AI Chip startups vs big companies
Dv engineer with 1 yoe looking to learn more about the ai chip scene looking for advice from DV OGs on blind. Do any of the new chip startups have a chance of becoming the next NVIDIA or do developers just love NVIDIA too much ? If some company even makes a chip that beats them will the hesitation to move infrastructure and the ability of Nvidia to eventually catch up stop developers ? Any notes on tenstorrent and mythic ?
Hardware is a tough industry and semi is even harder. Most likely the companies that might be on the fringe of individual success will get acquired by the big semiconductors.
One thing that most ai startups are slowly starting to realize is: sw is as hard a problem as hw for machine learning acceleration. So startups that create architecture well suited for a workload really needs to deliver those TOPS to actual usable perf which needs a solid sw stack. That’s one area nv is doing a great job!
That is true. I actually feel that it is easier for nvidia to adopt the new hot architecture compared to the startups convoncing customers to trust them
I picture future as all those big tech will have their in-house chip team to do customized solutions, like TPU for tensorflow. AI is fed on data and only big tech have that scale of data to do really great AI application at the end. When you need a mega scale data center, it becomes reasonable to build it from scratch like what google and facebook is doing. Then you will provide packed research resources to academia to close the loop and make students become familiar with the toolz
Many of them have a chance. Collectively, I believe that at least one of them will eventually displace Nvidia as the defacto solution for running AI. The question is which one? All of them have a long way to go and it is unclear which, if any, of the existing crop will succeed. Most of them have a better architecture, but whether they can win developers and deliver the product at scale is the question. It is just too early to tell.
Appreciate the insight. What are your thoughts on the fact that big companies like aws, google, fb are all making their own chips and providing cloud services ? If ml is a game of how much data one has is it better to bet on their chips ?
The CSPs are developing chips for their own use (mostly - google is sort of trying to sell TPUs) not to sell to others. So while these may be come prevalent in their instances I don't any of them taking over the market broadly.