Nvidia reveals new A.I. chip, says costs of running LLMs will ‘drop significantly’::Currently, Nvidia dominates the market for AI chips, with over 80% market share, according to some estimates.
Nvidia has always been a tech company that also happens to make consumer graphics cards.
Yeah, I misspoke there, but for most of recent memory they’ve been doing big things besides consumer graphics cards. Nvidia launched its professional oriented graphics Quadro product line in 2000. They launched CUDA architecture in 2006 which opened up parallel processing capabilities of GPUs for use in science and research. They entered the data center and cloud computing market in the early 2010s, and in 2015 they launched the DRIVE product line.
And before that it was a bit mining company, with a side line of gaming graphics hardware.
I’m sure the cost to the consumer will remain exactly the same, or somehow increase.
I’m not worried about that. There will be open competition, because most of this stuff is open-source. Cheaper hardware will open the door for anyone like you or me to set up our own services. Anyone can set up a server with their own hardware (or rent it from Amazon or wherever) and run their own chatbot (with blackjack! and hookers!) instead of using ChatGPT.
This is already possible on consumer hardware, just not with the biggest and best networks. Right now, if I wanted to run, say, BLOOM (an open-source LLM), I’d need to spend close to $100K on hardware. Obviously, that’s out of reach for a hobbyist, so I’m limited to using smaller, less advanced networks like LLaMa or GPT-J. Cheaper hardware will help break the hold that the big players currently have over the industry.
if I wanted to run, say, BLOOM (an open-source LLM), I’d need to spend close to $100K on hardware
Doesn’t that dozens of notes with over a terabyte of RAM each? And state of the art networking?
Sounds closer to $100M than $100K.
If you want to train your own network like they did, you’d want something like that, yeah, but to run the trained network you “only” need ~360GB of memory.
For context, even if you wanted to run this in CPU, there are currently no A5 mobos (Ryzen 7000 series) that support more than 192GB of memory. You literally can’t even run it on high-end consumer hardware.
I’m liking AMD still. They’re not perfect of course but they seem to have far less fuckery going on than Intel and Nvidia, and they have open source drivers that play nice with Linux.
I always have this thought in the back of my mind too, but the issue is that while raw performance is a bit better than the counterparts, Nvidia still offers more features for the money, and I don’t always have money to throw away. Typically i’d upgrade my gpu once every 5 years or so
AI might not survive the next decade? I already use it every day at work. The productivity gains are enormous and far from saturated. I think it’s more likely that AI will survive and consumers (humans) will not survive.
I think people simultaneously overestimate the capability of current machine learning models while underestimating their long term impact. These models are going to be in everything. They are very resource hungry and will absolutely be a driver of hardware innovation for the next decade and probably longer.
How are they killing their consumer market? If they change their mind and put out a better gpu people will buy it.
You’ve answered your own question. They used to release upgraded hardware with a reasonable generational boost almost yearly. Now the gap has widened, and they’re iterating on old hardware, by giving it more juice and a larger cooler. Not to mention the astronomical prices that have outclassed previous top-end cards at the current mid-range
