3 ms·

If they arent making 11bn on 19bn as the article says, then I'm wrong and the whole thing is withdrawn -- that's just a premise. I take it that they are howeve

If they arent making 11bn on 19bn as the article says, then I'm wrong and the whole thing is withdrawn -- that's just a premise. I take it that they are however, since their pricing seems extraordinarily out of keeping with the cost of manufacture. But if the article is wrong, or misleading, so be it. I don't see signs of a undue market power at 10-20% profit, just a little inefficient in the competitive landscape

paulmd 3y ago
11bn on 19bn is only because of AI sales - when you are selling $40k enterprise chips for $100k it's going to push margins up. The 2022 numbers are before AI kicked in, and those numbers are roughly comparable to what Radeon Group makes - and that's with NVIDIA mixing in an already-decent amount of enterprise sales (which pushes margins up). Their pricing is pretty much in-line with the way costs have risen in the industry. GTX 670 launched at $399 for a 300mm2 cutdown in 2012, GTX 1070 launched at $449 (the "non-founders MSRP" was widely decried as fake at the time since partners didn't follow it) for a 300mm2 cutdown in 2016. $599 for 4070 is really about on the overall cost-trend curve for a product that's launched on a leading node. https://old.reddit.com/r/Amd/comments/4navtq/amd_this_is_how_nvidia_launched_their_pascal/ https://old.reddit.com/r/Amd/comments/4navtq/amd_this_is_how... https://hardforum.com/threads/geforce-gtx-1080-most-bizarre-secret-paper-launch-ever-h.1899892/ https://hardforum.com/threads/geforce-gtx-1080-most-bizarre-... https://forums.anandtech.com/threads/hardocp-geforce-gtx-1080-most-bizarre-secret-paper-launch-ever.2473396/ https://forums.anandtech.com/threads/hardocp-geforce-gtx-108... Ampere and Turing both were trailing-node products - NVIDIA was trying to keep the cost down at a moment when TSMC was cranking prices and 7nm capacity was scarce. But once they moved back to the leading edge, prices "snapped back" from under-market to at-market. And I think you can certainly argue that this was a mistake, because it mis-calibrates people's expectations and the snap-back feels much worse all at once than as a gradual 3-gen increase. This isn't to say stuff like 4080 weren't obscene, 4070 Ti and 4060 Ti were only good by comparison. But by the time we are getting to 4060 and 4070 launches NVIDIA is pretty much following the historical cost trends, and AMD is not able to drastically undercut these prices (because they're not particularly out-of-line). That is the problem - people look at trailing-node products like Ampere and Turing and Maxwell and say "well, these dies were bigger and cost less" and yeah, if you are accepting something that's 2 nodes behind leading-edge it will cost less. You also don't get a 4090 or GTX Titan-class product that towers 50% above the 4080/680, and you get much worse efficiency, etc. And of course the dies will inherently be bigger, because they're lower density... 2060 being 440mm2 or whatever is absolutely not the norm for a x60 class card, it's a massive outlier cherrypicked to make a point. And remember - it's not like AMD is giving you a 440mm2 die on a x60 class product either, historically or presently. For some reason the expectation of 440mm2 dies on x60 being a perpetual thing only applies to NVIDIA. Nobody made a 2-part video essay series about how the RX 5700XT was a "midrange product being sold at flagship prices" or whatever - and it's significantly smaller than 1080! Literally half the size of a 2060 - surely that’s a low-end card? https://www.youtube.com/watch?v=Jd1bp9eSfwo https://www.youtube.com/watch?v=Jd1bp9eSfwo RX 5700 XT was $399 for a 251mm die in 2019. 6700XT was $479 for a 335mm2 die in 2021 - and while yes, "mining prices", it probably wasn't inflated as much as people want it to be (the "it was mining" has become a Rorschach test for people to project whatever prices they want to imagine it should have been). And historically AMD has had to come in underneath NVIDIA anyway: if AMD was charging $480 then NVIDIA would have been charging $550 or $600 for the same "silicon tier". (Oh, and of course that's all pre-inflation, etc - $399 in 2012 is $539 in 2024. $399 in 2019 is $477 in 2024. And wafer prices have risen a lot since even 7nm!) That's the problem - Ada prices are not really "extraordinarily out of keeping with the cost of manufacture". People don't see that prices have already been creeping up over time (people don't even bother adjusting for CPI let alone the trends of silicon cost), and the cost trend has very much accelerated in the finFET era (post-28nm). Everyone is like "what if moore's law ends" and doesn't realize that it's ended 10 years ago already, and vendors are out of tricks to avoid the cost spiral. Yeah, 4080 was a gouge, but 4090/4070/4060 prices are pretty much just what a leading-edge product looks like/costs now. And it's almost exactly the same price and size as GTX 670, for example, 10 years later. It is the same problem as climate change - a large amount of people are blind-faith believers that moore's law will never end, and vendors devote a significant amount of effort to keep prices down (see: turing, ampere) which further allows people to continue believing this. Then we run out of runway and have a "snap-back to reality" moment - we're blowing past the worst-case estimates for transistors-per-$ gains and you have a large group of the population that still doesn't realize it's happening at all. And of course there is a marketing machine devoted to the whole "maybe moore's law is dead for them, but not for us!!!" type shit (not just AMD but they're a convenient and credible example), but once you get past the headline, even AMD is saying that costs are ballooning and future products won't meet the cost trends that define moore's law. They just mean performance will keep scaling - but so will prices. https://www.pcgamer.com/amd-moores-law-aint-dead-its-just-a-whole-lot-more-expensive/ https://www.pcgamer.com/amd-moores-law-aint-dead-its-just-a-... > "It's not that there's not going to be exciting new transistor technologies. Actually, I can see exciting new transistor technology for the next—as far as you can really plot these things out, is about six to eight years, and it's very, very clear to me the advances that we're going to make to keep improving the transistor technology, but they're more expensive." > The difference now, Papermaster explains, is that where you used to get double the transistor density every same year while costs remained largely the same for a given chip size, the cost per area of silicon is increasing with each successive production node. Computer chips of a given size are becoming much more expensive. That is the blunt reality: cost-per-transistor is no longer coming down much at all. Shrinking increases density, but also increases cost at almost the same rate. Packaging can fix yields and reduce power, but it doesn't fix rising wafer costs entirely. Even if you yield at 100%, adding more silicon still costs more. So in the long term you won't be able to get higher performance levels unless you raise performance-per-transistor. And that's literally been a 10+ year effort for NVIDIA - Kepler and Maxwell deconstructed the entire accelerator down to barebones, really too far (and some of the ways they stripped it down had to be backed out in Pascal). Turing brought in DLSS and tensor to try and provide a "general-purpose accelerator". Ada brought in framegen. And in fact gamers have generally pissed and moaned about how none of this was good enough, and NVIDIA needed to just keep scaling raw raster. But the low-hanging fruit has already been picked from 2014-2016, you can't squeeze infinite performance gains year after year from a fixed number of transistors. https://www.anandtech.com/show/5699/nvidia-geforce-gtx-680-review/3 https://www.anandtech.com/show/5699/nvidia-geforce-gtx-680-r... https://www.anandtech.com/show/8526/nvidia-geforce-gtx-980-review/2 https://www.anandtech.com/show/8526/nvidia-geforce-gtx-980-r... Anyway: yeah they're making bank on the AI craze. Rightly so - they've led the way on bringing GPGPU to the masses, pushed GPGPU as a concept and built the software ecosystem, and bet heavily on AI at a time (2014-ish) when it was not an obvious call at all. They weren't just in the right place at the right time with the right product, it's not even that they innovated some hardware advancements to make it happen, they literally built the entire segment from scratch so yeah, they're the defacto leader when there suddenly is a killer app and their decades of work pays off. They're making bank in the short term because of that. https://news.ycombinator.com/item?id=32983297 https://news.ycombinator.com/item?id=32983297 But this generally has nothing to do with your GeForce prices. Nor are they particularly out of line anymore - you're just miscalibrated on what things cost in 2024. There's no more 1600AF for $85 either. Things cost more now. Electronics cost a lot more now. Some of it is TSMC, some of it is physics. And the market is generally efficient enough that AMD can't afford to drastically sweep in under NVIDIA pricing.

paulmd 3y ago

11bn on 19bn is only because of AI sales - when you are selling $40k enterprise chips for $100k it's going to push margins up. The 2022 numbers are before AI kicked in, and those numbers are roughly comparable to what Radeon Group makes - and that's with NVIDIA mixing in an already-decent amount of enterprise sales (which pushes margins up). Their pricing is pretty much in-line with the way costs have risen in the industry. GTX 670 launched at $399 for a 300mm2 cutdown in 2012, GTX 1070 launched at $449 (the "non-founders MSRP" was widely decried as fake at the time since partners didn't follow it) for a 300mm2 cutdown in 2016. $599 for 4070 is really about on the overall cost-trend curve for a product that's launched on a leading node. https://old.reddit.com/r/Amd/comments/4navtq/amd_this_is_how_nvidia_launched_their_pascal/ https://old.reddit.com/r/Amd/comments/4navtq/amd_this_is_how... https://hardforum.com/threads/geforce-gtx-1080-most-bizarre-secret-paper-launch-ever-h.1899892/ https://hardforum.com/threads/geforce-gtx-1080-most-bizarre-... https://forums.anandtech.com/threads/hardocp-geforce-gtx-1080-most-bizarre-secret-paper-launch-ever.2473396/ https://forums.anandtech.com/threads/hardocp-geforce-gtx-108... Ampere and Turing both were trailing-node products - NVIDIA was trying to keep the cost down at a moment when TSMC was cranking prices and 7nm capacity was scarce. But once they moved back to the leading edge, prices "snapped back" from under-market to at-market. And I think you can certainly argue that this was a mistake, because it mis-calibrates people's expectations and the snap-back feels much worse all at once than as a gradual 3-gen increase. This isn't to say stuff like 4080 weren't obscene, 4070 Ti and 4060 Ti were only good by comparison. But by the time we are getting to 4060 and 4070 launches NVIDIA is pretty much following the historical cost trends, and AMD is not able to drastically undercut these prices (because they're not particularly out-of-line). That is the problem - people look at trailing-node products like Ampere and Turing and Maxwell and say "well, these dies were bigger and cost less" and yeah, if you are accepting something that's 2 nodes behind leading-edge it will cost less. You also don't get a 4090 or GTX Titan-class product that towers 50% above the 4080/680, and you get much worse efficiency, etc. And of course the dies will inherently be bigger, because they're lower density... 2060 being 440mm2 or whatever is absolutely not the norm for a x60 class card, it's a massive outlier cherrypicked to make a point. And remember - it's not like AMD is giving you a 440mm2 die on a x60 class product either, historically or presently. For some reason the expectation of 440mm2 dies on x60 being a perpetual thing only applies to NVIDIA. Nobody made a 2-part video essay series about how the RX 5700XT was a "midrange product being sold at flagship prices" or whatever - and it's significantly smaller than 1080! Literally half the size of a 2060 - surely that’s a low-end card? https://www.youtube.com/watch?v=Jd1bp9eSfwo https://www.youtube.com/watch?v=Jd1bp9eSfwo RX 5700 XT was $399 for a 251mm die in 2019. 6700XT was $479 for a 335mm2 die in 2021 - and while yes, "mining prices", it probably wasn't inflated as much as people want it to be (the "it was mining" has become a Rorschach test for people to project whatever prices they want to imagine it should have been). And historically AMD has had to come in underneath NVIDIA anyway: if AMD was charging $480 then NVIDIA would have been charging $550 or $600 for the same "silicon tier". (Oh, and of course that's all pre-inflation, etc - $399 in 2012 is $539 in 2024. $399 in 2019 is $477 in 2024. And wafer prices have risen a lot since even 7nm!) That's the problem - Ada prices are not really "extraordinarily out of keeping with the cost of manufacture". People don't see that prices have already been creeping up over time (people don't even bother adjusting for CPI let alone the trends of silicon cost), and the cost trend has very much accelerated in the finFET era (post-28nm). Everyone is like "what if moore's law ends" and doesn't realize that it's ended 10 years ago already, and vendors are out of tricks to avoid the cost spiral. Yeah, 4080 was a gouge, but 4090/4070/4060 prices are pretty much just what a leading-edge product looks like/costs now. And it's almost exactly the same price and size as GTX 670, for example, 10 years later. It is the same problem as climate change - a large amount of people are blind-faith believers that moore's law will never end, and vendors devote a significant amount of effort to keep prices down (see: turing, ampere) which further allows people to continue believing this. Then we run out of runway and have a "snap-back to reality" moment - we're blowing past the worst-case estimates for transistors-per-$ gains and you have a large group of the population that still doesn't realize it's happening at all. And of course there is a marketing machine devoted to the whole "maybe moore's law is dead for them, but not for us!!!" type shit (not just AMD but they're a convenient and credible example), but once you get past the headline, even AMD is saying that costs are ballooning and future products won't meet the cost trends that define moore's law. They just mean performance will keep scaling - but so will prices. https://www.pcgamer.com/amd-moores-law-aint-dead-its-just-a-whole-lot-more-expensive/ https://www.pcgamer.com/amd-moores-law-aint-dead-its-just-a-... > "It's not that there's not going to be exciting new transistor technologies. Actually, I can see exciting new transistor technology for the next—as far as you can really plot these things out, is about six to eight years, and it's very, very clear to me the advances that we're going to make to keep improving the transistor technology, but they're more expensive." > The difference now, Papermaster explains, is that where you used to get double the transistor density every same year while costs remained largely the same for a given chip size, the cost per area of silicon is increasing with each successive production node. Computer chips of a given size are becoming much more expensive. That is the blunt reality: cost-per-transistor is no longer coming down much at all. Shrinking increases density, but also increases cost at almost the same rate. Packaging can fix yields and reduce power, but it doesn't fix rising wafer costs entirely. Even if you yield at 100%, adding more silicon still costs more. So in the long term you won't be able to get higher performance levels unless you raise performance-per-transistor. And that's literally been a 10+ year effort for NVIDIA - Kepler and Maxwell deconstructed the entire accelerator down to barebones, really too far (and some of the ways they stripped it down had to be backed out in Pascal). Turing brought in DLSS and tensor to try and provide a "general-purpose accelerator". Ada brought in framegen. And in fact gamers have generally pissed and moaned about how none of this was good enough, and NVIDIA needed to just keep scaling raw raster. But the low-hanging fruit has already been picked from 2014-2016, you can't squeeze infinite performance gains year after year from a fixed number of transistors. https://www.anandtech.com/show/5699/nvidia-geforce-gtx-680-review/3 https://www.anandtech.com/show/5699/nvidia-geforce-gtx-680-r... https://www.anandtech.com/show/8526/nvidia-geforce-gtx-980-review/2 https://www.anandtech.com/show/8526/nvidia-geforce-gtx-980-r... Anyway: yeah they're making bank on the AI craze. Rightly so - they've led the way on bringing GPGPU to the masses, pushed GPGPU as a concept and built the software ecosystem, and bet heavily on AI at a time (2014-ish) when it was not an obvious call at all. They weren't just in the right place at the right time with the right product, it's not even that they innovated some hardware advancements to make it happen, they literally built the entire segment from scratch so yeah, they're the defacto leader when there suddenly is a killer app and their decades of work pays off. They're making bank in the short term because of that. https://news.ycombinator.com/item?id=32983297 https://news.ycombinator.com/item?id=32983297 But this generally has nothing to do with your GeForce prices. Nor are they particularly out of line anymore - you're just miscalibrated on what things cost in 2024. There's no more 1600AF for $85 either. Things cost more now. Electronics cost a lot more now. Some of it is TSMC, some of it is physics. And the market is generally efficient enough that AMD can't afford to drastically sweep in under NVIDIA pricing.