$64B Gamble: SoftBank Arm Plan to Launch AI Chip in 2025
geekynews.org
geekynews.org
And… that page is plagiarized from https://www.ihg.com/crowneplaza/hotels/us/en/dublin/dblbl/ho...
This looks to be a mix of stolen/plagiarized copy and LLM-regurgitated crap.
If you want a suggestion, I guess this is one: don’t steal other people’s words and republish them as your own.
My limited visibility to them usually doesn’t involve a lot of good news.
And, CPUs prioritize hiding latency with all sorts of caches, and GPUs prioritize cores and bandwidth to hide latency, so there’s different tradeoffs about memory bandwidth versus latency.
More relevant to the article - I'm highly skeptical that our current architecture - LLMs in particular - will scale to AGI, let alone super intelligence. Humans in particular literally use orders of magnitude less energy and the smartest among us are still more capable. IMHO AGI, let alone superintelligence will require quantum computing getting somewhere, some sort of organic/biological computing or something else not invented yet.
Arm does seem like a natural fit, though. It seems everyone want to make their own chips these days.
this is such a dumb question but can arm chips even run AI models? it seems like AI on the edge is still 10-15 years away with everyone trying to build a moat
I can't imagine more than 2 years for GPT4 level LLM on edge devices.
The question is, will ppl want GPT4 on the edge when GPT 6 is one request away?
"According to Precedence Research of Canada, the AI chip market is poised for explosive growth, with projections indicating a rise from $30 billion this year to over $200 billion by 2032.
Another research firm DataHorizzon Research even projected the AI Chip market to be 1,114.3 billion by 2032."
Ok, few here will deny the AI (chip) market is exploding. And CEOs will have visions of datacenters packed with expensive AI chips, extracting value from consumers. But if I were to take a guess:
AI models will be optimized the heck out of. Their architecture is only at the beginning of a long road of innovation (and in particular: simplification - aka race to the bottom). With so many heavyweights getting into the game, competition will be fierce.
Bottom line: yes, AI chips will be everywhere. But they'll also be cheap(-ish). Not unlike current day cpu's, flash, RAM etc.
With that in place, the next revolution: local-running AI models in everybody's hands. Cheap, easy to use & tinker with like mobile/PC apps today.
But that's NOT what wework was.
isn't it like AirBnB but for offices? isn't that good? like obviously there are complaints about AirBnB and stuff but it let your average person list their spaces
No.
[1] https://www.arm.com/products/silicon-ip-cpu/ethos/ethos-u85
Seems like there are 2 plays here, there is inference training and then there is the end user inference acceleration. I expect they'll avoid the training side in gen1 but make some standard instruction additions that will become available on all ARM parts in the next couple years.
Right now ARM is worth >50% more than SoftBank itself. It seems odd that they are still holding 90%, and also committing so much cash for future investments.
Again, if ARM needs money, ARM needs to sell something ARM owns or to take out a loan. SoftBank's selling something SoftBank owns does not by itself increase ARM's cash even a little. (A company can create new shares in itself at any time, so one of the things ARM can sell is newly-created shares in ARM. As soon as ARM creates new ARM shares, SoftBank's ownership share in ARM goes down from 90% to less than 90% unless SoftBank buys some of those new shares.)
e.g. tsn200 40W 500TOPS tsn800 300W 2000TOPS
Similar products [1,2] are approx $1-$2/TOPS (this is theoretical 8bit int ops, so multiply by x2-x4)
[1] https://www.axelera.ai/ [2] https://hailo.ai/#discriminative-ai