Don't care about the argument here, just want to know what this means for gqmers.
It means we'll have to wait for more info.
All this speculation is exciting, but at the moment all that we really know for certain is that we don't know enough to make any purchase decisions yet.
And the stock price is what nVidia really cares about. They won't sell enough of the x80's to really move revenue, it's a halo product. But they want it to kick butt so you think of nVidia first and as the premium product when you go to buy whatever it is you are going to buy. And they will sell hundreds of thousands of x60's and x70s - where the real money is, and where they actually have competition.
Nvidia isn't the only company with a stock price that they care about. Micron, Hynix, and Samsung are publicly traded, too. They would all love to announce that they have GDDR6X memory available to sell before their competitors, or that it is sampling before their competitors. Samsung is in a lot of other markets, but Micron and Hynix are pretty much memory companies. If announcements aren't forthcoming, it's probably because the memory isn't ready.
Actually, I might as well explain what I mean by "sampling". The first time you try to make a new chip, it basically never works well enough to sell it commercially. You might have some chips that are good enough to sell, but most of them aren't. You might have some that work, but not at the intended clock speeds. You might not have any that entirely work at all.
Once you have physical chips, you can test them to see what is wrong, then modify your design to try to fix it. It can take months for chips to go through the fabs, so you really only get one new test every couple of months or so, at least if you want to see what went wrong with the previous design before creating a new one.
Chips are "sampling" when you can say, we have some working chips, but just not very many of them. Sampling lets you give a handful of chips to your partners so that they can test them with their own chips. That is, here are some working GDDR6X memory chips so that you can test them with a GPU that is intended to use GDDR6X memory. That way, Nvidia doesn't have to wait until the new memory is in mass production before they test their own memory controller--and potentially have to do a respin or two of their own if it doesn't work.
If only 10% of the memory chips work, that's plenty enough for sampling. That's not enough for mass production, as you'd have to build 10 chips for everyone one you sell. So while sampling is a major milestone, it's nowhere near being done. After you announce that chips are sampling, you go back and try to fix whatever is wrong. Maybe you do a respin and then get 50% yields--a lot better, but still not good enough for mass production. Then you do another respin and get 90% yields. Then you decide that's good enough for mass production, you scale up production dramatically, and a few months later, you have large amounts of chips available to sell.
The time gap between sampling and mass production can vary a lot. It's generally going to be at least months at a bare minimum, as even if the first design that is good enough for sampling is also good enough for mass production, you don't start thousands of wafers for a production run until after you have back the first wafers to know that it works. Depending on how many respins you need to do to get good enough yields for mass production, the time gap between sampling and mass production is likely to be several months, and can be over a year.
If memory isn't yet sampling, then you can't build working video cards based on it at all. Not review samples, not engineering samples, not demos, nothing. You can build a mock up that doesn't do anything, but that's about it. The memory you need doesn't yet exist in the real world.
If the memory you need is sampling but not yet in production, then you can build a handful of working cards based on it. That's plenty enough for a demo or even some review samples for a paper launch. You need a real production run for more than that.
Even once memory is in full production, that doesn't mean that you can immediately do a hard launch of cards based on it. It takes time to transport memory chips, attach them to PCBs, ship cards around the world, and so forth. It also takes time to accumulate a significant stock of cards so that they don't immediately sell out on launch day.
All I am saying is that cards will launch with GDDR6X on them in September from Nvidia period. Not GDDR6.
Quizzical said: There are two options, and GDDR6X isn't one of them.
I would be shocked if Ampere is all GDDR6X throughout the entire lineup, or even just for the consumer cards. Look at what vendors have done in the past when potentially waiting on new memory standards. AMD's RV770 chip used the then very new GDDR5 for the Radeon HD 4870, but GDDR3 for the Radeon HD 4850 that launched at the same time. Similarly, Nvidia's GP104 chip used then-new GDDR5X for the GeForce GTX 1080, but more mature GDDR5 for the GeForce GTX 1070. Meanwhile, their lower end cards in the lineup used only the older memory standards. For example, the Radeon HD 4670 used GDDR3 and not GDDR5, while the GeForce GTX 1060 used GDDR5 and not GDDR5X. In both cases, they later launched additional cards that used only the new memory standard.
I would expect Nvidia to do something like that with Ampere, and for the same reasons. If GDDR6X is delayed, then Nvidia doesn't want to be stuck with nothing to sell for several months after their intended launch. It's better to at least have something that works based on the older memory standard.
Furthermore, GDDR6X is probably going to be expensive precisely because it's so new. For high end cards where you have to have it (or something more expensive yet such as HBM2) in order to get the performance you need, you pay what it costs and pass that cost on to consumers. But for lower end cards, trying to get consumers to pay $250 for a video card that isn't any better than your competitor's $200 card doesn't tend to work very well.
AmazingAveryAge of Conan AdvocateMemberUncommonPosts: 7,188
I would expect GDDR6X on the top SKUs sure. Like from the 070 and up and Ti variants. The fact there are already pictures out there showing that.
Also expect RT throughout the whole stack with much improved performance.
AMD could still upstage them with big Navi as I don't expect much availability for these cards until 2021, especially the lesser-priced ones. The availability of GDDR6X will limit it and Nvidia will push the high end cards where the profit margin is high.
I agree, but there's a double standard. AMD has to be faster and cheaper in order to be considered better.
I don't know exactly why that is, but it just is. If AMD isn't the fastest, even if that isn't the card you will be buying, then it's a loser.
AMD has a really nice card in the 5700. But it couldn't really compete against a 1080 or 1080Ti, so a lot of people wrote them off this generation.
In most low to mid tiers in the last few generations, AMD has had the better performing part for equal or less money -- except the highest end, where it just hasn't been able to compete recently.
nVidia commands a premium for their brand, and people are happy to pay it. The halo effect is not the only thing going on there, but it's definitely real. I still love the driver argument that inevitably comes up.
I know a lot of people are just going to look at the 3090 or whatever they call it, and compare it to AMD's Navi 2. And if AMD doesn't blow it out of the water performance-wise, will keep their Team Green jersey, even if it does cost $1,500 MSRP, has next to no availability, and they have no intention of buying it.
I've got 3 gaming computers in my house (myself, wife, kid). At least two of them are due for upgrades soon. The oldest is long past due (it's running a RX470), the other is starting to flake out (GTX980). I won't rule out nVidia, one of those two that needs an upgrade is an nVidia now - it's all a matter of what cards are available at price/performance levels when I go to buy. Most of the time, I'm buying in a bracket where AMD wins out, but not always.
Comments
All this speculation is exciting, but at the moment all that we really know for certain is that we don't know enough to make any purchase decisions yet.
Actually, I might as well explain what I mean by "sampling". The first time you try to make a new chip, it basically never works well enough to sell it commercially. You might have some chips that are good enough to sell, but most of them aren't. You might have some that work, but not at the intended clock speeds. You might not have any that entirely work at all.
Once you have physical chips, you can test them to see what is wrong, then modify your design to try to fix it. It can take months for chips to go through the fabs, so you really only get one new test every couple of months or so, at least if you want to see what went wrong with the previous design before creating a new one.
Chips are "sampling" when you can say, we have some working chips, but just not very many of them. Sampling lets you give a handful of chips to your partners so that they can test them with their own chips. That is, here are some working GDDR6X memory chips so that you can test them with a GPU that is intended to use GDDR6X memory. That way, Nvidia doesn't have to wait until the new memory is in mass production before they test their own memory controller--and potentially have to do a respin or two of their own if it doesn't work.
If only 10% of the memory chips work, that's plenty enough for sampling. That's not enough for mass production, as you'd have to build 10 chips for everyone one you sell. So while sampling is a major milestone, it's nowhere near being done. After you announce that chips are sampling, you go back and try to fix whatever is wrong. Maybe you do a respin and then get 50% yields--a lot better, but still not good enough for mass production. Then you do another respin and get 90% yields. Then you decide that's good enough for mass production, you scale up production dramatically, and a few months later, you have large amounts of chips available to sell.
The time gap between sampling and mass production can vary a lot. It's generally going to be at least months at a bare minimum, as even if the first design that is good enough for sampling is also good enough for mass production, you don't start thousands of wafers for a production run until after you have back the first wafers to know that it works. Depending on how many respins you need to do to get good enough yields for mass production, the time gap between sampling and mass production is likely to be several months, and can be over a year.
If memory isn't yet sampling, then you can't build working video cards based on it at all. Not review samples, not engineering samples, not demos, nothing. You can build a mock up that doesn't do anything, but that's about it. The memory you need doesn't yet exist in the real world.
If the memory you need is sampling but not yet in production, then you can build a handful of working cards based on it. That's plenty enough for a demo or even some review samples for a paper launch. You need a real production run for more than that.
Even once memory is in full production, that doesn't mean that you can immediately do a hard launch of cards based on it. It takes time to transport memory chips, attach them to PCBs, ship cards around the world, and so forth. It also takes time to accumulate a significant stock of cards so that they don't immediately sell out on launch day.
I would expect Nvidia to do something like that with Ampere, and for the same reasons. If GDDR6X is delayed, then Nvidia doesn't want to be stuck with nothing to sell for several months after their intended launch. It's better to at least have something that works based on the older memory standard.
Furthermore, GDDR6X is probably going to be expensive precisely because it's so new. For high end cards where you have to have it (or something more expensive yet such as HBM2) in order to get the performance you need, you pay what it costs and pass that cost on to consumers. But for lower end cards, trying to get consumers to pay $250 for a video card that isn't any better than your competitor's $200 card doesn't tend to work very well.
The fact there are already pictures out there showing that.
Also expect RT throughout the whole stack with much improved performance.
I don't know exactly why that is, but it just is. If AMD isn't the fastest, even if that isn't the card you will be buying, then it's a loser.
AMD has a really nice card in the 5700. But it couldn't really compete against a 1080 or 1080Ti, so a lot of people wrote them off this generation.
In most low to mid tiers in the last few generations, AMD has had the better performing part for equal or less money -- except the highest end, where it just hasn't been able to compete recently.
nVidia commands a premium for their brand, and people are happy to pay it. The halo effect is not the only thing going on there, but it's definitely real. I still love the driver argument that inevitably comes up.
I know a lot of people are just going to look at the 3090 or whatever they call it, and compare it to AMD's Navi 2. And if AMD doesn't blow it out of the water performance-wise, will keep their Team Green jersey, even if it does cost $1,500 MSRP, has next to no availability, and they have no intention of buying it.
I've got 3 gaming computers in my house (myself, wife, kid). At least two of them are due for upgrades soon. The oldest is long past due (it's running a RX470), the other is starting to flake out (GTX980). I won't rule out nVidia, one of those two that needs an upgrade is an nVidia now - it's all a matter of what cards are available at price/performance levels when I go to buy. Most of the time, I'm buying in a bracket where AMD wins out, but not always.