NVIDIA ‘Ampere’ 8nm Graphics Cards

[Twinz]

In fact, the successor to the 1080 Ti is clearly the 2080 super, not the Ti, and the successor to the 1080 is the 2070 Super, and the successor to the 1070 is the 2060 super. You can see what I have done in my chart annotations below.

The 2080S is still out of the pattern. It's also good to graphicly establish how bad the original Turing offerings were.

The 2080Ti's *performance* was actually in line with historical progress. It's the price that made it "not fit".

Just graph the the Ti's performance over the generations and bring the 2080Ti's price down to $750. I think it would fit right into the pattern that way.

 

[danlightbulb]

@Twinz yes I agree with you. Certainly the major problem was the bed they'd made with the 2080 Ti on the 1st Turing release. However I think it is fair to say that the rest of the 2nd release turing range is actually a fair step up at an equivalent price to previous generation cards.

Certainly after compiling these progressions, I am disregarding 1st release Turing and the 2080 Ti cards, and I now see the super range as reasonable value, consistent with history.

It sets my expectations well for the 3000 series, because if we see the right pattern coming up next (around +3500 on passmark scores and +/- £50 to £100 on pricing across the upgrade path) then that would be a release consistent with the past progressions. This also indicates to me that the 3080 Ti/3090 whatever its called, is not part of this sequencing any more.

 

[Joxeon]

Firstly, you may not agree but I have eliminated the 1st release of Turing completely. It just doesn't fit with the historical pattern and if you consider the super range as what Turing should have been, things make much more sense.

It still doesn't look great for Turing super especially when you factor in a gap of over 3 years between them and the gtx 10 series, I don't think nvidia would have even released the supers had it not been for the 5700/XT.

 

[gpuerrilla]

The problem with that is we haven't really seen anything really super impressive regarding RT on the consoles. If the consoles don't match like 70%+ the RT quality and effects of Ampere, Nvidia marketing will make the gap seem even bigger and they get to keep their prices unless RDNA 2 on the desktop can make a dent. I could be wrong but I doubt Nvidia has any warmth for anything other than cold hard cash(like most companies).

Yeah I agree. I am not being over-optimistic as we have been burned in the past by AMD shortcomings. nvidia will be the power player here as they were first to market.

 

[Devrij]

Awesome charts @danlightbulb, really helpful to get a feel for the change over time. Fully agree @Twinz that the 2080ti just seems obviously overpriced in the light of these charts. It almost would have made more sense if they'd called it the 2090 and then just released the 2080s as the 2080ti, 2070s as 2070ti etc.

 

[gpuerrilla]

Yeah its nearly as clear cut as that @Devrij only like @Joxeon points out - historical releases may not have the 3 year gap so you cannot roll over so easily.

 

[LeMson]

It still doesn't look great for Turing super especially when you factor in a gap of over 3 years between them and the gtx 10 series, I don't think nvidia would have even released the supers had it not been for the 5700/XT.

No bugger bought the non supers, the owners threads on here were barren

 

[Dirk Diggler]

you may not agree but I have eliminated

A chart with data removed to make it look better, that's helpful.

 

[Fyter]

Right Ive made some more modifications which to me, make sense as to how to represent the progression.

Firstly, you may not agree but I have eliminated the 1st release of Turing completely. It just doesn't fit with the historical pattern and if you consider the super range as what Turing should have been, things make much more sense.

Secondly I have eliminated the 2080 Ti. Again, it just does not fit and the progression from Pascal to Turing super range makes much more sense. In fact, the successor to the 1080 Ti is clearly the 2080 super, not the Ti, and the successor to the 1080 is the 2070 Super, and the successor to the 1070 is the 2060 super. You can see what I have done in my chart annotations below.



My final comment is that whilst the percentage increases are reducing from generation to generation, the increments in raw scores are actually pretty similar. As anyone who works with percentages knows, you do have to be careful interpreting them because as the numbers get larger, the percentages look smaller for a given increase. This is exactly what we're seeing here.

Also Turing super release is cheaper, along the upgrade path I have indicated, than Pascal was.




Yes I agree with you, the increment between 1080 Ti and 2080 Super was poor as you can see on my latest chart above. We would have expected it to be further to the right. However then it would start interacting with the dud card 2080 Ti which is a major outlier in this sequencing. I think that is the crux of the matter. 1st Turing was clearly awful.

Nice graphs mate! Well done. Would love to see updated version once ampere hits. It will help me decide on whether to get 3070/80/90. Gaming @ 1440P 165hz here.

 

[LeMson]

Firstly, you may not agree but I have eliminated the 1st release of Turing completely. It just doesn't fit with the historical pattern and if you consider the super range as what Turing should have been, things make much more sense.

Should include them for full disclosure though, might highlight how crappy the prices of next cards are

 

[CAT-THE-FIFTH]

It also shows how the mainstream has gotten screwed with Turing,Super or no Super. For one thing you haven't include the RTX2060 and the RTX2070 in the chart too,which are current models,and fill the gaps between the Super models.

The RTX2060 series is considered a mainstream GPU by Nvidia. Now look at how the mainstream has started to shift up higher and higher and higher. The upper levels of the mainstream are now priced at what used to be enthusiast levels in the product stack. Also the problem is by removing the RTX2080TI model,it is also not showing another problem. The whole range is being shifted upwards one model at a time. This is what the RTX2080TI did,ie,now you get an 80 series variant GPU for the price of a 80TI one,and a 70 series variant for the price of an 80 series one,etc. Now a 60 series variant replaces the 70 series ones at the price point.

Why do you think there is an RTX3090?? Instead of the 80TI being the second fastest model,now its the 3rd fastest model. So the 80TI now will replace the 80/80 Super in the line-up,and the old 90,where the 80TI was.

 

[five8five]

It still doesn't look great for Turing super especially when you factor in a gap of over 3 years between them and the gtx 10 series, I don't think nvidia would have even released the supers had it not been for the 5700/XT.

I don't think it was just the AMD cards that prompted them to release the Supers. I feel it was also the amount of greif they got for the price / performance of the original Turing cards.

 

[danlightbulb]

Should include them for full disclosure though, might highlight how crappy the prices of next cards are

They're on the previous iterations if you scroll back a bit. I wanted to put the additional information on showing the clear progression paths and the 'odd generation' that is 1st turing release was getting in the way.

It also shows how the mainstream has gotten screwed with Turing,Super or no Super. For one thing you haven't include the RTX2060 and the RTX2070 in the chart too,which are current models,and fill the gaps between the Super models.

You'll have to look back through the last few pages to see the iterations and the reasoning on why the 1st turing release was shown then eliminated.

Also the problem is by removing the RTX2080TI model,it is also not showing another problem. The whole range is being shifted upwards one model at a time. This is what the RTX2080TI did,ie,now you get an 80 series variant GPU for the price of a 80TI one,and a 70 series variant for the price of an 80 series one,etc. Now a 60 series variant replaces the 70 series ones at the price point.

I don't think this matters. What matters is that at a given price point there is progression, and that's what we're seeing between Pascal and Turing super. I think the marketing trick Nvidia have pulled here is to make you think that the 2080 Ti was the successor to the 1080 Ti, when the charts actually show, if following historical progression paths, that it really wasn't.

 

[Joxeon]

I don't think this matters. What matters is that at a given price point there is progression, and that's what we're seeing between Pascal and Turing super. I think the marketing trick Nvidia have pulled here is to make you think that the 2080 Ti was the successor to the 1080 Ti, when the charts actually show, if following historical progression paths, that it really wasn't.

Only in terms of price, in terms of performance it is exactly what you would expect as a successor to the 1080ti.

Nvidia saw how the 1080ti jumped up to £900+ and was still selling out during the mining boom and thought let's run with it on the 2080ti.

 

[Kirby Wurm]

It's at the upper echelons of the chart where Nvidia are exposed, even without the original Turing release. The 2080S is awful whichever way one considers it, and the card that historically would have been in it's rough position is priced way off the chart.

It's interesting to note that about 3.5 years separated the 780 Ti and 1080 Ti releases with the latter launching at the same price with around double the performance. Roughly the same time has passed since the 1080 Ti launch and there's surely no way the RTX 3000 entry at $700 will match that increase in performance. I suppose it's up to consumers to decide whether RT and DLSS are valuable enough to offset the difference in typical performance progression.

 

[CAT-THE-FIFTH]

You'll have to look back through the last few pages to see the iterations and the reasoning on why the 1st turing release was shown then eliminated.

I don't think this matters. What matters is that at a given price point there is progression, and that's what we're seeing between Pascal and Turing super. I think the marketing trick Nvidia have pulled here is to make you think that the 2080 Ti was the successor to the 1080 Ti, when the charts actually show, if following historical progression paths, that it really wasn't.

The GTX1080TI was 75% faster than a GTX980TI,the GTX1080 was nearly 70% faster than a GTX980. The GTX780TI(500~600MM2 large die) was nearly double the performance of a GTX580(500~600MM2 large die),and the GTX680/GTX770(300MM2 small die) was double the performance of the GTX560TI(300MM2 small die).

That is with a new node. So that is what Ampere should at least give us over Turing,around 70% to 100% more performance,at each tier in my estimations.

If you move to Maxwell which was on the same node,the GTX980TI was 40% faster than a GTX780TI,and the GTX980 was around 30% faster than a GTX780,but Nvidia was sneaky with the GTX980,as it was a smaller die(the GTX780 wasn't) and realistically a GTX770/GTX680 successor. If you compare it to the GTX770,it is over 50% faster.

The RTX2080TI was the GTX1080TI successor,but priced higher. For one they have the same sized memory bus and the same 11GB of VRAM. It's also around 35% faster which was acceptable,being on the same node,but is on the low end of refresh using the same node(TSMC 12NM was basically TSMC 16NM with a large reticle limit) on a per model basis.

Saying the RTX2080 and RTX2080 Super are successors of the GTX1080TI I don't entirely agree with. Because for one,neither show more than 20% extra performance over the GTX1080TI,which is terrible by historic comparisons. More importantly they also cut down VRAM(8GB vs 11GB) and moved to smaller buses,which are similar in size to typical 80 series GPUs. The GTX680,GTX980,GTX1080,RTX2080 and RTX2080 Super had 256 bit memory controllers and multiple of 2GB VRAM quantities. The RTX2080 and RTX2080 Super make more sense as GTX1080 successors,ie,40% to 50% faster which sounds about right for a generation without a totally new node.

The whole point of companies pushing up tiers,is to increase ASP,and also to make more tiers. So you end up spreading the performance improvements over more and more models.

 

[danlightbulb]

Only in terms of price, in terms of performance it is exactly what you would expect as a successor to the 1080ti.

Yes, you said that before I think along with Twinz, and you are correct. My take on it is that it is strong visual representation showing the 1st turing release was a dud across the board, then partially rectified by the super release however it did not fix the issue of progression specifically from the 1080 Ti. The 2080 S was not a big step change in performance (although was the same price so its still better performance per $) and the 2080 Ti was a massive price outlier.

It's interesting to note that about 3.5 years separated the 780 Ti and 1080 Ti releases with the latter launching at the same price with around double the performance

I again would suggest caution when comparing performance. Going by absolute scores, each generation has typically seen 3000 to 4000 passmark score increases across the upgrade paths. But as numbers get higher, percentages will get smaller for the same absolute increases, hence can be very misleading to use them.

 

[Joxeon]

The whole point of companies pushing up tiers,is to increase ASP,and also to make more tiers. So you end up spreading the performance improvements over more and more models.

This is probably why there is now speculation of a 3090.

 

[CAT-THE-FIFTH]

This is probably why there is now speculation of a 3090.

Yes,because now the 80TI instead of being the second fastest tier,will be the 3rd fastest tier eventually.

The fact is you can just look at what Apple and Samsung have done - they just introduce multiple models,which have pushed what essentially is the true high end tier,higher and higher.

So people buying by price,tend to get worse and worse relative upgrades. Which is partly why the smartphone market is now slowing down,as people are waiting longer and longer to upgrade.

 

[danlightbulb]

The GTX1080TI was 75% faster than a GTX980TI,the GTX1080 was nearly 70% faster than a GTX980. The GTX780TI(500~600MM2 large die) was nearly double the performance of a GTX580(500~600MM2 large die),and the GTX680/GTX770(300MM2 small die) was double the performance of the GTX560TI(300MM2 small die).

Not sure where those numbers are from but going by passmark scores, the 1080Ti was 28% faster than the 980Ti. Obviously there are different benchmarks around and I picked passmark for ease of use but caution has to be taken with percentages because the absolute numbers underlying them can affect the scale of the percentage number.

Saying the RTX2080 and RTX2080 Super are successors of the GTX1080TI I don't entirely agree with. Because for one,neither show more than 20% extra performance over the GTX1080TI,which is terrible by historic comparisons. More importantly they also cut down VRAM(8GB vs 11GB) and moved to smaller buses,which are similar in size to typical 80 series GPUs. The GTX680,GTX980,GTX1080,RTX2080 and RTX2080 Super had 256 bit memory controllers and multiple of 2GB VRAM quantities. The RTX2080 and RTX2080 Super make more sense as GTX1080 successors,ie,40% to 50% faster which sounds about right for a generation without a totally new node.

Yeah I get what you're saying but if you pick a price point (+/- £50) and read across the chart, you'll see what cards you come across. So for example if you start with the 770 at $400 and read across, you get the 970 (+62% perf, $70 cheaper) then the 1070 (+38% perf, $120 more expensive) then the 2060 Super (+22%, $50 cheaper). While it looks like the performance increases are diminishing, they aren't in absolute terms - its about +3-4000 score each iteration.

When new models are slotted in which change the product stack, I think all you can then do really is compare across a price band, at least initially. I do agree that the 1080ti to 2080 Super increment is out of whack with the historical leaps, and out of whack with the leaps in the lower cards in the same generation. That is clearly a hangover from the 1st dud Turing release. Once the 2080 Ti was set, they then couldn't introduce a 2080 S which had the 'right' performance jump because it would have undermined the Ti. That's how I see it.