8 ms·

Once I got my hands on stable diffusion, I bought a 3090 because I was afraid that they would go up in price, but my power supply wasn't big enough and I return

Once I got my hands on stable diffusion, I bought a 3090 because I was afraid that they would go up in price, but my power supply wasn't big enough and I returned it for a 2060. But now I see that 3090s are dropping like 40% in price since then. This article does a great job explaining these dynamics. Metaverse and crypto will be a bust, but democratized AI is going to explode, the big risk for NVDA is that models and tools get small and efficient enough that they don't need $1,000+ hardware.

flyinglizard 4y ago
$1,000 is a very reasonable enterprise price point. After all, datatcenter CPUs are much more expensive.
- bongobingo1 4y ago
  I think democratizing AI means more than "a very reasonable enterprise price point" of $1000.
  - smoldesu 4y ago
    It doesn't, you can run this on $150 GTX 1060s if you're willing to wait a few moments. The interesting angle here is that there is a consumer segment of hardware (gaming GPUs) that can be used to drive cool artistic experiments. I think for most people, even non-gamers, getting dedicated acceleration hardware is not any more expensive than an entry-level analog synthesizer.
    - aidenn0 4y ago
      With only 4GB, I'm RAM limited for a lot of models. I'm thinking of buying a dedicated GPU just for ML stuff.
- Dennip 4y ago
  Won't enterprise always go for the quadro cards which are more like $4000+
- vineyardmike 4y ago
  Democratic means fit on an iPhone if not on $50 throw-away hardware.
jeffparsons 4y ago
> the big risk for NVDA is that models and tools get small and efficient enough that they don't need $1,000+ hardware. You will always be able to do more and better with the premium hardware. Maybe for video games the marginal differences aren't enough to sway most consumers towards the high end cards, but I expect there will be a lot of people willing to pay top dollar to be able to run the best version of X AI tool.
- solardev 4y ago
  With games, the best card might make sense cuz it could be the difference between 60+ fps and 40. For workloads though, does it really matter if it completes in 20 seconds instead of 15? Unless there's some feature difference in the higher end cards (like AI-specific subchips), not just more of the same & faster, then a lower card of the same generation shouldn't impact your ability to run something or not. Maybe RAM size...
  - squeaky-clean 4y ago
    How many times do you have to complete that 20/15 second task in a year? Multiply that by 5 seconds to get the "wasted time". Factor in the employees salary by that wasted time, how much are you paying them to wait on that slower hardware? Is it more money than the price difference between the better GPU and the cheaper GPU? Or maybe it's an automated thing that runs constantly? The difference between 20 and 15 seconds is 480 additional iterations per 8-hours. How much is that worth to you? There's also the unmeasurable benefits. What if you're losing potential good employees applying because word on the net is your company uses last-gen hardware for everything, and your competitors use the newest.
    - solardev 4y ago
      I don't think it's that binary a thing. It's not like either you have a supercomputer or you have last decade's junk. There are so many different models and price points, each company/team/dev can find their own sweet spot and upgrade cadence. It's not necessarily the case that 20% better performance is 20% more money, and 5 seconds times like 100 times per week times a year is still just a single workday. I doubt any employee works effectively 365 days a year... and the cost difference of the highest end card vs a medium one is definitely more than I get paid in a day, at least. I mean, it's no different than CPUs and RAM and such. A middle-of-the-road laptop from the last year or two is good enough for most users... no company I've ever worked for gets the top-of-the-line model every single refresh for every single dev. That'd be an absurd waste of money, IMO. Personally, I wouldn't want to work for a company that upgrades hardware all the time for single-digit performance increases. That money is better used elsewhere, whether employee salaries/benefits or donating it to the community or just saving customers some money or paying dividends to shareholders. Even if you're running massive continuous workloads, it is probably more cost-efficient to just parallelize a bunch of midrange systems (optimize for perf/dollars and/or perf/watt) than to just chase the highest-end chips for no reason.
ronsor 4y ago
I actually don't think Metaverse will be a bust, but I think the current corporate implementation of it will be. Companies are trying to cash in too early and it's not really working out for them, particularly Zuckerberg's Meta.
- metadat 4y ago
  What other implementation will come to fruition? AFAICT the Metaverse is a shoddy VR clone of Second Life. The Metaverse is already a bust, essentially a stillborn fetus. The market doesn't actually want it.
  - not2b 4y ago
    The market doesn't want Zuckerberg's version. Maybe some day someone will get it right in a way that's good enough to be interesting.
    - metadat 4y ago
      Yes, maybe someday, somewhere else.
  - quantumwannabe 4y ago
    A spiritual successor to Second Life with modern graphics and VR support would likely be very successful. Dumbed-down versions like Fortnite Creative and Roblox have been very popular. Zuckerberg won't be the one to build a successful metaverse game though, not because Meta doesn't employ skilled game developers, but because culturally the leadership is completely out of touch. Second Life thrived because of its uncontrolled player creation environment, which Meta would never allow. It let you 3D model and script anything you could think of using in-game (and external) tools and had a free market economy where the developer only took a cut when converting to/from real world currency. The whole debacle with the "virtual 'groping'" and the personal space bubble has demonstrated that even if Facebook's Metaverse did have good creative tools, all it would take is a single article written by a journalist with an axe to grind for them to cripple it. They've already removed legs from avatars in Horizon Worlds for "safety", which is ridiculous.
WFHRenaissance 4y ago
>Metaverse and crypto will be a bust I love when HN netizens just throw stuff like this out without further explanation. It's as if they're monks reciting the dogma. Sigh.
- flumpcakes 4y ago
  It is probably the default safe position. I think there would be a larger need in explaining how they won't be a "bust" compared to saying that they will. The internet needed explaining to many people to see it's real advantages before it took off.
- wpietri 4y ago
  People have viewpoints. They have enough of them that explaining them in detail every time they come up would be tedious. You and I both could write up plausible explanations of what he might mean by that, as both those topics are already discussed to death here. There's no particular reason to think he's being dogmatic, just pithy. I'll also note that it would be easy to make the same accusation about your own posts. E.g.: "The contemporary democratic nation state is a facade over a distributed and decentralized oligarchy." https://news.ycombinator.com/item?id=32837598 https://news.ycombinator.com/item?id=32837598 But presumably you see yourself as holding that position for non-dogmatic reasons, so I'd encourage you to do the same for others.
  - WFHRenaissance 4y ago
    Granted, my post is a reference to history (the iron rule of oligarchy is well-documented... polemic tone aside hehehe). OP's post is speculative.
- anthonypasq 4y ago
  Every single person on this forum would have predicted every social media company to be dead on arrival.
mw888 4y ago
Democratized AI will not explode, it will thrive as centralized services until the day real time AI is expected by consumers.

flyinglizard 4y ago

$1,000 is a very reasonable enterprise price point. After all, datatcenter CPUs are much more expensive.

bongobingo1 4y ago
I think democratizing AI means more than "a very reasonable enterprise price point" of $1000.
- smoldesu 4y ago
  It doesn't, you can run this on $150 GTX 1060s if you're willing to wait a few moments. The interesting angle here is that there is a consumer segment of hardware (gaming GPUs) that can be used to drive cool artistic experiments. I think for most people, even non-gamers, getting dedicated acceleration hardware is not any more expensive than an entry-level analog synthesizer.
  - aidenn0 4y ago
    With only 4GB, I'm RAM limited for a lot of models. I'm thinking of buying a dedicated GPU just for ML stuff.
Dennip 4y ago
Won't enterprise always go for the quadro cards which are more like $4000+
vineyardmike 4y ago
Democratic means fit on an iPhone if not on $50 throw-away hardware.

jeffparsons 4y ago

> the big risk for NVDA is that models and tools get small and efficient enough that they don't need $1,000+ hardware. You will always be able to do more and better with the premium hardware. Maybe for video games the marginal differences aren't enough to sway most consumers towards the high end cards, but I expect there will be a lot of people willing to pay top dollar to be able to run the best version of X AI tool.

solardev 4y ago
With games, the best card might make sense cuz it could be the difference between 60+ fps and 40. For workloads though, does it really matter if it completes in 20 seconds instead of 15? Unless there's some feature difference in the higher end cards (like AI-specific subchips), not just more of the same & faster, then a lower card of the same generation shouldn't impact your ability to run something or not. Maybe RAM size...
- squeaky-clean 4y ago
  How many times do you have to complete that 20/15 second task in a year? Multiply that by 5 seconds to get the "wasted time". Factor in the employees salary by that wasted time, how much are you paying them to wait on that slower hardware? Is it more money than the price difference between the better GPU and the cheaper GPU? Or maybe it's an automated thing that runs constantly? The difference between 20 and 15 seconds is 480 additional iterations per 8-hours. How much is that worth to you? There's also the unmeasurable benefits. What if you're losing potential good employees applying because word on the net is your company uses last-gen hardware for everything, and your competitors use the newest.
  - solardev 4y ago
    I don't think it's that binary a thing. It's not like either you have a supercomputer or you have last decade's junk. There are so many different models and price points, each company/team/dev can find their own sweet spot and upgrade cadence. It's not necessarily the case that 20% better performance is 20% more money, and 5 seconds times like 100 times per week times a year is still just a single workday. I doubt any employee works effectively 365 days a year... and the cost difference of the highest end card vs a medium one is definitely more than I get paid in a day, at least. I mean, it's no different than CPUs and RAM and such. A middle-of-the-road laptop from the last year or two is good enough for most users... no company I've ever worked for gets the top-of-the-line model every single refresh for every single dev. That'd be an absurd waste of money, IMO. Personally, I wouldn't want to work for a company that upgrades hardware all the time for single-digit performance increases. That money is better used elsewhere, whether employee salaries/benefits or donating it to the community or just saving customers some money or paying dividends to shareholders. Even if you're running massive continuous workloads, it is probably more cost-efficient to just parallelize a bunch of midrange systems (optimize for perf/dollars and/or perf/watt) than to just chase the highest-end chips for no reason.

ronsor 4y ago

I actually don't think Metaverse will be a bust, but I think the current corporate implementation of it will be. Companies are trying to cash in too early and it's not really working out for them, particularly Zuckerberg's Meta.

metadat 4y ago
What other implementation will come to fruition? AFAICT the Metaverse is a shoddy VR clone of Second Life. The Metaverse is already a bust, essentially a stillborn fetus. The market doesn't actually want it.
- not2b 4y ago
  The market doesn't want Zuckerberg's version. Maybe some day someone will get it right in a way that's good enough to be interesting.
  - metadat 4y ago
    Yes, maybe someday, somewhere else.
- quantumwannabe 4y ago
  A spiritual successor to Second Life with modern graphics and VR support would likely be very successful. Dumbed-down versions like Fortnite Creative and Roblox have been very popular. Zuckerberg won't be the one to build a successful metaverse game though, not because Meta doesn't employ skilled game developers, but because culturally the leadership is completely out of touch. Second Life thrived because of its uncontrolled player creation environment, which Meta would never allow. It let you 3D model and script anything you could think of using in-game (and external) tools and had a free market economy where the developer only took a cut when converting to/from real world currency. The whole debacle with the "virtual 'groping'" and the personal space bubble has demonstrated that even if Facebook's Metaverse did have good creative tools, all it would take is a single article written by a journalist with an axe to grind for them to cripple it. They've already removed legs from avatars in Horizon Worlds for "safety", which is ridiculous.

WFHRenaissance 4y ago

>Metaverse and crypto will be a bust I love when HN netizens just throw stuff like this out without further explanation. It's as if they're monks reciting the dogma. Sigh.

flumpcakes 4y ago
It is probably the default safe position. I think there would be a larger need in explaining how they won't be a "bust" compared to saying that they will. The internet needed explaining to many people to see it's real advantages before it took off.
wpietri 4y ago
People have viewpoints. They have enough of them that explaining them in detail every time they come up would be tedious. You and I both could write up plausible explanations of what he might mean by that, as both those topics are already discussed to death here. There's no particular reason to think he's being dogmatic, just pithy. I'll also note that it would be easy to make the same accusation about your own posts. E.g.: "The contemporary democratic nation state is a facade over a distributed and decentralized oligarchy." https://news.ycombinator.com/item?id=32837598 https://news.ycombinator.com/item?id=32837598 But presumably you see yourself as holding that position for non-dogmatic reasons, so I'd encourage you to do the same for others.
- WFHRenaissance 4y ago
  Granted, my post is a reference to history (the iron rule of oligarchy is well-documented... polemic tone aside hehehe). OP's post is speculative.
anthonypasq 4y ago
Every single person on this forum would have predicted every social media company to be dead on arrival.

mw888 4y ago

Democratized AI will not explode, it will thrive as centralized services until the day real time AI is expected by consumers.