The Register Home Page

back to article AMD acquires AI chip startup Taalas to boost inference performance by etching models into silicon

In AMD’s latest bid to upset Nvidia's dominance in AI hardware, the House of Zen has acquired AI chip company Taalas, which bakes model weights directly into silicon in a process that promises to boost inference performance by an order of magnitude or more. The deal, announced at market close on Thursday, appears to be framed in …

  1. sebacoustic
    Happy

    Happy days - I worr for a mask maker, that means more revenue for us, right?

    1. short

      Dunno, but I work for an IC socket manufacturer...

  2. EricM Silver badge

    Really inflexible?

    I'm currently trying to keep pace with the flurry of releases of local models on Hugging Face ...

    As the article mentions, this idea seems to lend itself best for _really_ static environments.

    Maybe if they come up with a modular design, where you can clip in different model-chips into a base card already integrated into a server? Ideally from outside of the server like via an external USB port?

    Interesting concept, nevertheless...

  3. jScrag
    Terminator

    Game changer for small scale, local AI

    Scale this way down and we could be looking at small, cheap, specialised AI nodes trained for specific use cases e.g. home automation

  4. Neo-Isshin

    i suppose taalas must be the future, it will help devices with low performance such as mobile/Edge caculator/controller pad do more things. Agent relies on ex high performance chips now and waning too much power which mobile devices cannot afford barely, but taalas will change it

    1. cdegroot

      Not sure what the future holds but GPUs were more an accident than a conscious design choice for LLMs so if the coming years don't see a flurry of completely new architectures, well, paint me surprised.

  5. TimMaher Silver badge
    Coat

    Yeah but…

    … can I plug it into my Raspberry Pi?

    1. trevorde Silver badge

      Re: Yeah but…

      ... will it run Crysis?

  6. Anonymous Coward
    Anonymous Coward

    Makes sense…

    …as an base LLM to train your more advanced LLMs.

  7. Rattus

    So ROM based.

    Is there a specific type of mask-ROM that is orders of magnitude faster than FLASH to access? I don't see a particular reason why a similar FLASH fabric would impede performance.

    I suspect that the reason for mask ROM has been selected is because the same data storage on FLASH would require a a considerably larger die.

  8. This post has been deleted by its author

  9. MONK_DUCK

    All those 100s GW of electricity investments for the data centre build outs, which tend to have a 15/20 year payback period..... suddenly feel lot more risky.

POST COMMENT House rules

Not a member of The Register? Create a new account here.

  • Enter your comment

  • Add an icon

Anonymous cowards cannot choose their icon