(architectlabs.com)
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?
> Another issue with multi-core CPUs is the difficulty of programming them. Debuggability is a major issue, and entirely new classes of issues arise, such as race conditions. However, in interaction nets, all active pairs are independent, and so can actually be executed in any order. This property is known as confluence, and means that no matter the order in which things are executed, the result will be the same. This makes debugging straightforward and eliminates issues like race conditions. This allows Vine to be parallel out of the box; the programmer does not have to do anything special with threads, processes, CUDA kernels, etc. to get parallel execution.
Seems to describe something analogous instruction level parallelism, which traditional CPUs made this millennium have as well via instruction pipelines, branch prediction, and speculative execution. Some of which turned out to not work too well so now we generally run hobbled CPUs with Spectre/Meltdown mitigation enabled.
The reason why we tend to reach for difficult abstractions for core level parallelism is that most automatic/safe approaches tend to stumble pretty hard on Amdahl's law, where most parallel code is consequently almost entirely designed around minimizing critical sections.
I don't fully see how inets would resolve Amdahl's law.
So there is 'no sharing of state' baked into the computational model by design, enforced by this very simple and elegant design choice.
On the optimistic side, this looks like a kind of "bigger step" in the direction that GPU/shader programming took, where you change the way you think about and break down a problem to get insanely fast results. Feels like if you target this interaction net structure and give it to this CPU that uses them natively, we could amplify the speed of computations to possibly allow for things that are currently way to slow to do. Realtime physics sims with magnitudes faster steps. That's pretty cool. And all of the downstream effects that would allow for, in product design, scientific modelling, etc. And that's just one train of thought.
On the pessimistic side, though, I have no clue how to "program for an interaction net". Seems like the authors took a "you don't need to, an LLM will do it for you" position, and I can get behind that, but that's a pretty huge step in the direction of "Natural Language Programming" in that I literally can't program it without using an LLM and natural language, because I don't understand the underlying details and probably won't have any structural support to understand it for a while, as the tooling slowly rolls out and gets iterated on, until it can do something we can recognize as "Debugging". I imagine there's some kind of "program it in C, and translate it" thing, too, but it feels like the point here is not to have me be able to debug it, and a detour through C just wastes LLM tokens.
So those are the initial impressions, but then I'm making a TON of assumptions in both sides of that. So I don't even know if my initial impressions are conceptualizing the right paradigms, much less if they are worth discussing with people who are seriously considering this topic. Only speaking for myself here, but at least my own silence is due to not having enough information myself, and not seeing enough in this paper to interest me in research in the topic just yet. Maybe if they've got something I can play with, I could get my head around it more. But, for now, it's just neat.
I mean, it's basically what out-of-order CPUs already do: read lots of instructions in at the same time, dispatch instructions as their inputs become ready to execute.
Sort of the reality of modern processors is that we have the units available to do the work, we just don't have dispatch logic to keep them fed with work, given how slow memory is. When you have very regular parallelism--dense linear algebra is the canonical example here--SIMD works very efficiently to minimize the dispatch logic needed to keep units fed. When you've got irregular parallelism, dataflow architectures seem to work well at first, until you start trying to make actual chips and you find you've been consistently underestimating how much overhead your dispatch logic is actually chewing up.
If it is true that what is offering is some kind of "super-simd" processor that is built in a way that everything run through it must be similar to simd style code, but because it's hardware, it will run it much faster, then my questions are definitely targeted toward what the speed up is vs a highly-performant SIMD application. I would hope that their "coming next" stuff includes comparisons against the types of SIMD-boosted dense linear algebra you referenced. What is the speed up compared to that. And if it's less than (arbitrarily, say) 10x, how much can be gained by just having an LLM target breaking more stuff down in to multiplex-able instructions? But then, if it's more than that, my questions start turning towards "how much would it take to prototype" and "how much has to be changed from current tooling to produce the chips at scale" and "what is a good showcase target for an initial run? what would be something that could be made actually (in real dollars/hours) cheaper/faster/better by using this chip?"
if it is, that has been tried before and failed (with itanium/etc), because the optimization burden on compilers means that they never end up getting the projected speeds over conventional execution
I.e. they are on the same level as Turing machines and Lambda Calculus. Arguably they are more elegant than LC as beta-reduction doesn't look like a primitive operation.
E.g. if try to 'run' Turing machine using pencil and paper, a primitive step is to update one step. With IN one step can also be 'local'.
OTOH LC beta-reduction might be impossible to do in-place - you might need a lot of new pages to do just one beta-reduction.
People had a lot of novel ideas for nomenclature back then in the 1940s, that just don't work 80 years later, it's a shame they still clung on to it in 2026.
Even if people know they are it doesn't help that a good chunk of the article is dedicated to their AI methods, and there are very little technical explanations of how this actually works.
[1] https://en.wikipedia.org/wiki/Dataflow_architecture
Every step there is local, every dependence is explicit and every value has only one owner.
> same as interaction nets, no?
Exactly my point. But well researched and has compilers from high level languages (Id and some others). > except they are strongly confluent and have incremental sharing
Data flow is an inverted representation of data demand (lazy evaluation, which is strongly confluent) [1] and I do not see why it cannot (be constrained to) have strong confluence.[1] https://dl.acm.org/doi/pdf/10.1145/93548.93578
Have authors compared interaction nets to the data flow machines? No.
They did not even compared their results to the Transport Triggered Architecture [2] automatic core construction workflow.
[2] https://web.archive.org/web/20051125195128/http://ce.et.tude...
In some cases (FFT), TTA's automatic synthesis was able to provide power-per-operation reduction of three orders of magnitude, when compared to specialized ASIC hardware.
Also, the article title is one that heavily emphasizes the AI autoresearch aspects, and doesn't mention interaction nets at all, so it makes sense that people would go into the article primed to think about the autoresearch aspects.
If this stuff works, it'll surface first as some PCIe extension device that work can be dispatched to when it makes sense. Then maybe it'll get integrated into SoCs and be exposed via new CPU instructions or devices. Perhaps compilers can then start to identify blocks of code that can be reduced to interaction nets, compile the net and then run them on Frostbite style chips opportunistically.
Llms are pretty good at porting software.
I think AI sentiment is charged and skews negative amongst culture, but very negatively charged amongst engineers because of the "impending doom" and how our jobs are obviously changing. The craft is arguably gone. Interesting news just becomes another rung in the ladder of job insecurity.
Nice to see you on the interwebs Chris-san :)
If you can program it in rustish language with the intermediate representation, that makes everything portable to it without having to figure out the parallelism, potentially.
Very impressive and potentially revolutionary. Micron or Samsung needs to see this, they are having to reinvent compute anyway.
> without having to figure out the parallelism...
Parallelism can be regular (matrix or tensor operations are examples of that) and irregular (Barnes-Hut as a very simplified example). To extract something like Barnes-Hut from the code with quadratic complexity, one have to make some interesting assumptions. E.g., one has to figure out something.One may think that regular parallelism is a solved problem, but even nested loops are not always parallelizable without figuring out how to make them to work in parallel.
Then there comes Microsoft Word. Can we make a CPU where Microsoft Word gets executed with maximum IPC achievable? If you do not like MS Word, take a look at Linux Kernel and userland programs as a whole.
My question: if these new developments roll out steadily for enough years, what does that mean for nvidia or amd or intel? Do they benefit? Probably? They already have the manufacturing pipelines
Several new kinds of architectures have been tried: the iAPX 432, the Mill, the Reduceron, but things just keep coming back to the out-of-order modified Harvard (a.k.a. fake von Neumann) machine being the most efficient general-purpose computer in our laws of physics.
Really? They let a swarm of AI agents loose and you're sure they didn't draw upon more than 4 references?
Did they just ignore Taelin and HVM?
This is a very interesting project, but unfortunately at this point in time anything with “AI” in the title invites too many hot takes and personal opinions about AI that are unrelated to the story.
On Hacker News (and Reddit and every other social news site) the title drives the comments more than the article content because, frankly, most people don’t even try to read the article. If you had submitted this with a headline that didn’t mention AI and instead said something about interaction nets the tone would probably be very different in the comments.
Please don’t be discouraged, though. This content is very interesting to many of us, but it takes time to read and appreciate it, whereas firing off some “AI bad” hot takes in the comments without reading the article is an easy way to get some upvotes and ignite large comment chains without reading the article. That asymmetry always biases the comments in one direction, but don’t take the comment section bias as a sign that it wasn’t appreciated as a submission. It’s currently the #1 story which is a much better signal of interest than the comments.
In short, how can we be sure it will work and be stable, and that we're not just watching another AI hallucination?
Will be cool to see if Linux can run on a QEMU or FPGA based implementation of this and how the efficiency numbers look then because it's not quite as interesting if it's not general purpose computing ready.
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.
I think that describes the entire lifespan of consumer computing.
I’m old enough to remember Windows 98 running on a Gateway PC. That cost $2000 (without taking inflation into account!) and worked much less of the time than today’s OSes do. Modern OSes are frustrating in various ways but I think it’s easy to take for granted how stable they are compared to the past.
A lot of this is simply because of microsoft firing all their QA and pushing it onto customers, which hasn't really worked all that great
I'd love for Apple to stop releasing new OSes every single year in lockstep with iOS.
Snow Leopard was the peak for me. Everything since has been a regression trying to merge iOS with macOS when they have completely different ui/ux paradigms.
Partly this was too "encourage" adoption of newer, faster, better machines, partly it was to ensure when that wave of fresh new hardware hit the streets, Windows was ready to take full advantage of it.
I know windows 2000, XP, etc followed this strategy, I think 98 overshot, which required 98SE to do some repair
What does that even mean, other than "faster"? Windows 11 created a mountain of e-waste, consisting of perfectly good hardware.
Why does Windows 11 take several seconds to open the calculator on my 2025 work pc, a 6C/12T machine with 16 GB RAM? Do I need next gen hardware just so Win11 stops wasting my time?
This wasn't a given for XP.
That is, literally, what that means, yes. That was quite obvious from the context.
I'm not sure if you've used any "nextgen" stuff lately but faster is not how I'd automatically characterize them.
My guess? Its probably some slop-infested Electron app that has to spool 1GB of garbage from disk. And, you know, "copilot" crap.
Should'a been a 100KB application, instant open, and just works.
The answer is almost certainly because it is a "work PC". They are typically loaded up with so much shitty monitoring software with hooks in everything, making everything slower than it should be.
Why couldn't a more streamlined, simpler, less resource-hungry system run much faster on "newer, faster, better machines"?
> to ensure when that wave of fresh new hardware hit the streets, Windows was ready to take full advantage of it.
So why doesn't Windows take advantage of it?
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...
The idea of removing humans from the loop entirely feels like some sort of AI apocalypse.
1) First-mover wins: burn VC capital to get a product out free
2) Establish brand and market dominance
3) Integrate advertising to keep the product free and grow the user base
4) Crush competition and capture users due to "mind share" and "network effects"
5) Enshittification until no one is happy with your product and making a replacement is impossibleAt the end of the day, in a well-policed society "profit extraction" == "what you can convince people to willingly spend their money on"
Turns out people are generally willing to spend their money on ads to grow their own businesses, and are generally not willing to spend their money on un-enshittified software replacements for what they're used to.
You might not agree with their choices but that's just how people are. I don't think it would be a better world to force them to choose differently.
2. I (like most people I feel) am generally very on board with political action to reduce fraud & coercion. I'm not on board with "let's just overthrow capitalism", which just introduces coercion on people's preferences, rather than remove it.
The challenge is more about providing actual, objective data to learn this.
The LLM will not be able to decide what is "good UX" and "bad UX", it will have to ingest this information in vast detail from judgement provided by humans, which is also varying alot.
Case in point: Coming from that field, I personally don't have a good user experience using iOS, a majority of end-users will strongly disagree.
An LLM would need to be trained to handle both cases to judge the same human interface, so it would need alot of data to understand both views...
(also: UX is not UI, but I guess this goes without saying...)
A perfectly scoring vending machine UX applied on Adobe Premiere will tank the UX for Premiere users.
We respond with delight or disdain based on how they feel.
The only way for AI to know how humans feel about something is to be trained on what we’ve already said about it.
Neither do ~80% (or more) of software developers.
~80% (or more) of UX designers don't know what "good backend design" is.
Even worse: ~80% (or more) of frontend developers don't know what "good backend design" is.
Not sure what's the point here...
"AI does X kinda bad" "Well most humans do X even worse so shut up and eat it up"
That is, roughly, how all automation works: You build the machine that does the humans job, barely, badly. Then you improve the machine. Producing shitty output on that journey is better than producing no output, if the domain allows for it. You learn more.
Frequently automation happens only after someone realizes that to automate the process, the process has to be completely reinvented.
I’d say that the majority of successful automation doesn’t follow an iterative path starting from build anything that does exactly what a human does because human physiology isn’t easily replicated with rotational motion.
with current gen frontiers, the limitations seem to be requiring a human in the loop, which is quite ironic
That’s not how automation works. The first thing you build is already taking care of a cumbersome part of the human jobs, even if limited in scope. The improvement part is in reliability and maintenance. Because the core problem is already solved. All that’s left are engineering issues.
So even if the first contraception is horrible, it should be clearly visible which immediate need it intents to solve, even if the full scope of changes is not known yet. But with those, there’s no need behind them, just imagined problems.
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.
It was not. Just by using QT you were on the fast track to support multiple platforms. And by 2010, most companies were only interested in Windows and MacOS. But today, it’s just laziness. Devs that only know web tech wants to use web tech everywhere, even when it does not fit the job. Hammer and nail style.
> 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 who will code these tests? To design good tests, you have to know the product quite well, and to know the product, you have to build it at least once. Because specs are full of knowns unknowns and unknowns unknowns. That’s why the best spec is the code itself.
Even when you know the requirements, you still have to design the test harness, and that depends on the products too. With the same caveat on knowledge.
Windows 98 constantly crashed and froze. It would even freeze during shutdown, and that was the standard experience. People during that period were paranoid about saving every change they made to their documents, and that was for good reason.
Windows 9x didn't even have proper memory protection, the most basic building block of any modern operating system. The fundamentals were as shaky as it could possibly have been.
If you frequent sites like Hacker News or other spaces where projects and announcements are shared, it’s going to look like this.
My friends who don’t follow these spaces have no idea, because those high volume, average quality projects aren’t newsworthy.
The highest quality projects aren’t out here advertising that they were built with AI. Smart teams know that saying something was built with AI is not an attractive quality for consumers. It’s only the teams trying to increase their status and make more content for their resumes who are blasting out announcements about how much AI they’re using.
From here on out, you can assume that any product you use will have had AI used in the development, likely heavily. You’re going to start running into products that blow your mind with their quality, UX, and polish where teams will have quietly employed AI to a high degree. Human input still remains a significant a valuable part of the process, but high performing and attentive teams are actually the ones most able to extract the best results from AI. They’re also the least likely to advertise that fact.
All of the “made entirely by AI” projects you see are going to continue being fun experiments to see what AI can do or little exercises in people trying to advertise their AI skills so employers will hire them or give them raises. When the really great products come along, they’re not going to come with “made by AI” advertisements at all.
> I hope we have products thanks to AI that compete on reliability of software and hardware
That will only work if most people select for quality and it's very well documented that people don't
If they sell it to a big corp it will become slop in a matter of weeks.
And even to replicate this - developers of this calibre is 1 in a million.
The traditional corporation consists mostly of information movers, sitting behind relatively few accountability-accepters and action-takers.
The traditional corporation will be replaced by a smaller version consisting purely of high-agency accountability-accepters/action-takers.
These may grow, to be "large companies" - but with a fundamentally different makeup.
Not even close IMO, a massive number of humans are part of this process... their collective input is what the model is!
Sorry, but the complexity of these devices is much higher than anything that came before it.
What are you comparing this against, and why do you think it would be reasonable to expect higher quality at a similar cost?
Something tells me that a generation of TikTok influencers aren’t going to turn around and demand quality from their computing devices.
The best you can hope for is that the Swiss cheese of OS security is patched up a little more. At least those problems make headlines.
My solution is to have an independent ratings agency that gives OS' and individual apps a score. Then software makers could compete by advertising their scores.
No human in the product making. No human in the product use.
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.
"main KPI, productivity will be used to deliver the same amount of work (or less!), at lower cost."
Maybe you personally don't need to be worried but I know many who are. Lower costs really mean less headcount right now.
That explains so much. Everyone (or at least, everyone we hear about) is using AI for faster, which leads to more, but not better.
I agree. Considering these AI companies promise us everything at perfection, it is shocking how the quality isn't anywhere near perfect.
Which part of the entire history of our industry (and other industries) leads you to have a hope like this?
Software quality (or the lack of it) It's all about the process that produces it. Most companies don't run a quality oriented process. Accelerating the pace of change is only going to make things worse faster.
How does that make any sense? How is really good UX, peak performance and optimization driven by the absence of humans in the process?
This is complete nonsense.
Something simple becomes complex by increasing the number of relationships. Binary, DNA, Amino Acids, forces, our universe.
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.
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.
On the other hand, the performance per area of the "traditional" CPU that is used for comparison, which I assume that it is some kind of RISC-V, is very poor, so that the performance per area of the new CPU is likely to be also poor, even if much better than that of the strawman.
Their "traditional" CPU has an area that is slightly bigger than that of Cortex-A720 made with a CMOS process almost twice as dense.
So if they were made with the same process, the "traditional" CPU and the new CPU would have an area somewhat bigger than a half of a Cortex-A720.
But Cortex-A720 is an out-of-order 4-wide superscalar core with advanced SIMD and scalable vector extensions, with a good branch predictor and with many other enhancements, for which the area ratio in the same process is expected to be much greater than 2 in comparison with the CPUs evaluated in TFA, while its performance must be several times greater than that of a 2-wide in-order CPU core.
So both the "traditional" CPU and the Frostbite CPU appear to have a much bigger area than expected for their performance.
Moreover, the performance of in-order CPUs is very sensitive to the benchmark selection and to the quality of the compiler optimizations.
So the performance obtained for the benchmarks selected by them, like FFT or LU decomposition, is expected to be much better than for benchmarks that would execute an application not dominated by such regular loops, but having lots of hard to predict branches, e.g. a compiler benchmark.
They say that the ISA will be published. At that time it will be easier to appreciate whether this CPU architecture is really innovative. For now, they have done a lot of work, by fully implementing the CPU core, but the data presented is not sufficiently convincing to demonstrate that it is indeed an architecture that will be able to compete with SOTA CPU cores, e.g. with those from the Arm company.
Generally this has not been my experience. Typically both very slow CPUs like the ATMega328 and high-performance CPUs like an Intel i7 are in the neighborhood of 1 nanojoule per instruction. (For operations more than 8 bits wide, the AVR is significantly worse.) Low-power CPUs like the STM32L line are commonly around 0.2nJ/insn. Ambiq's subthreshold logic gets them closer to 0.05nJ/insn.
I agree with the rest of your comment.
The fairest comparison is between chips from major vendors in the same year. If you're just doing FP32 matrix multiplies, a 2026 nVidia GPU is ~1 pJ / FLOP, an Apple M4 is ~30 pJ, and a 2026 Intel Xeon is ~50 pJ. The quality of process and optimization is a little different between those 3 companies, but not by as much as 2x. But from those numbers, you can reliably conclude that the GPU architecture is much better for matrix multiplies.
By this I mean: currently, transforming a program written in a typical high-level language to optimized ISA code involves what can only be described as black magic inside the compiler. If you don’t know the implementation details of modern superscalar OOO branch-predicting microarchitectures then you may have no clue why two very similar-looking functions could have dramatically different running times. Cue Matt Godbolt tech talk…
What I see as the more promising side of this type of development is a tighter coherence, and thus more transparency in expected runtime performance, between the software above the ISA and the hardware below it. Essentially we now have a linear machine with a linear source language, and with the popularity of Rust nowadays we actually have a critical mass of developers who are already used to applying ownership discipline in program design.
I shouldn't have to read this far into the tea leaves to figure out what performance claims a company is making, but these facts combined sound like they're claiming much better absolute performance, too?
> Each Frostbite quad (a two-by-two grid of tiles) paired with four L3 nodes is 0.792 mm².
> This TCPU [the one they benchmarked against] is 0.895 mm² at the same process.
> Across all the workloads, we average an 8.77x performance-per-area
It got 8.77x performance per area, and was 88% of the area, so that's 7.76x actual performance, right?
Not that "a parallel processor outperformed a single-core processor on parallelizable tasks" is the most earth-shattering relevation in the world, especially when they won't say which single-core processor "for legal reasons".
> On a set of 25 initially chosen workloads, including 15 held out until the final evaluation, we achieved a 4.59x performance-per-watt multiple when compared to a modern, silicon-proven superscalar (dual-issue, in-order) CPU core.
They do start claiming performance per area in the following paragraph, but that's not an interesting claim for a couple of different reasons:
- As I said above, MISC designs get dramatically better performance per area than any mainstream CPU.
- We're already in the dark-silicon era, where even the relatively low performance-per-area designs used in laptops and cellphones must thermally throttle on a regular basis to avoid burning themselves up, and contain many specialized accelerators used only a small fraction of the time (AES-NI, AVX-512, saturating 8-bit arithmetic, popcount, etc.) instead of additional small general-purpose cores that could run all the time.
Their claims would be a lot more interesting if they provided some kind of absolute numbers to compare against.
- 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...
They also compare their architecture simulation language to RTL, instead of comparing it to existing arcitecture sim languages.
Generated by ChatGPT, probably.
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.
Put that "seed" in a soup of chemicals, and within few days a chip builds up.
It’s similar to how proprietary source code is rapidly losing its value. The most valuable code now is open source, and whoever is best at building processes around open source becomes the winners.
One caveat is that new AI architectures could change the equation, at least temporarily.
We’re converging on intellectual property losing its value in general. What ends up being valuable can be boiled down to culture, human processes and physical materials.
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.
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_lawRight 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.
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.
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.
I think China is working towards making the models commodities.
The real breakers will be access to cheap raw materials and to cheap energy.
(RIP Arvind)
The technology may be different but the conclusion is similar.
https://nerdralph.blogspot.com/2017/02/inside-amd-gcn-code-e...
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.
He's been quite critical because hw architects don't talk a lot to software developers.
The architecture doesn't even have a program counter or call stack, any application running on it would basically have to be programmed in Vine.
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?
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.)
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.
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.
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?
Mostly because of SoC designs and dedicated/specialized compute units and versatile instructionsets yadda yadda
Very true.
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.
I didn't make this claim that he wants me to prove:
> "Does it, or is that because of the large caches and memory bandwidth?"
He made that claim, not me.
The fact that single core performance has been improving consistently in a compounding fashion in the Apple series of M and A chips is a historical fact.
M5 S-core is 3.4mm2 including private L2.
But Intel's Lion cove is on N3B while M5 is on N3E which is about 10% less dense. do the math to scale up Lion Cove to N3E and it becomes around 5mm2.
That's 25-35% bigger for worse performance on the same node.
AMD could have gone with 3nm, but went with 5nm (N4P is 5nm++) instead. 3.72mm2 for the core and 4.77mm2 with private L2 cache.
N3E has the same size SRAM as N5/N4. Theoretical logic scaling is around 1.5x (216 MTr/mm² and 143.7 MTr/mm), but high-performance layouts don't get anywhere near that number due to the extra spacing needed for wires and transistors.
If we assume there's zero SRAM anywhere in the core (not true with large register files, L1, uop cache, etc), we get about 2.51mm2 for the core plus another 1.05mm2 (no cache shrink) for L2 giving 3.56mm2 (bigger than M5 S-core). With a more realistic shrink, we could be talking a total area of 3.9-4.2mm2.
TL;DR -- even in the absolute best-case scenario, both Zen5 and Lion Cove are bigger than Apple's M5 S-core.
Honest question again <waves to whoever was ticked off by that>.
I could ask a LLM but HN could use some actual discussion.
Apple doesn’t need uop cache while x86 designs do which means more power and area.
ARM has looser memory ordering which improves performance. x86 can do this too, but they have to add an entire memory speculator and all the circuitry to handle when the predictions are wrong. Again, that’s extra area and power.
More controversially, x86 decode is still a significant part of the core area. ARM A715 cut decoder area by 75% just by dropping ARMv7 support (and it’s nowhere near as complex as x86). If the decoder is just 5% of the core area (and basically always on), cutting that by 75% means reducing core size and power by a significant amount. I’d also note that a company doesn’t choose complex 2x4 or 3x3 decoder implementations with massive complexity and inferior worst-case performance unless something is forcing them.
The situation isn’t quite as bad is it might otherwise seem though because 512-bit SIMD as it adds to core size (though arguably doesn’t add very much to most workloads over 128-bit aside from some great new instructions). We can calculate this with Strix Point which uses 256-bit “double pumped” SIMD. The core plus L2 area drops to about 4.2mm2 (or about 3.1mm2 for just the core). Using perfect 1.5x scaling for the core and no scaling for the L2 gives around 3.2mm2. Assuming more reasonable 1.3x scaling (which might still be too generous) gives 3.5mm2 which is still bigger for worse performance on hypothetical N3E vs M5 S-cores.
Even in the best scenario, x86 is way behind in total performance, perf/watt, and area. If we assume the engineers at AMD and Intel are good at their job (a good assumption given how long they’ve been beating everyone else on the planet), it only leaves the ISA holding them back complicating things with legacy bloat.
You’ll see a lot of pushback to this idea, but never much of substance to the assertion.
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)
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.
A calculator is superhuman. In speed of calculating numbers. Then again, Wikipedia lists calculators as AI, for just that reason.
I don't personally believe that intelligence is one dimensional, to the point of doubting that "general intelligence" exists or is even a valid concept. It's like saying "general locomotion." There is no such thing. There are wings for flying through air, fins and rotors for water, legs and wheels for land, rockets for space. A wheel doesn't work well in the air or space and a rocket is a poor locomotion method in the water.
This implies, BTW, that I don't think humans have general intelligence, because I doubt the existence of such a thing categorically. We get all up in our egos thinking we do because we feel so smart at all the things we're smart about, and of course we can't see the other things because we can't see them. Might not even be able to imagine them.
A chess algorithm is superhuman at playing chess. A human beats AI at writing prose or telling a joke. AI can beat a human at bug-hunting in a giant code base or constructing a large mathematical proof. Humans can beat an AI at dreaming up entirely new directions for mathematics. AI can beat humans at taking those new directions and fleshing them out and relating them to current art.
And so on.
We in fact tend to build machines to be hyper-good at things we are not good at, because that makes them useful to us. So we make AI that is better than us at fast thinking, huge context windows, huge scale associative memory, etc., just like we made classical algorithmic computers that were hyper-good at doing math.
There has been an open question in psychology for almost 100 years now: can you design a solvable abstract task for human performance that isn't correlated with this principal component? No one seems to be able to do it, and the AI models have only reinforced this. Making AI models smarter makes them better at almost EVERYTHING. The scaling properties of LLM models, and the scaling properties of neurons in the brain, ensure a general improvement of pattern matching that seems to be task invariant.
Personally, I am convinced by the speed hypothesis here. Since IQ in humans is very correlated with thinking speed, the ability to generate and evaluate more hypotheses per unit of time yields general improvement.
Also, this is an early stage technology that might be promising and it should probably be compared against x86, ARM, and RISC-V in their early stages. I feel like we are comparing a young up-start against established legends.
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.
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.
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.
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.
When they actually make a CPU and offer it for use, then we'll talk about their magic LLM-designed CPU.
It's not like Intel dug x86 from some mine in Oregon
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?
It's not especially hard to get chips that win in some metrics on some inputs, especially when you get to design the ISA.
I don't see any indication that they got their chips fabricated, which is the "last 90%" of the work.
... WHY ARE THE PRICES SO DAMN HIGH COMPARED TO THREE YEARS AGO?
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.
Cache hierarchies also exist because memory latency and capacity tend to be inversely proportional to each other, and directly proportional to cost. The underlying process and physics of chip design works against you on this front.
I'd love to have 1 GB of L1 cache, but that is simply not viable because your cache would be as big as your core, way more expensive, and by the sheer size, it would become a NUMA cache, which defeats the point of a cache and the "illusion" of UMA.
I think the greatest benefit of rethinking CPU architecture at the level of the ISA will be to make performance more transparent to developers without the layers of black magic.
Citation needed. After 10 minutes of searching I can't find anything to support this statement.
Because parallelism is not that needed, there’s always contention starting from physical constraint. A user can only have one body, a printer can only output one piece of paper at a time,… There are natural areas for parallelism and it’s been already implemented. Memory as a bottleneck has not been a things for decades.
Breaking down things in terms of tasks is not hard. The hard part is discovering that those parts need to communicate. And sometimes they want access to same limited resources. We already have things like array programming that have a parallel view of computation. But no one is rushing to get things done with them.
In a way I see the various cloud computing services as the organically grown (i.e. haphazard and clunky) answer to this problem.
But even that feels lacking, even though I'm really impressed with Seastar. I love the "bring operations to data" thinking.
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
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)
As an evidence, ask anyone of the programmers using LLMs to code. Who gives up first and needs a break to come back next day?
This money is generally not spent on things that AI can easily make cheaper.
Could be verrry interesting .
Current "home lab" 3D printing is fundamentally different from chip manufacturing (additive manufacturing versus lithography), so it's unsurprising that it's incapable of producing chips.
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.
If we use the very abstract term “the future” that is.
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.