Predictions on when GPUs will be back at RRP.

[Saracen999]

Coming soon, 4070!

Very probably.

At what price point, though?

Seeing as it's a 4070 not a 4080 'variant', it'll be a lot cheaper, right? Yeah, quite so.

We'll see.

[kalniel]

I presume there will be a bit of reduction, as I expect that's why they wanted to call it a 4080 in the first place, to go with the higher price. But it very much depends on what the competition do and indeed, shoppers - if the demand isn't there (and everything points to massively reduced demand) then doesn't matter if there's competition or not - people are voting with their wallets.

[CAT-THE-FIFTH]

I didn't even notice it uses the 104 chip, which as you say, is normally the x060 card. Sheesh, the greed has well and truly got to them.

Coming soon, 4070!

But also as a 104 chip it is tiny - the last time we had such a small chip was the GK104(GTX680) and the 104 series chips before that used to be more midrange(for example the GTX460 and GTX560 used a 104 series chip). This is further proven by the 192 bit memory bus.

In terms of die size and memory bus it is closer to the GA106(272MM2 vs 295MM2),so you could call it an RTX4060/RTX4060TI at best. So even an RTX4070 would be generous and still would mean a big price rise for an RTX3060 successor,because we all know Nvidia will make the RTX4070 more expensive than the RTX3070 or RTX2070.

Well, they could argue that the card name is more about comparative performance than chipset and that 4080 12GB is a reaso ..... aw, nuts, even I can't sell that one. I mean, I don't mind the occassional light troll with a dollop of devil's advocate but there's the ridiculous and then there's .... that.

Not that card name reflecting performance is inherently a bad idea. As with the CPU clock speed discussion elsewhere, the only metric that actually much counts is real-world performance, but given that different people have different use cases, a metric that's already extremely hard to define gets eeven harder. e.g. how much difference does extra memory make? It depends what games you play, or will play, and/or what non-gaming uses you have.

I guess "unlaunching" the 12GB card is, overall, a Good Thing. It did, after all, at a minimum appear to be about being deceitful. nVidia don't get many brownie points from me for it, though - it's kinda like catching a burglar in the act and him putting your stuff back, then trying to sell the notion that all's well. Oh look, we tried to rip people off, got called on it so we're going to pretend it never happened? It's in the mens rea, nVidia.

The problem is that relative to the high end,increasingly over the last decade,the midrange/mainstream dGPUs have been progressively getting worse and worse. This is because repeatedly Nvidia(and even AMD) are just pushing up the price of smaller die dGPUs upwards,and this has been matched by a big increase in margins. So any cost increases due to the node and RAM,have been massively outpaced by all these shenanigans.

You can see that the successors to many midrange/mainstream dGPUs basically offer exactly the same price/performance than the dGPUs they replaced,ie,any performance increase has been matched by the increase in price. It also means to actually get an improvement in performance you need to upgrade to a dGPU up the range,which often is the proper successor.

In terms of the RTX4080 12GB at $900,if you look at the die size and memory bus,it is closer to an RTX4060. In fact Nvidia has only used 192 bit memory buses on mainstream/midrange 60 series dGPUs. The RTX3060,RTX2060,GTX1060,GTX660TI all had 192 bit memory buses.

Its the same renaming nonsense you see with other electronics like smartphones and its a way of quietly hiding price increases. Like all the nonsense "Pro" or "Ultra" additions on the name. So you have one model replaced by two or three.

This is why people comment the replacement for their phone doesn't offer a big improvement,because in many cases more models are introduced,and its often one of the higher priced models which is the actual replacement. Nvidia started all thise nonsense with dGPUs,back in the Kepler days.

I have been telling people for years this has been happening - its a form of tech shrinkflation.

[DanceswithUnix]

The RTX3060,RTX2060,GTX1060,GTX660TI all had 192 bit memory buses.

Careful with that, 4000 series have more cache like AMD's infinity cache. Now for a given die size they have about 3 times the transistors to play with, so they need that cache to avoid HBM being mandatory, but it does make direct comparisons pretty meaningless. If Nvidia got it right, then they will fit the width of memory needed to feed the shaders; no more (as that would waste money) and no less (as that would hurt performance).

In previous gens, more shaders were added and memory became faster so they sort of scaled at the same rate.

[CAT-THE-FIFTH]

Careful with that, 4000 series have more cache like AMD's infinity cache. Now for a given die size they have about 3 times the transistors to play with, so they need that cache to avoid HBM being mandatory, but it does make direct comparisons pretty meaningless. If Nvidia got it right, then they will fit the width of memory needed to feed the shaders; no more (as that would waste money) and no less (as that would hurt performance).

In previous gens, more shaders were added and memory became faster so they sort of scaled at the same rate.

But the difference AMD went with a 256 bit bus at the top-end. Also you have to remember,there is an AD103 too which is similar area to the GA104. The 103 is the new 104.

The RTX4090 has a 384 bit memory bus like all the previous high end Nvidia dGPUs since the GK110. So a 192 bit memory bus and a sub 300MM2 die(the smallest 104 series die since the GK104) is the definition of midrange in the Ada Lovelace line-up. It's just under half the size of the AD102. By every metric it's an RTX3060 successor and midrange dGPU.

Edit!!

4NM/5NM


The AD104 is 48.5% the area of the AD102(192 bit bus). AD103(RTX4080 16GB and 256 bit bus) is 62.3% of the area of the AD102.


8NM:


The GA106 was 44% the area of the GA102(192 bit bus). The GA104(256 bit bus) was 62.4% of the area of the GA102.


14NM/12NM


The TU106 was 59% of the area of the TU102(192 bit bus). The TU104(256 bit bus) was 72.3% of the area of the TU102.


The GP106 was 42.45% the area of the GP102(192 bit bus). The GP104(256 bit bus) was 66.67% of the area of GP102.


28NM:


The GM206(128 bit bus) was 38% the area of the GM200. The GM204(256 bit bus) was 66% of the area of the GM200.


The GK106 was 39% of the area if the GK110(192 bit bus). The GK104(256 bit bus) was 52.4% the area of the GK110.


45NM:


The GF106(128 bit bus) was 45.4% the area of the GF100/GF110. The GF104(256 bit bus) was 63.8% the area of the GF100/GF110.

The AD104 is the smallest 104 series dGPU relative to the 102 series dGPU in the last 12 years. The 104 series has now shifted to the third position where the 106 used to be.

All the 106 series/third in line-up dGPUs had less than 256 bit memory buses. The second line-up dies always had at least a 256 bit memory bus.

By modern standards,its relative size is closer to a 106 series dGPU. The 104 series tends to 2/3 the area of the top chip.

The "AD104" is basically an "AD106" rebranded one level up because there is an AD103.

When it gets released,lets see performance relative to the RTX4090. I expect it will be more like a 60 series dGPU.

[kalniel]

The 2080 was a 104 chip don't forget, so not totally out there that the 4080 was as well and the 2070 was a 106. Both sides have been gradually reducing mem width as well as VRAM gets faster and caches bigger.

[CAT-THE-FIFTH]

The 2080 was a 104 chip don't forget, so not totally out there that the 4080 was as well and the 2070 was a 106. Both sides have been gradually reducing mem width as well as VRAM gets faster and caches bigger.

They introduced a 103 series this time. So the 104 series is now the exact same place the 106 series dGPU would have been. Just look at relative size to the top die,the AD104 fits entirely the same sort of relative size the 106 series used to be. I also expect performance relative to the top die,will be more like a 106 series based dGPU.

Nvidia are masters of rebranding stuff. People are wondering why the midrange/mainstream dGPUs seem to stagnate relative to the top end- that is the reason why. The RTX4060 most likely will be a fourth die,ie,the 106 will be where the 107 used to be.

If you go back further to Fermi and so on,the GTX460/GTX560 dies used a 104 series die,not a 106 series die and it was nearly 2/3 the same of the top die. The GF116 was 45% the size of the top die.

If you look above,consistently the "second die" has been around 2/3 the size of the large die. The one below it around the 45% mark.

The AD104 fits the classic size of the third die,because it is the third die,not the second. Nvidia is just changing names now.

They are selling smaller and smaller dies for more and more money. This is how their margins have been going up so much during the last 12 years!

It is so important AMD can compete - when AMD competed,it forced Nvidia to base the RTX3080 on its best die,with a stronger memory bus. I just hope Nvidia doing this,does not mean RDNA3 isn't as fast we think! They tend to do this when AMD can't compete!

[Iota]

You would think they would do the opposite, make the 4060 so fast that Intel's stuff only manages 4050 performance and gets labelled as low end.

I guess they really aren't bothered by Intel.

I mean really speaking they could have gone RTX 4060 (12GB) / RTX 4070 (16GB) / RTX 4080 Ti (24GB) with the cards they're releasing. They would still have room in the product stack based on cores alone for 4060 Ti / 4070 Ti / 4080 / 4090 and/or Super variants and/or x660 variants if they really wanted to fill out the product stack completely. Instead it look like they opted for the increased margins on what the cards should really have been by just naming them higher up the food chain. Sure they've got increased wafer costs, but there was still room in there for improved yields and more product releases while also kicking Intel and AMD squarely with performance. Enough so that any response would have to involve Intel and AMD going for the low to mid range. Maybe AMD will surprise us with RDNA3, I suspect that it'll leave AMD scope to also increase margins as well though.

The problem is that relative to the high end,increasingly over the last decade,the midrange/mainstream dGPUs have been progressively getting worse and worse. This is because repeatedly Nvidia(and even AMD) are just pushing up the price of smaller die dGPUs upwards,and this has been matched by a big increase in margins. So any cost increases due to the node and RAM,have been massively outpaced by all these shenanigans.

Ignoring the die sizes which are going to get smaller going from 12nm > 8nm > 4nm even with increases in transistor counts, this seems a repeat of the 2xxx side of things (I'm sure they've also done this earlier as well), where they increase pricing compared to performance because of the stock still available for prior generation cards, mainly to keep AIB cards selling so they don't end up with having to keep the AIB partners happy in some way. 3xxx was an anomaly because of the addition of scalpers deciding to include dGPUs in their portfolio (pandemic + mining as well) to keep cards scarce and push up pricing. Really all Nvidia has done is switch miners for scalpers to buy their products. It wouldn't surprise me if they directly supplied scalpers with 4090 cards just so they could say "they sell really well, so we've got the pricing right and the demand is there" to shareholders.

I'm getting cynical in my old age.

The 2080 was a 104 chip don't forget, so not totally out there that the 4080 was as well and the 2070 was a 106. Both sides have been gradually reducing mem width as well as VRAM gets faster and caches bigger.

280 GT200 512 bit bus.
480 GF100 384 bit bus.
580 GF110 384 bit bus.
680 GK104 256 bit bus.
780 GK110 384 bit bus.
980 GM204 256 bit bus.
1080 GP104 256 bit bus.
2080 TU104 256 bit bus.
3080 GA102 320 bit bus.

4080 (12GB) AD104 192 bit bus.
4080 (16GB) AD103 256 bit bus.

Pick the odd one out.

It's going back quite a bit and Nvidia naming of the chip used aside, you are correct that the bus used for memory has gradually been getting smaller as everything has progressed. The 3080 10GB was a little bit of an anomaly where they jumped upwards to a 320 bit bus (384 bit bus for the 3080 12GB and 12GB Ti variants, 320 for the 3080 Ti 20GB variant).

The 16GB 4080 variant seems to be the correct bit bus for this generation if I'm correct in considering the memory is the same GDDR6X used, I would suspect the 192 bit bus would restrict the memory bandwidth available on the AD104 chip a little too much?

[CAT-THE-FIFTH]

Ignoring the die sizes which are going to get smaller going from 12nm > 8nm > 4nm even with increases in transistor counts, this seems a repeat of the 2xxx side of things (I'm sure they've also done this earlier as well), where they increase pricing compared to performance because of the stock still available for prior generation cards, mainly to keep AIB cards selling so they don't end up with having to keep the AIB partners happy in some way. 3xxx was an anomaly because of the addition of scalpers deciding to include dGPUs in their portfolio (pandemic + mining as well) to keep cards scarce and push up pricing. Really all Nvidia has done is switch miners for scalpers to buy their products. It wouldn't surprise me if they directly supplied scalpers with 4090 cards just so they could say "they sell really well, so we've got the pricing right and the demand is there" to shareholders.

For over a decade its been:

102/100>104>106>107

Now its:

102/100>103>104>106

The top end dies have remained quite consistently between 550MM2 to 650MM2 in area.

Now look at the area relative to the top die,the AD104 fits into the same percentage range as the previous 106 series dies,ie, 40% to 50% of the top die. The 104 series were typically 2/3 the size of the top die,which surprise,surprise the AD103 is.

In terms of transistors the GA106 had 42.4% of the transistors of the GA102. GA104 had 61.5% of the transistors of the GA102.The AD104 has 46.96% of the transistors of the AD102. The AD103 has 60.2% of the transistors of the AD102.

The GP104(GTX1080) had 61.02% of the transistors of the GP102(GTX1080TI).

So even looking at transistors,the AD104(RTX4080 12GB) does not look like enthusiast level die - it looks more like a mainstream die.

If you look at shaders,the AD104 only has 41.67% of the shaders of the AD102.

Plus people are getting bamboozled by the jump - remember there is a much better node too.

But what you are getting in the AD104 is less of a GPU relative to the AD102,than the GA104 was relative to the GA102. The mainstream dGPUs are going to be powered by what would be considered a 107 series sized AD106.

It's almost like they are trying to make the mainstream dGPUs relatively less performant so they can see more high end sales.

Sure,Nvidia needs to ditch its Ampere stock,but why bother rebranding a GA106 successor as a AD104? Because they intend to push the real product stack up again. They did it sucessfully with Kepler. A 104 series dGPU used to power a 60 series card. But a relatively less performant 106 series die was used for years.

It is only because RDNA1 and RDNA2 performed well,than it forced Nvidia to use a larger 104 series die for the RTX2060 Super and RTX3060TI. We actually had cards with more than 6GB of VRAM and larger memory buses.

My concern is RDNA3 might not be performing as well as expected so Nvidia feels it can do this!

[kalniel]

280 GT200 512 bit bus.
480 GF100 384 bit bus.
580 GF110 384 bit bus.
680 GK104 256 bit bus.
780 GK110 384 bit bus.
980 GM204 256 bit bus.
1080 GP104 256 bit bus.
2080 TU104 256 bit bus.
3080 GA102 320 bit bus.
4080 (12GB) AD104 192 bit bus.
4080 (16GB) AD103 256 bit bus.

Pick the odd one out.

It's the 3080 isn't it You can see the general trend towards 104 class chips (though it's just a name, as Cat is highlighting) and lower memory buses, then along comes some serious competition from AMD (+Samsung fab issues) and Nvidia had to pull out a much higher chip class and wider mem bus.

My concern is RDNA3 might not be performing as well as expected so Nvidia feels it can do this!

I'm a bit more optimistic - I think Nvidia just assumed people will take DLSS3 frame rate (including interpolated frames) as the performance level and so thought/hoped they could pretend a massive leap generationally (as opposed to a simply large one) and class all their cards up a tier - which fits the problem of 3000 inventory too.

The backlash over the 4080 12GB and fairly well understood analysis of DLSS3 means people aren't going to be as taken in as expected, so RDNA3 probably stands a very good chance when comparing like for like.

On frame interpolation: I would not be at all surprised if AMD come up with a non-propriatory way of doing this - it's not rocket science, it's far easier than upscaling resolution (though adjusting for artefacts is still tricky), and AMD traditionally have an excess of shader performance which they could perhaps use to good effect.

[Iota]

My concern is RDNA3 might not be performing as well as expected so Nvidia feels it can do this!

I'm fairly optimistic about RDNA3, especially when you add in the refinements over time with drivers from AMD improving performance.

I'm a bit more optimistic - I think Nvidia just assumed people will take DLSS3 frame rate (including interpolated frames) as the performance level and so thought/hoped they could pretend a massive leap generationally (as opposed to a simply large one) and class all their cards up a tier - which fits the problem of 3000 inventory too.

I honestly wouldn't want to take the hit with latency in games using interpolated frames. Sure if you're completely tied in with Nvidia reflex, it probably isn't quite as bad, but if you're not then you would probably notice it. I'd rather have decent frame rates and low latency, case in point I actually limit my 3090 to 120 frames per second. Two reasons, first is power draw and the second is I don't notice any improvements over 120 (I definitely notice it dropping to 60 though).

[Percy1983]

Interesting time ahead, AMD 7900xt and xtx pricing has to effect everything, they are both outside my acceptable price.

But at those prices the whole 6950xt downwards stack can't remain at the same prices, to which used must get cheaper.

As much as the 6700xt as just about getting into my £320 budget I am going to wait a little longer as a 6800 or 6800xt may be the new target.

Generally it seems mining is getting ever bigger on losses in the UK with increasing energy prices but it seems many are holding cards mining hopium but with cards values dropping with new and more efficient cards (for mining or gaming) making used mining cards worth even less to which a race to the bottom could start if they stop mining hopium and start mining reality.

[Output]

... if they stop mining hopium and start mining reality.

It took me a few moments to get that with the first mention of 'hopium'.

At first I was thinking "Ugh, they've still somehow found new crap to mine?"

[Saracen999]

...

Generally it seems mining is getting ever bigger on losses in the UK with increasing energy prices ...

My heart bleeds for the poor things.
/snigger

[Iota]

My heart bleeds for the poor things.
/snigger

Mine doesn't, they've likely already made their money, meanwhile pushing up card pricing for everyone else who actually uses them for something more "useful". Like gaming.... hmmm.

[Saracen999]

Mine doesn't, they've likely already made their money, meanwhile pushing up card pricing for everyone else who actually uses them for something more "useful". Like gaming.... hmmm.

Yeah, that was kinda my point. With the caveat that I'm NOT talking about individuals mining part-time on their gaming/prductivity systems, but ONLY about miners/scalpers that so screwed the rest of us by bulk buying and driving prices into the stratosphere (or above) .... may the <bleeps> rot in hell for eternity.