140
you are viewing a single comment's thread
view the rest of the comments
[-] tal@lemmy.today 5 points 5 hours ago

One thing I've been wondering about, after seeing an article a while back about Nvidia talking to TMSC about manufacturing memory is whether they might do an ASIC aimed at being a neural net or part of one.

The limiting factor today is memory bandwidth. But that's because you have to ship all the data over a long, narrow memory bus with big fat traces.

If you can stick a bunch of adders and multipliers on a single chip and put memory on the same chip, then you don't have to slow things down and generate the heat that shipping everything off-chip requires.

If so, that might be one form that AI-optimized memory might take.

[-] Buelldozer@lemmy.today 4 points 4 hours ago* (last edited 4 hours ago)

Nvidia bought GRoQ for exactly this. 13,000 tokens a second from an expansion card that uses 200 watts. AMD bought Taalas for the same reason.

[-] brucethemoose@lemmy.world 3 points 4 hours ago* (last edited 4 hours ago)

Nah, Groq is toast. Those numbers are for toy models distributed across tons of cards, and splitting MoE LLMs across so many cards in SRAM an inefficient approach IMO.

The best model they serve is the kinda obsolete GPT-OSS 120B, at 350 token/s: https://openrouter.ai/openai/gpt-oss-120b

Which you can get with Deepseek, Muse or other open weights models these days, if you really need it.

The only ASIC maker I kind like is Cerebras, especially their scheme for hybrid inference in conjunction with AMD. Architecturally, that makes a lot of sense.

EDIT:

Upon further investigation, they host more models than I thought:

https://console.groq.com/docs/models

They have MiniMax 2.7 at 200 tokens/s, which isn't too bad... but still largely obsolete now.

[-] brucethemoose@lemmy.world 3 points 5 hours ago* (last edited 4 hours ago)

Yep. Samsung is already researching this. They have been for some time.

Also, I feel Bitnet would be useful here, as it would only require small adders, no huge multipliers.

They could even take a hybrid approach, processing higher-precision attention and small parts of the LLM on the GPU 'traditionally' while specialized adders in memory handle the huge ternary MLPs (or whatever the sparse equivalent is in other ML architectures). This would also give them architectural flexibility (as attention schemes are evolving wildly, the MLPs kinda aren't).

this post was submitted on 20 Nov 2026
140 points (96.7% liked)

Technology

89921 readers
3908 users here now

This is a most excellent place for technology news and articles.


Our Rules


  1. Follow the lemmy.world rules.
  2. Only tech related news or articles.
  3. Be excellent to each other!
  4. Mod approved content bots can post up to 10 articles per day.
  5. Threads asking for personal tech support may be deleted.
  6. Politics threads may be removed.
  7. No memes allowed as posts, OK to post as comments.
  8. Only approved bots from the list below, this includes using AI responses and summaries. To ask if your bot can be added please contact a mod.
  9. Check for duplicates before posting, duplicates may be removed
  10. Accounts 7 days and younger will have their posts automatically removed.

Approved Bots


founded 3 years ago
MODERATORS