3 ms·

Then why wouldn't they just make more of them? The excuse is supposed to be fab capacity, but the 3070 has better performance than the 4060 etc. and is built on

Then why wouldn't they just make more of them? The excuse is supposed to be fab capacity, but the 3070 has better performance than the 4060 etc. and is built on the older process which should no longer be in short supply. There is actually very little margins in the midrange consumer discrete GPU market. The market for discrete GPUs have been shrinking since the mid 2000s.[0] Most GPUs are sold as integrated such as SoCs and in consoles nowadays. In a shrinking market, the midrange and low-end products will cease to be profitable. Hence, Nvidia's 60/70 offerings are lackluster because they don't make much money from them. They want you to buy the 80s and 90s cards. Furthermore, node advancements have stopped scaling $/transistor. So the transistors aren't getting cheaper, just smaller. Lastly, Nvidia wants to allocate every last wafer they pre-purchased from TSMC to their server GPUs. [0]https://d15shllkswkct0.cloudfront.net/wp-content/blogs.dir/1/files/2022/12/GPU_Dev_graph_01.png https://d15shllkswkct0.cloudfront.net/wp-content/blogs.dir/1...

AnthonyMouse 3y ago
> Most GPUs are sold as integrated such as SoCs and in consoles nowadays. But those are mostly AMD, and doesn't really have anything to do with what features someone puts on their discrete gaming cards, except insofar as it implies gamers don't need the cards some amateur ML hobbyist might buy. > In a shrinking market, the midrange and low-end products will cease to be profitable. That's assuming the products have high independent development costs, but that isn't really the case. The low end products are essentially the high end products with fewer cores which use correspondingly less silicon -- which have higher yields because you don't need such a large area of perfect silicon or can sell a defective die as a slower part by disabling the defective section, making them profitable with a smaller margin per unit die area. > Furthermore, node advancements have stopped scaling $/transistor. So the transistors aren't getting cheaper, just smaller. Which implies that they can profitably continue producing almost-as-good GPUs on the older process node. > Lastly, Nvidia wants to allocate every last wafer they pre-purchased from TSMC to their server GPUs. Which is why the proposal is for them to make as many GPUs as the gamers could want at Samsung.
- aurareturn 3y ago
  Customers who want midrange or low end GPUs are price sensitive. Therefore, midrange and low end GPUs are a low margin business. Hence, there's little to no reason to provide a very compelling product in those categories. While midrange GPUs are a cut from highend GPUs, they're still significantly more expensive to manufacture than say CPUs, at a transistor to transistor level. Look at an AMD 7950x transistor count, and then an RTX 4060 transistor count. The GPU has ~50% more transistors but sell at half the price. In addition, the GPU requires RAM, a board, circuitry, and a heatsink fan. The margins simply aren't there for lowend GPUs anymore. Previously, Nvidia and AMD can make it up through volume. But again, the market has gotten much smaller going from 60 million discrete GPUs per year sold to 30 million. That's half! Based on your logic, AMD should feast on midrange and low end discrete GPU market because Nvidia does not have value products there. But AMD isn't feasting. You know why? Because there's also no profit there for AMD either. Once you stop thinking like an angry gamer, these decisions start to make a lot of sense.
  - AnthonyMouse 3y ago
    > Customers who want midrange or low end GPUs are price sensitive. Therefore, midrange and low end GPUs are a low margin business. Customers who want petroleum are price sensitive. Therefore, petroleum exporting is a low margin business. This is why the Saudis make the profit margins they do. Wait, something's not right here. > Look at an AMD 7950x transistor count, and then an RTX 4060 transistor count. The GPU has ~50% more transistors but sell at half the price. You're comparing the high end CPU to the mid-range GPU. The AMD 8500G has more transistors than the RTX 4060 and costs less. > In addition, the GPU requires RAM, a board, circuitry, and a heatsink fan. The 8500G comes with a heatsink and fan. The 8GB of GDDR6 on the 4060 costs $27 but the 4060 costs $120 more. A printed circuit board doesn't cost $93. > Previously, Nvidia and AMD can make it up through volume. But again, the market has gotten much smaller going from 60 million discrete GPUs per year sold to 30 million. That's half! That's not because people stopped buying them, it's because they shifted production capacity to servers. > Based on your logic, AMD should feast on midrange and low end discrete GPU market because Nvidia does not have value products there. But AMD isn't feasting. You know why? Because there's also no profit there for AMD either. But they do though. You can find lower end AMD GPUs from the last two years for $125 (e.g. RX 6400) whereas the cheapest RTX 3000 or 4000 series is around twice that. And anyway who is talking about the bottom end? The question is why they don't produce more of e.g. the RTX 3070, which is on the old Samsung 8LPP process, has fewer transistors than the RTX 4060, is faster, and is still selling for a higher price.

AnthonyMouse 3y ago

> Most GPUs are sold as integrated such as SoCs and in consoles nowadays. But those are mostly AMD, and doesn't really have anything to do with what features someone puts on their discrete gaming cards, except insofar as it implies gamers don't need the cards some amateur ML hobbyist might buy. > In a shrinking market, the midrange and low-end products will cease to be profitable. That's assuming the products have high independent development costs, but that isn't really the case. The low end products are essentially the high end products with fewer cores which use correspondingly less silicon -- which have higher yields because you don't need such a large area of perfect silicon or can sell a defective die as a slower part by disabling the defective section, making them profitable with a smaller margin per unit die area. > Furthermore, node advancements have stopped scaling $/transistor. So the transistors aren't getting cheaper, just smaller. Which implies that they can profitably continue producing almost-as-good GPUs on the older process node. > Lastly, Nvidia wants to allocate every last wafer they pre-purchased from TSMC to their server GPUs. Which is why the proposal is for them to make as many GPUs as the gamers could want at Samsung.

aurareturn 3y ago
Customers who want midrange or low end GPUs are price sensitive. Therefore, midrange and low end GPUs are a low margin business. Hence, there's little to no reason to provide a very compelling product in those categories. While midrange GPUs are a cut from highend GPUs, they're still significantly more expensive to manufacture than say CPUs, at a transistor to transistor level. Look at an AMD 7950x transistor count, and then an RTX 4060 transistor count. The GPU has ~50% more transistors but sell at half the price. In addition, the GPU requires RAM, a board, circuitry, and a heatsink fan. The margins simply aren't there for lowend GPUs anymore. Previously, Nvidia and AMD can make it up through volume. But again, the market has gotten much smaller going from 60 million discrete GPUs per year sold to 30 million. That's half! Based on your logic, AMD should feast on midrange and low end discrete GPU market because Nvidia does not have value products there. But AMD isn't feasting. You know why? Because there's also no profit there for AMD either. Once you stop thinking like an angry gamer, these decisions start to make a lot of sense.
- AnthonyMouse 3y ago
  > Customers who want midrange or low end GPUs are price sensitive. Therefore, midrange and low end GPUs are a low margin business. Customers who want petroleum are price sensitive. Therefore, petroleum exporting is a low margin business. This is why the Saudis make the profit margins they do. Wait, something's not right here. > Look at an AMD 7950x transistor count, and then an RTX 4060 transistor count. The GPU has ~50% more transistors but sell at half the price. You're comparing the high end CPU to the mid-range GPU. The AMD 8500G has more transistors than the RTX 4060 and costs less. > In addition, the GPU requires RAM, a board, circuitry, and a heatsink fan. The 8500G comes with a heatsink and fan. The 8GB of GDDR6 on the 4060 costs $27 but the 4060 costs $120 more. A printed circuit board doesn't cost $93. > Previously, Nvidia and AMD can make it up through volume. But again, the market has gotten much smaller going from 60 million discrete GPUs per year sold to 30 million. That's half! That's not because people stopped buying them, it's because they shifted production capacity to servers. > Based on your logic, AMD should feast on midrange and low end discrete GPU market because Nvidia does not have value products there. But AMD isn't feasting. You know why? Because there's also no profit there for AMD either. But they do though. You can find lower end AMD GPUs from the last two years for $125 (e.g. RX 6400) whereas the cheapest RTX 3000 or 4000 series is around twice that. And anyway who is talking about the bottom end? The question is why they don't produce more of e.g. the RTX 3070, which is on the old Samsung 8LPP process, has fewer transistors than the RTX 4060, is faster, and is still selling for a higher price.