Intel ARC GPUs, Xe Architecture for dGPUs [2022-]

[arandomguy]

I mean, good lord, we used to expect the new $400 GPU to match the last gen's highest end GPU! Even if that's not feasible anymore, there's still a huge middle ground between this and just getting a 15% discount at a given performance level after two freaking years...

Not as big as you might think given the time frames involved, and let's face it how you manipulate the numbers.

Pascal/1xxx is largely I think assumed as the last great value leap forward no? RTX 4070 Super was already a refresh price drop and not the initial release, $600 performance than was RTX 4070.

GTX 1060 has the same performance delta over the GTX 970 as 4070 Super performance would have over the regular 4070.

$600 -> $500 is a $100/16.7% drop.

$330 -> $250 was a $80/24.2% drop.

So we're really talking about pricing it as $450 for the same delta, or $475 if we split the difference.

 

[DegustatoR]

I mean, good lord, we used to expect the new $400 GPU to match the last gen's highest end GPU!

We also used to have a completely different production situation. What's anything of this has to do with the new products? You can't judge them based on what you were used to expect 10 years ago.

 

[techuse]

Not as big as you might think given the time frames involved, and let's face it how you manipulate the numbers.

Pascal/1xxx is largely I think assumed as the last great value leap forward no? RTX 4070 Super was already a refresh price drop and not the initial release, $600 performance than was RTX 4070.

GTX 1060 has the same performance delta over the GTX 970 as 4070 Super performance would have over the regular 4070.

$600 -> $500 is a $100/16.7% drop.

$330 -> $250 was a $80/24.2% drop.

So we're really talking about pricing it as $450 for the same delta, or $475 if we split the difference.

The 4070 offers about the same relative performance as the 1060 and costs 2.4x as much. Comparing a refresh to a new model is not really fair comparison.

 

[DegustatoR]

The 4070 offers about the same relative performance as the 1060 and costs 2.4x as much.

Relative to what?

 

[techuse]

Relative to what?

Their relative position in the performance stack. Both are about half the performance of the top GPU.

 

[Erinyes]

It's devastatingly depressing how much consumers seem to have forgotten what a new generation of GPU's is supposed to bring in terms of performance-per-dollar improvements.

What you're talking about is an absolutely abysmal generational improvement in value. Crumbs.

I mean, good lord, we used to expect the new $400 GPU to match the last gen's highest end GPU! Even if that's not feasible anymore, there's still a huge middle ground between this and just getting a 15% discount at a given performance level after two freaking years...

Well, that was also the time the highest end used to be $600-700, today it's $1500+ and the relative performance has increased significantly vs mid-range. And while I agree with what you say, there are many reasons for the generational improvements/performance-per dollar slowing down (general inflationary trends aside). Cost per transistor stopped scaling/has scaled much lower since 20nm. The 14/16nm FINFET nodes brought a one time significant improvement in power/performance which made up for the increase in transistor cost but ever since then, the cost of silicon per mm2 has been going up much faster than it did till 28nm. Along with increased costs for high speed memory and PCBs, cooling, etc.

Further adding to the situation in recent years is the competitive environment and demand for HPC/AI so there is simply no incentive to ship consumer silicon for lower margins. If anything, this is going to get worse and higher GPU prices are going to be the norm. And given Intel's cashflow issues of late, especially if they are using external foundry nodes, don't expect them to sell silicon for cheap/at a loss either.

Their relative position in the performance stack. Both are about half the performance of the top GPU.

But the top GPU at the time, the 1080ti was $700. The 4090 is $1600 so the price of the top GPU costs ~2.3X as much now.

 

[trinibwoy]

Both are about half the performance of the top GPU.

This is a completely meaningless metric as “top GPU” is not a product feature.

 

[techuse]

This is a completely meaningless metric as “top GPU” is not a product feature.

The top GPU of both generations are a near equally cut down slice of the 102 silicon. I’m using the 1080 Ti rather than the Titan to keep things as equal as possible.

 

[DegustatoR]

Their relative position in the performance stack.

So you're saying that 4070 has the same performance as 1060?

Both are about half the performance of the top GPU.

Which means absolutely nothing as "top GPU" is basically anything. Intel's "top GPU" at the moment is A770 for example.

 

[techuse]

So you're saying that 4070 has the same performance as 1060?


Which means absolutely nothing as "top GPU" is basically anything. Intel's "top GPU" at the moment is A770 for example.

Nvidia isn’t Intel and has been using the same silicon/performance tiers for well over a decade now. They don’t suddenly become irrelevant because they lower the value proposition they are offering customers.

 

[DegustatoR]

Nvidia isn’t Intel

And 4070 isn't 1060. The idea that they have anything in common aside from the company which designed the chips in them is just laughable.

the same silicon/performance tiers

I have no idea what that is. Are we talking about completely random chip codenames now as if they mean anything between different lineups?

They don’t suddenly become irrelevant because they lower the value proposition they are offering customers.

They don't become irrelevant, they always were irrelevant. Which chip gets sold in which products depends solely on competitive landscape and plans a company had when designing them.

 

[Seanspeed]

Well, that was also the time the highest end used to be $600-700, today it's $1500+ and the relative performance has increased significantly vs mid-range. And while I agree with what you say, there are many reasons for the generational improvements/performance-per dollar slowing down (general inflationary trends aside). Cost per transistor stopped scaling/has scaled much lower since 20nm. The 14/16nm FINFET nodes brought a one time significant improvement in power/performance which made up for the increase in transistor cost but ever since then, the cost of silicon per mm2 has been going up much faster than it did till 28nm. Along with increased costs for high speed memory and PCBs, cooling, etc.

Further adding to the situation in recent years is the competitive environment and demand for HPC/AI so there is simply no incentive to ship consumer silicon for lower margins. If anything, this is going to get worse and higher GPU prices are going to be the norm. And given Intel's cashflow issues of late, especially if they are using external foundry nodes, don't expect them to sell silicon for cheap/at a loss either.



But the top GPU at the time, the 1080ti was $700. The 4090 is $1600 so the price of the top GPU costs ~2.3X as much now.

Margins on GPU's are higher than ever. The prices we're paying for GPU's has little reflection on what they need to sell for, and everything to do with what they think they can get away with. We saw a more than doubling of price per mm² from 2020 to 2022 for Nvidia.

And nobody is asking anybody to sell anything for a loss.

You are probably right that we're never gonna see good consumer GPU pricing ever again, I'm just saying that's a sad state and we are largely responsible for that by giving in. It does not mean that the rest of us should just get accept it, much less be happy about it. In no world is 4070S performance in late 2024 for $500 anything to celebrate. That's lame. And if that's all they can do, then I dont believe Intel will ever make any kind of impact in the market. It's the same problem that AMD faces - they cannot simply have a worse overall product for a minor discount and think anybody will care. Especially in the $400+ space, where spending another $50-100 for the clearly better overall product is proportionately more easily justified.

 

[pharma]

Margins on GPU's are higher than ever. The prices we're paying for GPU's has little reflection on what they need to sell for, and everything to do with what they think they can get away with. We saw a more than doubling of price per mm² from 2020 to 2022 for Nvidia.

Hmm ... Sounds very similar to the increase in my some groceries.

 

[Qesa]

We saw a more than doubling of price per mm² from 2020 to 2022 for Nvidia

Wow, crazy. I'm sure this has no correlation to the price per mm² of their wafers - you wouldn't possibly have left out such pertinent information.