers, at runtime, which instructions actually depend on each other.

To this day I still wonder why people aren't seeing the immediate glaring red flag that the conventional sequential programming model presents. To this day people keep building inherently sequential languages and then wonder why parallelism is so hard. It's hard because you made a fatal assumption: There is only one global memory. How exactly is anyone supposed to reason about such an omnipotent and all powerful concept, where everything can potentially affect anything? If you have two cores, the entire idea of a singular memory screams like a bottleneck. Like, it's crazy to me that anyone would think in 2026 "yup, lets build another sequential language that makes our lives a living hell".

>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. This forces CPUs to maintain the illusion that every core sees the same memory, and keeping that illusion consistent and performant requires deep cache hierarchies and coherence protocols that cost area and power, and scale poorly.

Correct. Because everything potentially depends on everything, you need to figure out at runtime when things don't depend on each other, which is kind of insane when you think about it.

The assumption of a single global memory is what doomed C. It's the 21th century. Computers don't have global memories anymore, they pretend they do. You can't fix it. No amount of nostalgia towards C is going to fix this mistake.

Also before someone accuses me of e.g. being a C hater, I'll tell you that Zig's IO Backend parameters are the same type of flaw in Rust that probably will never be fixed. Implicit IO is a different kind of global state that Rust inherited and from the 2026 agentic coding perspective it's a pretty egregious flaw because agents don't care about passing in a parameter, meanwhile from the perspective of PL design, using a parameter instead of a full on effects system has this certain charm / beauty to it, how you don't need to special case everything to get the same results. You just use what is already there.

reply
actionfromafar 44 minutes ago
I'm with you, but I don't see a clear path from "here" (illusion of the same memory) to "there" (each processor has its own memory, I guess). It's really hard to reason about latency, synchronization and communication costs in the abstract.

In a way I see the various cloud computing services as the organically grown (i.e. haphazard and clunky) answer to this problem.

reply
chahmed312 22 minutes ago
[flagged]
reply
ElUncivio 12 minutes ago
[dead]
reply
ytxm 2 hours ago
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.
reply
xyzsparetimexyz 50 minutes ago
One of the worst things about the llm era is that it's very hard to tell the difference between ai-written breakthroughs and ai-written schizoposts.
reply
sph 43 minutes ago
i.e. signal-to-noise ratio has lowered dramatically and news platforms such as this cannot cope, unless they basically turn into clickbait and hype factories.

We used to say ‘show me the code’, we need a bit more proof today. Or just wait it out to see if the authors are still pursuing the same idea 3 months later or have they already moved on.

reply
xyzsparetimexyz 4 minutes ago
Yeah that makes sense to me honestly. Hard to tell at a glance if a fable oneshot is being dumped out, or if the creator has actual skin in the game
reply
Iolaum 26 minutes ago
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.
reply
mentos 13 minutes ago
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.
reply
rossng 2 hours ago
We will continue to invent Transputer-shaped things forever, it seems. Maybe AI actually makes it tractable/ergonomic though.
reply
luckystarr 45 minutes ago
Sadly, Feynman is dead already so he can't be asked for the buffer size again.
reply
einpoklum 2 hours ago
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.

reply
dist-epoch 58 minutes ago
it's pretty obvious their goal is to get aqui-hired by a big lab to work on their chips
reply
FrustratedMonky 52 minutes ago
Yes. This seemed aimed more at a proof that 'we'd like to be acquired'.
reply
badgersnake 2 hours ago
Naming is hard I guess.
reply
s20n 2 hours ago
I read the headline and thought they had built a CPU in EA's game engine
reply
badgersnake 2 minutes ago
If they did EA would surely can it and sack them all before anyone saw it anyway.
reply
smitty1e 2 hours ago
To say nothing of arriving at a "final" computer language via AI.
reply
globalnode 32 minutes ago
hardware slopocalypse inc
reply
raverbashing 31 minutes ago
Discovered? I don't think this means what you think it means

It's not like Intel dug x86 from some mine in Oregon

reply
FrustratedMonky 51 minutes ago
So far this thread is a little breathless about AI.

This looks cool and all.

But, can someone with CPU experience give an opinion? Does this blow your mind? Or is it trash?

Like how the new math solutions when read by a mathematician seem 'ugly', even if technically correct. Does this look like something actually alien and groundbreaking?

reply
rvz 2 hours ago
Once again, you are "not allowed" to invest in the frontier.
reply
Vivek-KY 2 hours ago
[dead]
reply
s08148692 2 hours ago
The future is going to be fascinating

If we have enough fabs theres a world where every single problem has a completely custom chip to maximise efficiency of whatever problem needs to be solved. End to end AI - Ask to solve a problem, AI designs the chips, fab delivers it, AI writes the software in a completely custom language. Nobody will understand how anything works other than knowing that it does work

reply
ywvcbk 19 minutes ago
AI so far can't really do anything that humans can't though? (obviously it's infinitely faster) if building extremely customized chips so far has proved to be suboptimal (even if we partially exclude the engineering salaries) why would AI change that?

This approach also implies that optimization and performance will be come highly important? But it's not (outside of a few use cases) and the software that even the SOTA AI labs are building is often an unoptimized mess just like human written software used to be (if not worse)

reply
Tuna-Fish 15 minutes ago
Right now, the cost that limits doing small custom chips is not designing the chip, it's the cost of making the mask sets for the chips and other foundry NRE, which is well above $50M for a single leading edge chip these days.

This money is generally not spent on things that AI can easily make cheaper.

reply
FridgeSeal 47 minutes ago
Surely this also potentially unlocks the scenario where “good enough” chip fabrication becomes easier, because not everything needs to rely on the single-most-cutting-edge node-fabrication capabilities, because they can now gain more from architectural customisations in return.

Could be verrry interesting .

reply
edwinfredy 60 minutes ago
Might allow further growth in chip performance now that simply making smaller transistors is seemingly approaching its limits.
reply
bentobean 58 minutes ago
Isn’t there an old episode of Star Trek where they visit a planet with a population that has become utterly dependent on a supercomputer? And as a result everyone has become very stupid? Will have to track that down…
reply
chrisjj 36 minutes ago
"The Return of the Archons".
reply
noduerme 50 minutes ago
That's the good old recurrent theme of the lotus eaters, I don't think any Star Trek series has missed taking a shot at it. Weirdly, you'd think that the main computer in Star Trek would have made them stupid, but they override it pretty frequently.
reply
techpression 14 minutes ago
Wouldn’t AI design a completely programmable base that can become any type of chip instead? It seems limiting that you’d need custom chips for different tasks, a chip could morph on demand, even include different architectures that change during runtime dynamically, or even reinvent itself as required.

If we use the very abstract term “the future” that is.

reply
toyg 2 hours ago
That "fab delivers it" does a lot of work, in that concept...
reply
exe34 2 hours ago
Unless it infringes on big tech patents or other wealth extraction measures, then it'll just be computer says no.
reply
chii 60 minutes ago
3D printing is the beginnings of this sort of capability. But current "home" lab 3D printing is incapable of producing chips - may be one day it will.
reply
jqpabc123 30 minutes ago
Modern chip fabrication occurs on a microscopic scale. It requires extreme precision and thus must be done in a highly sterile environment. The slightest speck of dust is enough to ruin the output.

Producing such extreme controlled conditions is unlikely to ever be practical at "home" --- certainly not any time soon. Packaging a nuclear reactor for home use would probably be easier.

reply
tourmalinetaco 2 hours ago
Which is, effectively, magic. A terrifying and yet exhilarating potential.
reply
noduerme 43 minutes ago
It is. In a sense it's very similar to magic.

It's a marvel that you can download a 12Gb file that's basically a neural net smart enough to discuss any age of history, identify photos, etc, on a local chip - all in the space of 3 or 4 high res movie files.

That shouldn't be possible; it means that we've found ways to compress knowledge into spaces smaller than logic would indicate we should able to, and now we have these tiny brains we can take around with us on a laptop that can recite history. When I was a kid there was an electronic encyclopedia you could buy, with condensed articles, and maybe a few megabytes of memory, and that was kind of "magic" ...it made me think of Hari Seldon, or the Hitchhiker's Guide.

I'm not sold on the idea that the compression is working as well as people say it is. But we're compressing on ideas now, not letters or words. Ideas are orthogonal and the damage or loss from compression isn't immediately visible.

reply
chrisjj 38 minutes ago
Uh, how will they know it does work?
reply
atomsandbits 51 minutes ago
Future = AI + Robots
reply