4 ms·

Let’s just hope that this time around the they don’t intentionally hold back innovations in favour of slow-dripping a predictable fixed X% gain per year like th

Let’s just hope that this time around the they don’t intentionally hold back innovations in favour of slow-dripping a predictable fixed X% gain per year like they did 2014 to 2020 in the absence of competition. Amazing how quickly Intel suddenly started leaping in performance/$ between releases once AMDs 5000 series came out.

benreesman 2y ago
I agree it's a weird world where Intel is our best hope for free, fair, and competitive markets. That's a really fucking weird world. But Gelsinger looks at a cursory glance to be someone who knows how to build chips, has a pretty clear comeback story, and realizes that complacency in a monopoly is what made them uncompetitive in the end, and therefore desperate enough to put a qualified person in charge [1]. I think he might keep the pressure on even if they start winning. [1] https://en.wikipedia.org/wiki/Pat_Gelsinger https://en.wikipedia.org/wiki/Pat_Gelsinger
paulmd 2y ago
> Let’s just hope that this time around the they don’t intentionally hold back innovations in favour of slow-dripping a predictable fixed X% gain per year like they did 2014 to 2020 in the absence of competition. idk why you would cite this as a downside when this is specifically what the tech review community has been begging for with gpus. Having a gen do 40% and the next do 0–10% is grating to consumers and disruptive to the market. What consumers need is exactly this smoothing of generational progression and the removal of huge outliers in terms of generational step. Smooth the curve - which inherently will be much closer to that 15-20% progress generationally, given how slow the transitor tech is moving nowadays. As Tim Hardwareunboxed says - if you don’t have a good uplift, don’t release the product. Intel made sure they had an uplift every year. You can’t change actual progress but it can be smooth.
- transcriptase 2y ago
  Why would I or anyone else care what the tech review community has to say about anything? What kind of idiot wants technological advances held back so they don’t get disappointed and the market doesn’t get disrupted?
  - paulmd 2y ago
    > Why would I or anyone else care what the tech review community has to say about anything? like, are you asking why thought leaders matter? > What kind of idiot wants technological advances held back so they don’t get disappointed and the market doesn’t get disrupted? Lots of people, have you never posted in an nvidia thread before? tons of people would much rather RTX, DLSS, and AI never existed at all, because it means they have to upgrade their GTX 1080 Ti or RX 480 or whatever. A lot of people bought their first GPU since that launch and have never lived through a hard upgrade cycle. recently it’s turned into this insane “if you can’t make cards like 3050 for $110 or get at least double the perf over 2 generations then don’t make a new generation at all” Neo-Luddism but it’s been brewing for a long time among the gaming public. https://m.youtube.com/watch?v=QbNwO4HXTIk&feature=youtu.be https://m.youtube.com/watch?v=QbNwO4HXTIk&feature=youtu.be It's obviously dumb and regressive, literal modern luddism, but it's crazy to pretend that it's not an extremely visible and popular modern trend, either. People are mad the low end is stagnating and the high end is continuing to progress, and the natural human response is “if I don’t get to have a good time I don’t see why anyone else should either”. “Don’t launch a 5090 until there’s a good card for $110” is a very mainstream phenomenon nowadays. Literally the drum the 2 biggest and most popular tech channels have been beating for a year now.
    - benreesman 2y ago
      I don’t think a vocal minority under the sway of what it sounds like you’re describing as ill-intentioned propaganda (I mean, that’s a serious word but the shoe seems to fit) is a particularly good proxy for the overall preferences and welfare of everyone in the ecosystem. To the extent that legitimate enthusiasts and early adopters and people trying to get the most out of their gear under resource constraints have a clear consensus, yeah, you want to be listening carefully. But that doesn’t really sound like what you’re describing. I hang out more on the “accelerated compute” parts of the “GPU Internet” than the “gaming” parts, but certainly over here in LocalLLaMA land, we want working, documented firmware, and cards that aren’t overkill on some things and weak AF on others (4090). That’s an effort to manipulate nosebleed margins by making the ostensible “prosumer/enthusiast” offering just kind of bad at everything (Quattro) except maybe CAD or Blender. In gaming it seems like a cost-constrained public would be best served by credible offerings ranging from “the minimum to play modern stuff at a very attractive price” (I remember the 1060 being very well-received here), up to “Crysis at a million FPS” at Veblen, elaborate bespoke case, over clock with active cooling, and a sensible set of options along the way. I don’t do enough gaming these days to have strong views, so please educate me if you’re an avid gamer and I’m full of shit, but both major gaming vendors seem to have good options for most budgets on mainstream platforms. That’s just nothing to do with the 2T cap and 55.58% Net Profit Margins us llama.cpp folks are up in arms about.
- benreesman 2y ago
  I have a certain sympathy for all involved in your argument, but I don't think it's a good way to do business for either businesses and their executives, their shareholders, or ultimately the consumer and public. The problem is short-term thinking around annual or even quarterly earnings reports and whatever drives that. I appreciate that a CEO or even a board may feel pressure up to and including losing their jobs if they leave literally anything on the table in the form of long-term investment, the "goodwill" (which is a real financial accounting line item that's sometimes a lot) of their customers, partners, etc. and the long-term sustainability of the business. But it's that ability to deliver that magic combination of a product that people ideally love buying, but at least don't mind buying, a team that is healthy, well-compensated, and inclined to go the extra mile for the employer at rare need because the mutual priority of the wellbeing of the partnership goes both ways (this is massive, when you treat people well when they're in a bind, they'll go the extra mile when you're in a bind with little if any pressure, most of the time), and a consistent, low-beta (for the investor's risk appetite) return on investment for shareholders. This is the kind of business that e.g. Warren Buffet at least talks about as the kind of business he learned to invest in. I'm not a Buffet scholar or watcher really, so I don't know how well he lives up to those ideals in general or especially recently, but it's certainly a pretty good way to describe the desired outcome of capitalism! I've heard lots of arguments for why boards and CEOs and the like simply have to skin the sheep once rather than shear him many times (to borrow the old saw): but never a very convincing one. It's basically absurd in the kinds of dual-class share structures that are common in tech now: it's not only extant but common for the founder/CEO to be basically unencumbered by quarterly earnings reports, what's e.g. Mark Zuckerberg going to do? Fire himself? And under this more "good for everyone over a long time horizon" model, you can just launch the next breakthrough card when it's done, and launch incremental improvements when you're out of inventory, or see a gap in your lineup, or whatever. Fix corporate governance in what should be the relatively uncontroversial ways I and many others have described, and you fix this (among many, many other) set of perverse incentives.
  - paulmd 2y ago
    I don't think anyone's under any illusions that people are going to upgrade more frequently, but the lack of predictability is pissing people off and miscalibrating expectations on what they "should" be getting every gen. Going 10%,40% is literally worse and more disruptive (in the bad way) to consumer psychology than if you'd just gone 20%,25%, because people will lock in on the most desirable possible combinations of perf gains and value offerings/prices. the x70 tier launched at $349 in 2010 dollars and has bounced between $349 and $399 for every generation, but the one people remember is the singular time it was $329. And the x70 die has almost always fluttered around 250-300 mm2, except for the trailing-node era (when NVIDIA got cheap TSMC/samsung and chased price) where the 2060 was almost as big as a 1080 Ti (445mm2 vs 471mm2) etc. But people have never shut up about how the 4070 isn't a real x70 because it's only 300mm2 etc - they latch onto the one singular time it was bigger etc. what you end up with in terms of node progress is similar - people tell themselves they deserve the gen when everything increases by 50%, at the best possible efficiency, with a bigger memory bus, but they also want the pricing from the value-focused generations. And they tell themselves that really things are actually advancing 50% per gen (because they see that sometimes/in some segments) and it's just evil corporations not letting them have it. If, instead, you smooth the progression, so you get 20% then 25% over those same 2 gens, there is no bones made about the actual progress being made etc. Consumers have predictability and information, and they can thus make optimal, informed decisions about which generations they are going to upgrade (which I think everyone agrees at this point isn't "every single generation" for literally anyone anymore). it's functionally no different in terms of progression, you are getting the same averaged gains year on year. It just means "don't launch the GTX 970 to be some super killer deal, because you have to top it again next year". And if next year you only get 10% over the absolute best sku in the lineup, you failed, your card is shit (or at least "disappointing"). you see this constantly with the 3060 Ti: the 4060 is deemed a failure because it doesn't leapfrog over the 3060 ti (the literal best-value card since RX 480 etc). And the 4060 Ti is a failure because it doesn't leave the 3060 Ti in the dust ("only" 3070 performance, with twice the VRAM). Etc etc. Stop doing that, just make normalized skus with consistent yearly improvement. That's what reviewers are asking for, and tbh they're not entirely wrong on that part (even though their overall performance/cost expectations are just completely bonkers). They're imagining it's going to be 50% a gen, not 20% a gen, but they're not wrong about how the public does need some consistency in their life (like any child) as far as generational gains. https://youtu.be/QbNwO4HXTIk https://youtu.be/QbNwO4HXTIk The market simply doesn't reward you for making a 3060 Ti or whatever anymore, because that is taken for granted. You can make the same argument about the trailing-node strategy (for 20/30 series) as a whole, which legitimately bent the price curve downwards for consumers. But consumers will never reward you for a price increase you didn't make - in fact they will be extra mad when you finally bounce back to the "normal" cost curve, because now it feels like an extra big increase all at once. Personally yeah, I'd rather have the god-tier value card and just understand that the successor is going to be feature/performance/efficiency focused and not value-focused. As high-information consumers we can make those decisions and claw back some consumer surplus. But that's not where the public is. And people have been hopping mad about it for years now, this is not an unusual sentiment. Intel basically exemplified this mindset perfectly: every generation you got X%, and they aligned X% so that they'd get a similar number on every single generation. Going (eg) Sandy Bridge to Skylake is much "lumpier". And people dump on that era, but they also are (functionally) clamoring for that same consistency in the GPU market. "You get 15% a year whether you need it or not". Right now they don't know what to expect, so they can imagine whatever they want, which only sets things up for disappointment.
- Nullabillity 2y ago
  > As Tim Hardwareunboxed says - if you don’t have a good uplift, don’t release the product. Intel made sure they had an uplift every year. You can’t change actual progress but it can be smooth. I don't think that quote means what you think it does. If you don't have meaningful uplift every year… don't try to release a new generation every year.

benreesman 2y ago

I agree it's a weird world where Intel is our best hope for free, fair, and competitive markets. That's a really fucking weird world. But Gelsinger looks at a cursory glance to be someone who knows how to build chips, has a pretty clear comeback story, and realizes that complacency in a monopoly is what made them uncompetitive in the end, and therefore desperate enough to put a qualified person in charge [1]. I think he might keep the pressure on even if they start winning. [1] https://en.wikipedia.org/wiki/Pat_Gelsinger https://en.wikipedia.org/wiki/Pat_Gelsinger

paulmd 2y ago

> Let’s just hope that this time around the they don’t intentionally hold back innovations in favour of slow-dripping a predictable fixed X% gain per year like they did 2014 to 2020 in the absence of competition. idk why you would cite this as a downside when this is specifically what the tech review community has been begging for with gpus. Having a gen do 40% and the next do 0–10% is grating to consumers and disruptive to the market. What consumers need is exactly this smoothing of generational progression and the removal of huge outliers in terms of generational step. Smooth the curve - which inherently will be much closer to that 15-20% progress generationally, given how slow the transitor tech is moving nowadays. As Tim Hardwareunboxed says - if you don’t have a good uplift, don’t release the product. Intel made sure they had an uplift every year. You can’t change actual progress but it can be smooth.

transcriptase 2y ago
Why would I or anyone else care what the tech review community has to say about anything? What kind of idiot wants technological advances held back so they don’t get disappointed and the market doesn’t get disrupted?
- paulmd 2y ago
  > Why would I or anyone else care what the tech review community has to say about anything? like, are you asking why thought leaders matter? > What kind of idiot wants technological advances held back so they don’t get disappointed and the market doesn’t get disrupted? Lots of people, have you never posted in an nvidia thread before? tons of people would much rather RTX, DLSS, and AI never existed at all, because it means they have to upgrade their GTX 1080 Ti or RX 480 or whatever. A lot of people bought their first GPU since that launch and have never lived through a hard upgrade cycle. recently it’s turned into this insane “if you can’t make cards like 3050 for $110 or get at least double the perf over 2 generations then don’t make a new generation at all” Neo-Luddism but it’s been brewing for a long time among the gaming public. https://m.youtube.com/watch?v=QbNwO4HXTIk&feature=youtu.be https://m.youtube.com/watch?v=QbNwO4HXTIk&feature=youtu.be It's obviously dumb and regressive, literal modern luddism, but it's crazy to pretend that it's not an extremely visible and popular modern trend, either. People are mad the low end is stagnating and the high end is continuing to progress, and the natural human response is “if I don’t get to have a good time I don’t see why anyone else should either”. “Don’t launch a 5090 until there’s a good card for $110” is a very mainstream phenomenon nowadays. Literally the drum the 2 biggest and most popular tech channels have been beating for a year now.
  - benreesman 2y ago
    I don’t think a vocal minority under the sway of what it sounds like you’re describing as ill-intentioned propaganda (I mean, that’s a serious word but the shoe seems to fit) is a particularly good proxy for the overall preferences and welfare of everyone in the ecosystem. To the extent that legitimate enthusiasts and early adopters and people trying to get the most out of their gear under resource constraints have a clear consensus, yeah, you want to be listening carefully. But that doesn’t really sound like what you’re describing. I hang out more on the “accelerated compute” parts of the “GPU Internet” than the “gaming” parts, but certainly over here in LocalLLaMA land, we want working, documented firmware, and cards that aren’t overkill on some things and weak AF on others (4090). That’s an effort to manipulate nosebleed margins by making the ostensible “prosumer/enthusiast” offering just kind of bad at everything (Quattro) except maybe CAD or Blender. In gaming it seems like a cost-constrained public would be best served by credible offerings ranging from “the minimum to play modern stuff at a very attractive price” (I remember the 1060 being very well-received here), up to “Crysis at a million FPS” at Veblen, elaborate bespoke case, over clock with active cooling, and a sensible set of options along the way. I don’t do enough gaming these days to have strong views, so please educate me if you’re an avid gamer and I’m full of shit, but both major gaming vendors seem to have good options for most budgets on mainstream platforms. That’s just nothing to do with the 2T cap and 55.58% Net Profit Margins us llama.cpp folks are up in arms about.
benreesman 2y ago
I have a certain sympathy for all involved in your argument, but I don't think it's a good way to do business for either businesses and their executives, their shareholders, or ultimately the consumer and public. The problem is short-term thinking around annual or even quarterly earnings reports and whatever drives that. I appreciate that a CEO or even a board may feel pressure up to and including losing their jobs if they leave literally anything on the table in the form of long-term investment, the "goodwill" (which is a real financial accounting line item that's sometimes a lot) of their customers, partners, etc. and the long-term sustainability of the business. But it's that ability to deliver that magic combination of a product that people ideally love buying, but at least don't mind buying, a team that is healthy, well-compensated, and inclined to go the extra mile for the employer at rare need because the mutual priority of the wellbeing of the partnership goes both ways (this is massive, when you treat people well when they're in a bind, they'll go the extra mile when you're in a bind with little if any pressure, most of the time), and a consistent, low-beta (for the investor's risk appetite) return on investment for shareholders. This is the kind of business that e.g. Warren Buffet at least talks about as the kind of business he learned to invest in. I'm not a Buffet scholar or watcher really, so I don't know how well he lives up to those ideals in general or especially recently, but it's certainly a pretty good way to describe the desired outcome of capitalism! I've heard lots of arguments for why boards and CEOs and the like simply have to skin the sheep once rather than shear him many times (to borrow the old saw): but never a very convincing one. It's basically absurd in the kinds of dual-class share structures that are common in tech now: it's not only extant but common for the founder/CEO to be basically unencumbered by quarterly earnings reports, what's e.g. Mark Zuckerberg going to do? Fire himself? And under this more "good for everyone over a long time horizon" model, you can just launch the next breakthrough card when it's done, and launch incremental improvements when you're out of inventory, or see a gap in your lineup, or whatever. Fix corporate governance in what should be the relatively uncontroversial ways I and many others have described, and you fix this (among many, many other) set of perverse incentives.
- paulmd 2y ago
  I don't think anyone's under any illusions that people are going to upgrade more frequently, but the lack of predictability is pissing people off and miscalibrating expectations on what they "should" be getting every gen. Going 10%,40% is literally worse and more disruptive (in the bad way) to consumer psychology than if you'd just gone 20%,25%, because people will lock in on the most desirable possible combinations of perf gains and value offerings/prices. the x70 tier launched at $349 in 2010 dollars and has bounced between $349 and $399 for every generation, but the one people remember is the singular time it was $329. And the x70 die has almost always fluttered around 250-300 mm2, except for the trailing-node era (when NVIDIA got cheap TSMC/samsung and chased price) where the 2060 was almost as big as a 1080 Ti (445mm2 vs 471mm2) etc. But people have never shut up about how the 4070 isn't a real x70 because it's only 300mm2 etc - they latch onto the one singular time it was bigger etc. what you end up with in terms of node progress is similar - people tell themselves they deserve the gen when everything increases by 50%, at the best possible efficiency, with a bigger memory bus, but they also want the pricing from the value-focused generations. And they tell themselves that really things are actually advancing 50% per gen (because they see that sometimes/in some segments) and it's just evil corporations not letting them have it. If, instead, you smooth the progression, so you get 20% then 25% over those same 2 gens, there is no bones made about the actual progress being made etc. Consumers have predictability and information, and they can thus make optimal, informed decisions about which generations they are going to upgrade (which I think everyone agrees at this point isn't "every single generation" for literally anyone anymore). it's functionally no different in terms of progression, you are getting the same averaged gains year on year. It just means "don't launch the GTX 970 to be some super killer deal, because you have to top it again next year". And if next year you only get 10% over the absolute best sku in the lineup, you failed, your card is shit (or at least "disappointing"). you see this constantly with the 3060 Ti: the 4060 is deemed a failure because it doesn't leapfrog over the 3060 ti (the literal best-value card since RX 480 etc). And the 4060 Ti is a failure because it doesn't leave the 3060 Ti in the dust ("only" 3070 performance, with twice the VRAM). Etc etc. Stop doing that, just make normalized skus with consistent yearly improvement. That's what reviewers are asking for, and tbh they're not entirely wrong on that part (even though their overall performance/cost expectations are just completely bonkers). They're imagining it's going to be 50% a gen, not 20% a gen, but they're not wrong about how the public does need some consistency in their life (like any child) as far as generational gains. https://youtu.be/QbNwO4HXTIk https://youtu.be/QbNwO4HXTIk The market simply doesn't reward you for making a 3060 Ti or whatever anymore, because that is taken for granted. You can make the same argument about the trailing-node strategy (for 20/30 series) as a whole, which legitimately bent the price curve downwards for consumers. But consumers will never reward you for a price increase you didn't make - in fact they will be extra mad when you finally bounce back to the "normal" cost curve, because now it feels like an extra big increase all at once. Personally yeah, I'd rather have the god-tier value card and just understand that the successor is going to be feature/performance/efficiency focused and not value-focused. As high-information consumers we can make those decisions and claw back some consumer surplus. But that's not where the public is. And people have been hopping mad about it for years now, this is not an unusual sentiment. Intel basically exemplified this mindset perfectly: every generation you got X%, and they aligned X% so that they'd get a similar number on every single generation. Going (eg) Sandy Bridge to Skylake is much "lumpier". And people dump on that era, but they also are (functionally) clamoring for that same consistency in the GPU market. "You get 15% a year whether you need it or not". Right now they don't know what to expect, so they can imagine whatever they want, which only sets things up for disappointment.
Nullabillity 2y ago
> As Tim Hardwareunboxed says - if you don’t have a good uplift, don’t release the product. Intel made sure they had an uplift every year. You can’t change actual progress but it can be smooth. I don't think that quote means what you think it does. If you don't have meaningful uplift every year… don't try to release a new generation every year.