How much easier is it to manufacture a low-end GPU?

[euskalzabe]

The question is simple: how much easier is it to manufacture a low-end GPU than a high-end GPU? Say, how much easier is it to manufacture a 106 GPU than a 102 die for Nvidia?

Now, I get how the die savings work: smaller GPU, you can fit many more dies on the same wafer, and being smaller means a lesser error rate, higher yield. I get the theory. What I don't know, is what about the rest of the components per graphics card. So, I'm wondering, what makes it easier (other than improved yields due to smaller GPU die) to manufacture a lower end GPU that normally sells at MSRP of $200-300?

Any links to websites that explore this would be welcome. It's good information to have in general, but also as a way to extrapolate and speculate how much stock we might expect to see of RX 6700 and RTX 3060 cards.

 

[CraptacularOne]

Don't really have any websites for you but obviously on lower end cards they don't require as robust a VRM area and can use lower end components and just generally the PCB is a simpler design with far less SMDs. Also the cooler doesn't require as much surface area or as many fans.

As for wanting to extrapolate this to estimate any supply or availability of upcoming cards, don't bother. There are other bottlenecks in the way like there is currently a very real substrate shortage (what they mount the GPU die on) there is also the issue these lower margin cards occupy the same die fabs as the higher end, higher profit margin dies so they may not get the same amount of wafers made. There is myriad of other issues that will be at play too like obviously if the card is cheaper it will appeal to an even wider audience so if they were to deliver a shipment for instance that may satisfy the demand for something like a RTX 3090, that same shipment wouldn't even make a dent in the demand for a far more affordable card. You are just setting yourself up for disappointment if you try to draw any conclusions to factors which you have no real direct information on.

 

[euskalzabe]

Thanks for that response! I'm not really trying to draw conclusion for what's coming, I think it's too complex and too many things are going on to "predict" anything - not that I'd be able to. I just had some curiosity as I was thinking about it, regarding what actually makes the lower end boards cheaper to manufacture - besides the obvious simpler GPU. I hadn't even thourhg about VRMs or the actual substrate until you mentioned it.

 

[MrGuvernment]

All that CraptacularOne one said. higher end cards have far lower yields because the level of quality they need to be able to hit and perform on is far higher. As you noted, any parts that do not make the cut for higher end cards can possibly be used in mid to lower end cards instead with features disabled and such.

 

[euskalzabe]

As you noted, any parts that do not make the cut for higher end cards can possibly be used in mid to lower end cards instead with features disabled and such.

Right. It makes sense that 3060 Tis are repurposed 3070 chips that didn't turn out good enough. I'm wondering if we'll see 3060s that are even lower binned 3070s. It sounds like that's the logical path besides an actual ga106 GPU, but I wonder if a 3070 chip binned down to a 3060 would actually lose money for Nvidia. It most likely would mean less profit, otherwise they wouldn't bother making simpler GPU chips and just bin everything down.

 

[madpistol]

Less complex PCB
Smaller GPU die
Less memory (sometimes)
Slower memory
Less robust cooler
Less components on PCB

Why are lower-end GPUs cheaper? Because they require... less.

 

[MrGuvernment]

euskalzabe they already made the wafer, so that money is already spent. Now think of the low / mid range cards they give to OEMs, or when they simply re-brand last gen cards as this gen (AMD has done this a crap ton! and NVIDIA may be doing it with 2000 series inventory now)

 

[E4g1e]

euskalzabe they already made the wafer, so that money is already spent. Now think of the low / mid range cards they give to OEMs, or when they simply re-brand last gen cards as this gen (AMD has done this a crap ton! and NVIDIA may be doing it with 2000 series inventory now)

True for the most part. Especially in the mobile space. For example, the most recent low-end discrete Nvidia MX series GPU, the MX450 from September 2020, is simply a GeForce GTX 1650 mobile GPU with half of its ROPs, NVDEC and NVENC disabled during manufacture.

 

[euskalzabe]

Less complex PCB

Now this is interesting. In the end, the card needs to connect to electricity and VRMs, and the GPU to memory - massive oversimplification obviously. So, if the basic blocks are the same - electricity - VRM - GPU - memory, other than having fewer VRMs due to lower power requirements and less memory blocks... what else can be simplified on the board? I guess it's the electrical engineering perspective that I'm missing here. The "less is cheaper" is obvious, I'm just trying to decipher what that "less" actually is.

True for the most part. Especially in the mobile space. For example, the most recent low-end discrete Nvidia MX series GPU, the MX450 from September 2020, is simply a GeForce GTX 1650 mobile GPU with half of its ROPs, NVDEC and NVENC disabled during manufacture.

That's interesting. Money saved when a low-binned GPU is sold as a cheaper unit makes sense - that GPU would've been trashed otherwise. But disabling ROPs and encoders/decoders that are - likely - not the hard manufacturing part, that work fine... that seems like it's wasted money. I would've guessed these parts are inserted into the wafer, but if you have to fuse them off... aren't they wasting those little parts each time they do that? Why put them there in the first place? Although perhaps that's when they just design the lower performing GPU chipset and avoid the "mutilation", once it starts to not be worth the trouble.

 

[RazorWind]

The question is simple: how much easier is it to manufacture a low-end GPU than a high-end GPU? Say, how much easier is it to manufacture a 106 GPU than a 102 die for Nvidia?

Now, I get how the die savings work: smaller GPU, you can fit many more dies on the same wafer, and being smaller means a lesser error rate, higher yield. I get the theory. What I don't know, is what about the rest of the components per graphics card. So, I'm wondering, what makes it easier (other than improved yields due to smaller GPU die) to manufacture a lower end GPU that normally sells at MSRP of $200-300?

Any links to websites that explore this would be welcome. It's good information to have in general, but also as a way to extrapolate and speculate how much stock we might expect to see of RX 6700 and RTX 3060 cards.

Are you talking about low end GPU, or a low end graphics card?

The low end GPU is cheaper to manufactuer because, as you say, you get more chips out of each wafer, so you have more chances for each one to be good.

The low end graphics card is cheaper to manufacture because it has a simpler PCB design, and generally fewer components. You don't tend to see monster 14 phase VRM designs on low end cards, and for every phase, you need at least a few dollars worth of parts. So, if you only have four phases, instead of 14, that's ten phases times three or four dollars that you aren't putting on the card. You can also get away with using simpler controller ICs, fewer memory chips and so forth. This is all on top of the greater yields of the GPU itself.

 

[euskalzabe]

Are you talking about low end GPU, or a low end graphics card?

That's an excellent point on terminology: I do mean graphics card, not GPU. I get how a GPU is cheaper. I wasn't seeing how the rest of the card could get much cheaper. I guess it really is as simple as less components, less cost, period. Perhaps I was expecting some sort of convoluted rationale, but I'm willing to see it for what it is: less is less

Hopefully this wasn't a waste of time, it still made me think about interesting parts of the whole card I had not thought about. What can I say, I have plenty of time to think (and perhaps overthink) things while we're stuck at home. It's been 10 months of pandemic, y'all. Closing in on one full year. Jeez. I hope you all have access to a vaccine relatively soon - I'm expecting mine sometime in February through my job. Good luck to all!

 

[RazorWind]

That's an excellent point on terminology: I do mean graphics card, not GPU. I get how a GPU is cheaper. I wasn't seeing how the rest of the card could get much cheaper. I guess it really is as simple as less components, less cost, period. Perhaps I was expecting some sort of convoluted rationale, but I'm willing to see it for what it is: less is less

Hopefully this wasn't a waste of time, it still made me think about interesting parts of the whole card I had not thought about. What can I say, I have plenty of time to think (and perhaps overthink) things while we're stuck at home. It's been 10 months of pandemic, y'all. Closing in on one full year. Jeez. I hope you all have access to a vaccine relatively soon - I'm expecting mine sometime in February through my job. Good luck to all!

There's also the lower cost of designing it. Every VRM phase on a modern graphics card typically has a QFN56 or QFN52 power stage IC. That means you have 57 traces for every single phase, just for that one IC. All those traces have to be laid out very carefully, and they have components that they connect to. The more you have, the harder it is to fit them all on the board.

Economies of scale are also a thing. As the number of copies of a card (or any other product) you make goes up, the per-unit cost goes down, because you can spread all the fixed costs out over more units.

 

[madpistol]

Now this is interesting. In the end, the card needs to connect to electricity and VRMs, and the GPU to memory - massive oversimplification obviously. So, if the basic blocks are the same - electricity - VRM - GPU - memory, other than having fewer VRMs due to lower power requirements and less memory blocks... what else can be simplified on the board? I guess it's the electrical engineering perspective that I'm missing here. The "less is cheaper" is obvious, I'm just trying to decipher what that "less" actually is.

Less connections between VRMs and other components
Less connections between GPU die and memory
Less components controlling signaling between parts.

The PCB on the 3080 FE Is very small for such a high-end part. That's not cheap. A lot of components are crammed into a small space. Some of these components require hundreds of leads to be traced between it and another component. Each of these leads cannot interact with another component before it reaches its destination, and the smaller the area is, the more layers the PCB may need to achieve this goal.

More design time = more expense.
More PCB layers = more expensive.

 

[euskalzabe]

Every VRM phase on a modern graphics card typically has a QFN56 or QFN52 power stage IC. That means you have 57 traces for every single phase, just for that one IC. All those traces have to be laid out very carefully, and they have components that they connect to. The more you have, the harder it is to fit them all on the board.

That blows my mind. I've always thought VRMs were just little blocks - of course I figured there was more to it, but I didn't know there were that many "lanes" (traces) to connect one by one.

The PCB on the 3080 FE Is very small for such a high-end part. That's not cheap. A lot of components are crammed into a small space. Some of these components require hundreds of leads to be traced between it and another component. Each of these leads cannot interact with another component before it reaches its destination, and the smaller the area is, the more layers the PCB may need to achieve this goal.

I figured the compactness of the most recent cards had to have an impact. Especially since you have to cram all those traces, and of course I'm sure you need to isolate them so there's no short-circuiting possible - or whatever the electrons "contaminating" each trace's signal is called, I think "cross talk" or something like that.

Again, for context - 17th century literature professor here, so I know very little about the engineering of this, but find it fascinating. It's why for years and years I've read Anandtech's fantastic architectural deep dives - I used to understand nothing (ROPs? TMUs?) and little by little researched every component until I can at this point read and understand most of it on my own. I do love learning new things, and this stuff is so complex at this point, it's like observing a new universe. Wasn't it Jobs that said that when a computer design is complicated enough, it becomes indistinguishable from magic? Well I'm here to rationalize, demystify and understand what makes the magic tic - or, in this case, raster, I guess. Or soon, ray-trace or path-trace

 

[Sycraft]

The biggest difference, cost wise, is the die size for sure. It isn't just error rate, it is because cost is per wafer and smaller chip = more chips per wafer. So suppose a 7nm wafer costs $50,000 (this is not a real number, just made up). If you have some absolute chonk of a chip and can only pack 4 on there, and on average 1 of those 4 is DOA then your manufacturing cost is $16,667/chip. However if you have a tiny chip and can pack 1000 on there, even if 10 of them are DOA on average you only a little over $50/chip.

It also means you can make them faster as you only get so many wafers/day from a fab. So the more you can get on a wafer, the more chips you get per day.

The other thing that can constrain supply currently is memory. Higher end cards use higher end memory and that tends to be in shorter supply, as well as more expensive.

As others have mentioned the rest of it costs more too, but in terms of supply constraints the chips are the big one. Thing is, that doesn't mean necessarily you'll see more lower end cards since they aren't independent of each other. Samsung gives nVidia a certain number of wafers per day, and any wafer used for one kind of chip is one they can't use for another. So while they can get more 3060s per wafer, they may not want to use that many wafers on them because they also want to sell the higher end cards.