Nvidia's 3000 Series RTX GPU [3090s with different memory capacity]

[DegustatoR]

What if one GPU is using high performance libraries and another high density? What if the HP library GPUs are made on a high contested (and therefore expensive) node, while the high density design is on a less performant (and thus cheaper per transistor) node? I think "performance per transistor" goes out the door as a relevant metric when one GPU designer is laying out 40-50 MT/mm^2 while the other is getting 65 MT/mm^2 on the same node.

Unless you're aiming at comparing GA102 with a GPU inside some midrange Qualcomm SoC for some reason all these "what ifs" are as relevant as saying that perf/transistor doesn't matter cause your GPU at 28nm can have higher performance than another GPU at 5nm.

We are talking about competing products aimed at the same market and being sold at the same time, not some made up situation.

 

[Deleted member 13524]

Performance per die cost is the obvious metric, with the minor problem that the few people who know the cost for sure aren't in a position to tell anybody.

I think performance-per-die-cost is the best metric as long as you're comparing GPUs with similar memory technologies and arrangements.
For example, a GPU that uses HBM2 instead of GDDR6 will provide higher bandwidth for the same die-area dedicated to PHYs. However, what is saved in die-area (and therefore die-cost) then translates into significantly higher memory and packaging costs.

Unless you're aiming at comparing GA102 with a GPU inside some midrange Qualcomm SoC for some reason all these "what ifs" are as relevant as saying that perf/transistor doesn't matter cause your GPU at 28nm can have higher performance than another GPU at 5nm.

What both me and @Qesa have been trying to explain to you is that the same node, and even the same chip can use both performance-optimized transistors and density-optimized transistors.
From this information, if you can't conclude that your perf/transistor metric is useless then there's nothing else to reason with.

 

[DegustatoR]

What both me and @Qesa have been trying to explain to you is that the same node, and even the same chip can use both performance-optimized transistors and density-optimized transistors.
From this information, if you can't conclude that your perf/transistor metric is useless then there's nothing else to reason with.

What you've been trying to do is add a little FUD into a simple metric which, among other things, determine the cost of competing GPUs. The fact that some GPU can have transistors made with different libraries means nothing for the end metric of perf/transistor. A smaller chip with higher performance will either win in perf/dollar or in margins. It's simple and yeah, nothing to reason with.

 

[trinibwoy]

What if one GPU is using high performance libraries and another high density? What if the HP library GPUs are made on a high contested (and therefore expensive) node, while the high density design is on a less performant (and thus cheaper per transistor) node?

What if one company has more highly paid janitors? Of course the devil is in the details but there’s no way for us to know all the details so there isn’t any point in starting down that rabbit hole.

 

[Deleted member 13524]

The fact that some GPU can have transistors made with different libraries means nothing for the end metric of perf/transistor.

Is this trolling or you can't really understand how wrong this statement is?

 

[DegustatoR]

Is this trolling or you can't really understand how wrong this statement is?

I've said all I could on the matter. Feel free to add more FUD about how wrong all this is when there are competing products made with 28nm and 5nm on the market at the same time.

 

[Qesa]

Unless you're aiming at comparing GA102 with a GPU inside some midrange Qualcomm SoC for some reason all these "what ifs" are as relevant as saying that perf/transistor doesn't matter cause your GPU at 28nm can have higher performance than another GPU at 5nm.

I *thought* it was clear that I was referring to RDNA and Ampere. Tottentranz's 28nm and 5nm example is obvious exaggeration, however when nvidia's managing similar transistor density to AMD while on a node with a much lower wafer cost, using transistor count as a proxy for die cost doesn't seem accurate. If GA102 had the transistor density of GA100 (and x100 and x102 have been similar in the past when on the same node) it'd be considerably smaller than Navi 21 despite the higher transistor count. A direct result of RTG opting to use high performance libraries and/or liberal use of decap cells to chase high clock speeds while nvidia's gone for density.

What if one company has more highly paid janitors? Of course the devil is in the details but there’s no way for us to know all the details so there isn’t any point in starting down that rabbit hole.

Sure, but I'm arguing against putting faith in a metric that is lacking those details, rather than trying to figure all of them out.

 

[DegustatoR]

I *thought* it was clear that I was referring to RDNA and Ampere. Tottentranz's 28nm and 5nm example is obvious exaggeration, however when nvidia's managing similar transistor density to AMD while on a node with a much lower wafer cost, using transistor count as a proxy for die cost doesn't seem accurate. If GA102 had the transistor density of GA100 (and x100 and x102 have been similar in the past when on the same node) it'd be considerably smaller than Navi 21 despite the higher transistor count.

That's down to more expensive janitors. You can't guarantee that even two chips made on the same process cost the same since that's up to specific contracts of these chips owners with the factory which make them. As I've said, the metric is simple and sound at its base. The rest is details.

A direct result of RTG opting to use high performance libraries and/or liberal use of decap cells to chase high clock speeds while nvidia's gone for density.

So? You get two chips with determined die sizes anyway, no matter what you've used to achieved these. You don't sell density or high clocks, you sell performance and the one which gets higher performance from lower number of transistors wins. The comparison is more complex right now of course since its been about 5 years since AMD and NV used the same process but this is also a result of both of them trying to find a way to make that performance cost lower in production, so essentially the same idea.

 

[Putas]

Perhaps we have two different perspectives here. I would choose a card according to ToTTenTranz's metric and I would judge chip design profficiency according to Degustator's metric.

 

[pTmdfx]

Seeing an assertion of "the one which gets higher performance from lower number of transistors wins", I feel obligated to link a recent article from David Kanter on this exact topic: Transistor Count: A Flawed Metric.

 

[DegustatoR]

Seeing an assertion of "the one which gets higher performance from lower number of transistors wins", I feel obligated to link a recent article from David Kanter: Transistor Count: A Flawed Metric.

Why?

The transistor density is intimately related to the overall objectives and design style. Comparing substantially different designs such as a fixed-performance ASIC (e.g., Broadcom’s Tomahawk 4 25.6Tb/s switch chip or Cisco’s Silicon One 10.8Tb/s router chip) and a high-performance datacenter processor (e.g., Intel Cascade Lake or Google’s TPU3) is misleading at best.

Guess it's good then that nobody here do that. Haven't read further.

 

[Deleted member 13524]

Imagine being so erroneously proud and self-righteous on your armchair expertise to the point of thinking you know better than David Kanter.

 

[Malo]

I think the point DegustatoR is making is that we're not comparing say an Ampere GPU to a Ryzen CPU.

 

[Deleted member 13524]

Seeing an assertion of "the one which gets higher performance from lower number of transistors wins", I feel obligated to link a recent article from David Kanter on this exact topic: Transistor Count: A Flawed Metric.

https://www.realworldtech.com/transistor-count-flawed-metric/2/
This table is pretty cool:

I wonder why Mariko has such a low density for a 12nm SoC. It seems to be almost on the level of a 28nm GPU (Fiji is 14.9 MTrans/mm^2), and SoCs do seem to carry a significantly higher density than GPUs (probably due to larger proportions of SRAM).

Perhaps it's pad-limited and smartphone/tablet SoCs with 64bit LPDDR can't really go much below 100mm^2 either way? OTOH, Apple has a couple of SoCs at below 85mm^2..

EDIT: Duh, that's not Mariko and it has nothing to do with the Nintendo Switch.

 

[DegustatoR]

Imagine having your head so far up your ass being so erroneously proud and self-righteous on your armchair expertise to the point of thinking you know better than David Kanter.

Imagine not being able to even understand what people are talking about. That's you.

 

[Deleted member 13524]

Imagine not being able to even understand what people are talking about. That's you.

Did you change your opinion about "GPUs with higher performance/transistor always winning" or is it still stuck inside?

 

[DegustatoR]

Did you change your opinion about "GPUs with higher performance/transistor always winning" or is it still stuck inside?

Have the laws of physics changed between then and now? As I've said already, there's nothing I can add to what I've said. You're just incapable of understanding.

 

[Deleted member 13524]

Still stuck, then.

 

[CarstenS]

This thread does not live up to what B3D should be like. IMHO. Come on guys, you are better than this.

 

[Frenetic Pony]

Since people were asking for 8gb+ vram usage... well here it is. Call of Duty Black Ops... whatever number (Cold War) using over 8gb even without raytracing:

See, ok, that's done. There's been other examples all around this forum of course. So the 3060ti and 3070 aren't perfectly ideal GPUs, nor perhaps is the 3080 as that's getting pretty close to above 10gb already.

But then, we don't live in an ideal world. After all its not like you can buy any of these new GPUs anyway, or a console for that matter. Heck maybe by the time you can, if you want to go to the extra expense and such cards as the 3080ti and 3070ti even exist, you can buy one of those. My only hope is to give the best buying advice I can to whoever comes around here being curious. Hell maybe AMD will suddenly put out better GPUs as well, and Intel's new CEO will somehow make their GPUs worthwhile. We'll wait and see.