Forgot your password?
Close
typodupeerror
Businesses AMD Hardware

AMD Buys AI Chip Startup Taalas That Hardwires AI Models Into Its Silicon (cnbc.com) 26

Posted by BeauHD from the mergers-and-acquisitions dept.
An anonymous reader quotes a report from CNBC: On Thursday, AMD said it's entered into an agreement to acquire Taalas, a Toronto-based startup that makes chips for inference. Taalas' accelerators are customized, or hard-wired for a single AI model, rather than being general purpose. In exchange for that loss of flexibility, Taalas' technology promises a less-expensive chip that it says can produce output for specific models thousands of times faster than a traditional GPU. An AMD representative declined to provide a purchase price for the transaction. Taalas has raised a total of $219 million in venture funding since its 2023 founding.

Taalas' current chip runs a small version of Meta's Llama 3.1 model, though the company is working on chips for bigger and more advanced models. It's manufactured using an older Taiwan Semiconductor Manufacturing Co. process, and uses speedy SRAM memory on the chip itself. Taalas CEO Ljubisa Bajic says on the startup's website that the company "developed a platform for transforming any AI model into custom silicon." "From the moment a previously unseen model is received, it can be realized in hardware in only two months," Bajic wrote.

AMD Buys AI Chip Startup Taalas That Hardwires AI Models Into Its Silicon More | Reply

AMD Buys AI Chip Startup Taalas That Hardwires AI Models Into Its Silicon

Comments Filter:

  • similar to Lisp and Java machines or transputers running Occam

    • Re: (Score:3, Informative)

      by Anonymous Coward

      Symbolics machines are mind bending

    • Depending on cost, I suspect. These would be ideal for home assistant style hardware, if that's where the price point falls. Instead of Alexa, buy one that natively runs your favorite AI. But if we're talking about something in the price range of a high-end Nvidia card, well, a 4-figure price would put it out of the home assistant market.
      • Now that I've actually looked, right now they're going with a PCIe card and it states "2.5 kW Server". That's 2.5x what my PSU provides, so I figure they aren't targeting the personal desktop or home assistant markets.

        • They're claiming 10x Cerebras (a 25kw system) performance.

          100x performance per watt.

          Even if they're overestimating / best casing the situation there is value to be had here.

          The current frontier models have use. This could be made for one and depending on hardware cost be useful for a year and replaced (Nvidia/AMD hardware are advancing fast enough to be obsolete pretty quick too). I can see a business built around a product based on an open weight model using these in servers.

          This is for end user/cloud prov

    • Re:This will be very niche (Score:5, Insightful)

      by HiThere ( 15173 ) <charleshixsn@NOspaM.earthlink.net> on Friday August 07, 2026 @12:41PM (#66277652)

      Yeah, quite niche. But it might be a VERY large niche. I see it as handling, among other things, reflexes for robots. Or recognizing voices. We've got several "special purpose" AIs native in our brains. One of them processes sounds. Some are simpler than others. Some are more reprogramable. like the reflexes used in playing a piano or driving.

  • Smaller, on device models are the future (Score:5, Insightful)

    by Registered Coward v2 ( 447531 ) on Friday August 07, 2026 @12:00PM (#66277562)

    Most tasks don't really need the big cloud based models. While on device models will lack the depth, breadth, and speed of cloud offerings, they have a number of advantages:

    • Private data can stay private and not get into the cloud models knowledge.
    • For specific types of tasks with well defined needs, the smaller models will be good enough
    • Costs are a lot less and more controllable since you aren't buying tokens at an unknown rate
    • Cloud models can still be accessed if needed via an API while still not revealing private data; for example you could use the cloud to do a lit search and use the data locally
    • They can be hosted on a private server or on the end user's machine
    • Speed isn't necessarily a big issue for many tasks vs costs; and speed can be bought

    Of course, if the silicon gets powerful enough to make local models more commonplace, the AI firms may be more reticent to release stripped down versions for free when they discover it's no longer the first hit is free model...

    • Honestly, if I had the full power of Claude 4.6 or 5, or Opus 5, on a chip, I'd be happy to buy it. I have $1500/month AI credits in work on CoPilot. Even if the chip was twice that, and I could plug it in to via USB to my Mac, and it gave me what I got with the cloud connection, it would be a huge massive saving.

    • Tried this hardware's demo to write code, and I can tell you it changes the paradigm of how you use it. Virtually, insanely instantaneous, way fast enough to be truly conversational.

      Frontier models are nice and all, but they're way too slow. Interacting with them is like running underwater. Developing software with the current models is painful and deliberate. It breaks your flow, slows down iteration. Maybe that's part of why devs are feeling somewhat demoralized when they switch to AI. That's super-import

  • Damn their chatbot is fast! (Score:4, Informative)

    by mileshigh ( 963980 ) on Friday August 07, 2026 @12:10PM (#66277574)
    Seems to be for real.Tried it on substantial prompts that couldn't possibly be pre-cached, 100ms response. Translated 1500 words to French in 50ms, and it's a good translation. Their demo's just an 8B model for now, but it's not hard to see how this can scale up. If nothing else, it can have a real-time conversation with no delay. That alone guarantees adoption in all sorts of applications where it interfaces with people. Methinks AMD got itself a bargain.

  • So no software updates then (Score:4, Insightful)

    by LindleyF ( 9395567 ) on Friday August 07, 2026 @12:47PM (#66277666)
    You're stuck with the model the chip was built for. The better one released a week later is out of reach.

    • Re:So no software updates then (Score:4, Interesting)

      by CommunityMember ( 6662188 ) on Friday August 07, 2026 @01:17PM (#66277708)

      You're stuck with the model the chip was built for. The better one released a week later is out of reach.

      Not all applications need the latest model. So while these chips would not be a good solution for running the frontier models which evolve rapidly, there is a lot of more mundane and stable uses where the dedicated chip might be a viable choice.

    • Would it be feasible to burn into the chip only a few layers of a transformer, for instance the first layers, and then train the rest? In this way you pair some speed-up and some trainability.

      I'm not an expert, but as far as I know there are contexts where basically only training the outer layer gives you a decent-enough neural network.

    • 2 words: "Bic pen" and all that suggests. 3-year h/w upgrade cycle isn't law. If a like-priced chip can do the job of 100 GPUs, you can afford to change it every month if you like and bank the power savings.

      Besides, this things's so freaking fast that sticking a decent-sized RAG bundle into the context won't kill performance. There's your update mechanism.

    • It comes in the form of a chip? then it could be housed in a cartridge-type form and then exchanged just as easy as those were with the Nintendo consoles of yesteryear, the C-64 etc. At least with those 2 types of computers I have personal experience running cartridges and it was already fine back then.

    • You're stuck with the model the chip was built for. The better one released a week later is out of reach.

      And? If a model works why do you need "the latest" one? We've been using the same unchanged AI model for some tasks at work for well over a decade. It works perfectly for its application (predicting machine failure based on process inputs) and there's no reason to change it in the slightest.

      If you have a model for translating language why do you need a new one? Is your lexicon being updated so quickly? Do you need your model to absolutely know what skibidi or rizz means because the kids these days think the

      • My Nest Cam regularly announces my car is driving away when it's just sitting there. Or describes my cat as a raccoon. And that's a cloud model. (The on-device model can distinguish animal from person but no more.)

        Any model is going to have issues and you can't always anticipate them.

  • So, in addition to badly trailing Nvidia in the AI GPU market, AMD is also aiming to badly trail Google, et al. in the ASIC space. The biggest customers for AI ASICs are the companies that design the ASICs themselves. That makes sense because those companies know exactly what they want the chips to do. It makes no sense for them to buy those chips from a third party because they already shift much of the supposedly non-core chip design to Broadcom, Marvell, or another company. ASICs only make sense in e

Slashdot Top Deals

System going down at 1:45 this afternoon for disk crashing.

Close