NVIDIA Pascal Thread

[tviceman]

In your example, 780Ti vs 980Ti, there is again a huge clock speed deficit on the 780Ti.

~25-30% clock speed gains on the 980Ti and extra vram which has an effect in many games from 2015 onwards. When you work it out, each Maxwell CC is effectively x1.2 to x1.3 Kepler CC only at the same clocks. Hence the 20-30% IPC.

If you want to claim Pascal has 30-50% IPC, with its ~20% clock speed advantage, GP104 will be 50-70% faster than Titan X.

Now, compare the last node shrink and uarch change, 580 -> 680. The 680 was ~25-30% faster. But the 580 chip itself was compute heavy and so it suffered perf/mm2 and perf/w, thus the 680 being a gaming focused chip, has even a better handicap.

This time, Titan X is already beast mode for gaming with gimped compute. It's already a lean-mean chip.

There's zero chance of GP104 being 50-70% faster. That's more for GP100. Because GM200 is already a gaming focused chip, I would say this time around, the delta will be potentially less than the 680 vs 580 comparison.

GP104 full ~Titan X + 20% is a good result on a small chip with the power savings.

However, Maxwell added back in compute performance vs. GK104 (and the less Kepler chips) and demonstrated better perf/transistor improvements on the same node than DP-stripped Kepler GPU's did on a new node vs. Fermi on an old node.

So trying to gather comparisons are getting to the point of moot. I think the best measurement will be to look at transistors and trying guess if there will be any improvement in perf/transitor. If GP104 has 50% more transistors than GM204 (5.2b x 1.5 = 7.8b, very plausible), even if perf/transitor stays the same, a 20% clock speed improvement will result in a 70-75% performance improvement over GM204, which ends up being 30-35% faster than Titan X at 4k.

Moving to finfets with a new / updated architecture throws comparisons out of the window. This isn't a typical shrink. There is so much up in the air to leave video card enthusiasts elated or disappointed, depending on how it all unravels.

 

[MrTeal]

However, Maxwell added back in compute performance vs. GK104 (and the less Kepler chips) and demonstrated better perf/transistor improvements on the same node than DP-stripped Kepler GPU's did on a new node vs. Fermi on an old node.

So trying to gather comparisons are getting to the point of moot. I think the best measurement will be to look at transistors and trying guess if there will be any improvement in perf/transitor. If GP104 has 50% more transistors than GM204 (5.2b x 1.5 = 7.8b, very plausible), even if perf/transitor stays the same, a 20% clock speed improvement will result in a 70-75% performance improvement over GM204, which ends up being 30-35% faster than Titan X at 4k.

Moving to finfets with a new / updated architecture throws comparisons out of the window. This isn't a typical shrink. There is so much up in the air to leave video card enthusiasts elated or disappointed, depending on how it all unravels.

Depends what compute you're talking about. Maxwell has less DP compute than GK104. GM204 has more FP32 compute than GK104 of course, but that's because it's a larger chip with more cores that runs faster.

Generally as well, performance per transistor doesn't usually go up unless specific features are used. For example, even in the most recent TPU that includes many of the GW games that are claimed to be crippling Kepler, the 980 is 72% faster than the 770.

The 980 has 47% more transistors though, and the max boost on stock cards is 12% higher. If you normalize for clock speed, the 980 is 53% faster for its 47% increase in transistors, or about 4% faster per transistor. That's with the ratio of FP64:FP32 dropping from 1/24 in GK104 to 1/32 in GM204, and later in the cycle as games are more optimized for Maxwell. At launch the GTX 980 was only around 50% faster than the GTX 770, so it was actually only about as fast as a 770 per transistor and was actually slower than the 770 is you normalize clocks.

 

[tential]

The faster they will announce/launch Pascal for desktop the faster people will start to forget how badly Maxwell is performing in 2016 games and especially in DX-12 titles.
NV seriously needs to steer away bad public perception from Maxwell in 2016 and in to a new DX-12 architecture with Pascal that will "promise" better DX-12 performance.

I think Nvidia knows how to manage public perception just fine. I mean not even bumpgate could phase Nvidia.

Don't worry Nvidia has something for public perception.

 

[CakeMonster]

What's 'bad' about Maxwell? My 980gtx is probably one of the chips that I've had in my almost 20 years of gaming that I'm the most happy with. Low power, low noise, very real world noticeable performance improvement over my previous 780gtx, as well as some impressive memory compression/bandwidth achievement, and all this on record long lasting 28nm process that we initially thought had peaked with the large chip 780*.

 

[Arachnotronic]

What's 'bad' about Maxwell? My 980gtx is probably one of the chips that I've had in my almost 20 years of gaming that I'm the most happy with. Low power, low noise, very real world noticeable performance improvement over my previous 780gtx, as well as some impressive memory compression/bandwidth achievement, and all this on record long lasting 28nm process that we initially thought had peaked with the large chip 780*.

I think the argument that the haters make goes something along the lines of: "blah blah, not the best bang for the buck, blah blah no Async Compute, blah blah NVIDIA sells overpriced crap."

 

[Bacon1]

What's 'bad' about Maxwell? My 980gtx is probably one of the chips that I've had in my almost 20 years of gaming that I'm the most happy with. Low power, low noise, very real world noticeable performance improvement over my previous 780gtx, as well as some impressive memory compression/bandwidth achievement, and all this on record long lasting 28nm process that we initially thought had peaked with the large chip 780*.

Mmmmm that sexy memory compression. Seriously though, that isn't a feature you'd actually notice on your computer.

You are also comparing nvidia to nvidia, so of course maxwell isn't 'bad', because its still an improvement over kepler. The point people are making is that AMD's GCN was clearly superior design. And that cool / low power? Thats from cutting FP64 support

 

[Arachnotronic]

Mmmmm that sexy memory compression. Seriously though, that isn't a feature you'd actually notice on your computer.

You are also comparing nvidia to nvidia, so of course maxwell isn't 'bad', because its still an improvement over kepler. The point people are making is that AMD's GCN was clearly superior design. And that cool / low power? Thats from cutting FP64 support

If GCN is so great, why is it such a power hog? Engineering is all about trade offs, and it looks like the ones NVIDIA made allowed it to sweep the gaming notebook market as well as the most lucrative portions of the high-end desktop market.

Yeah, I guess you can say "oh NVIDIA buyers are sheep/idiots/stupid/more money than brains" but that doesn't really hold up in my book.

 

[Magic Carpet]

I think the argument that the haters make goes something along the lines of: "blah blah, not the best bang for the buck, blah blah no Async Compute, blah blah NVIDIA sells overpriced crap."

Erm... GTX 980 is sexy. Personally, I like it. From whatever angle you look at it. I should of sold my GTX 670 when it worked and gotten that instead (I was offered one for $350 but I hesitated, lmao). Now, it's too little, too late

Luckily, my backup 270 can still perform somewhat, so I can play a few games in the meantime, haha.

 

[Arachnotronic]

Erm... GTX 980 is sexy. Personally, I like it. From whatever angle you look at it. I should of sold my GTX 670 when it worked and gotten that instead (I was offered one for $350 but I hesitated, lmao). Now, it's too little, too late

Yeah, the 980 is an awesome GPU and anybody who bought one (or two!) at launch is probably really happy with it. Great performance, great power consumption, massive overclocking headroom, etc.

 

[Magic Carpet]

Yeah, the 980 is an awesome GPU and anybody who bought one (or two!) at launch is probably really happy with it. Great performance, great power consumption, massive overclocking headroom, etc.

Yeah, perf per watt is really amazing. Still quite a speedy card. But now I've got a lil greedy, I would like the same performance but with a single 6 pin power connector. I don't demand too much, do I?

Waiting till july or whenever.

 

[USER8000]

I think the argument that the haters make goes something along the lines of: "blah blah, not the best bang for the buck, blah blah no Async Compute, blah blah NVIDIA sells overpriced crap."

Who cares?? Its only a graphics card not a religion. At each generation you have had winners and losers - in DX12 Maxwell is not doing well and that is coming from a GTX960 owner and AMD is doing better.

Who knows in next year,Nvidia will be doing better in DX12 and so on.

Some of you are obviously new to the whole enthusiast thing - going back 15 years,performance has flipped-flopped between companies. Nvidia was owned when DX9 first came out by ATI,and previous to that the Geforce 3 series was beating the ATI 8500 series. When DX10 first came out,Nvidia owned ATI.

Do you really think Jen-Hsun Huang or Lisa Su actually care about misguided brand loyalty for such fickle things like cards?? They are earning decent money while people are more worried about the brand name of the card they have.

Most of the engineering talent behind AMD,Nvidia or Intel products don't have brand loyalty - they flip between multiple companies or even at times have expressed some admiration of their compatriots abilities.

Graphics card brand loyalty is for silly people IMHO or those who really are not that technical when it comes to computers and just buy what is advertised the most.

 

[Arachnotronic]

Do you really think Jen-Hsun Huang or Lisa Su actually care about misguided brand loyalty for such fickle things like cards?? They are earning decent money while people are more worried about the brand name of the card thye have. Graphics card brand loyalty is for silly people IMHO or those who really are not that technical when it comes to computers and just buy what is advertised the most.

I have a pair of 980 Tis, the fastest consumer GPUs that money can buy at the moment. I am not "worried" about the brand of my cards

 

[Qwertilot]

Actually, it isn't doing remotely awfully in DX12. Just not as well as in DX11. The architecturally balanced comparison for the 970/80 is after all the 280/X.

They are mid range on quite a few metrics of course. The base complaint(s) seem to be down to two things. Firstly that the 970/80 would traditionally have dropped a chunk in price by now.

Quite true I guess - but then in a fully traditional world, I guess AMD would have managed to ship Fury in big volume say 6 months earlier than they did, retiring the 290/X. That probably would have probably pushed the 97/80 price down in response.

Also I think some resentment about the idea that anyone might care about anything other than performance/money. That brand of hatred is more reserved for the 960 though

 

[USER8000]

I have a pair of 980 Tis, the fastest consumer GPUs that money can buy at the moment. I am not "worried" about the brand of my cards

But you obviously are though as you seem very emotional:

I think the argument that the haters make goes something along the lines of: "blah blah, not the best bang for the buck, blah blah no Async Compute, blah blah NVIDIA sells overpriced crap."

It is a company and you seem very worried what people think about the company for some reason.

Edit to Post.

For example,I am massively into photography. I have Nikon FX gear,specialised lenses,etc. Some people think it is overpriced when compared to their £250 dSLR or their phone,etc. There are people who I know who think Canon is much better than Nikon. Knew someone with decent Pentax gear who was mocked since he did not shoot Canon or Nikon but won an international prize for his work...done on a Pentax and the others they won nothing.

I only bought it since it meets my needs now once I looked at all the alternatives at the time,but I will quite happily look at alternatives if required in the future if I believe Nikon are not delivering on what I NEED. I don't try and get emotionally involved with the companies since like a graphics card,all of these things are a tool to do a job. Companies like AMD and Nvidia want to sucker you in with all the emotional crap same with many companies:

http://www.bbc.co.uk/news/business-13416598

 

[IllogicalGlory]

Yeah, the 980 is an awesome GPU and anybody who bought one (or two!) at launch is probably really happy with it. Great performance, great power consumption, massive overclocking headroom, etc.

Not me. I completely regretted buying a 980 because it was overpriced and I didn't notice any difference compared to my 670 SLI, and now that I own a 290, I noticed no difference compared to the 980 either, even overclocked. Huge waste of $550 (or $80 net). Nothing about the 980 wowed or impressed me as an end user. Neither did the 290, but it cost me less than half what the 980 did.

 

[jpiniero]

Looks like 'press launch' with reviews. Actual launch could happen at or right after Computex though.

It almost has to be a paper launch though, even the GDDR5 models. Which is fine, I guess and pretty inevitable that would happen once AMD started talking about being several months ahead.

 

[Arachnotronic]

It almost has to be a paper launch though, even the GDDR5 models. Which is fine, I guess and pretty inevitable that would happen once AMD started talking about being several months ahead.

Unlikely.

 

[jpiniero]

Unlikely.

You really think it's going to be a hard launch? Have to figure all of the wafers they've been allocated up to now have been going straight to the P100. At some point they will get going on GP104, but it takes time to build the kind of volume needed for a real launch. This is just a paper launch to mess with Polaris.

 

[USER8000]

I expect both will be limited launches since Nvidia and AMD are probably more worried about getting cards to commerical customers and into prebuilt desktops and laptops.

 

[tential]

Actually, it isn't doing remotely awfully in DX12. Just not as well as in DX11. The architecturally balanced comparison for the 970/80 is after all the 280/X.

They are mid range on quite a few metrics of course. The base complaint(s) seem to be down to two things. Firstly that the 970/80 would traditionally have dropped a chunk in price by now.

Quite true I guess - but then in a fully traditional world, I guess AMD would have managed to ship Fury in big volume say 6 months earlier than they did, retiring the 290/X. That probably would have probably pushed the 97/80 price down in response.

Also I think some resentment about the idea that anyone might care about anything other than performance/money. That brand of hatred is more reserved for the 960 though

Why would the Fury retire the 290?
How does that make any sense? Even the 290x?

They aren't the same performance brackets or price levels so that makes zero sense....

 

[jpiniero]

I expect both will be limited launches since Nvidia and AMD are probably more worried about getting cards to commerical customers and into prebuilt desktops and laptops.

Polaris should be a hard launch since GloFo obviously isn't doing anything else on SS14 at this point. Of course, that might be part of the problem why Polaris isn't out yet, the yields are crap.

 

[Arachnotronic]

You really think it's going to be a hard launch? Have to figure all of the wafers they've been allocated up to now have been going straight to the P100. At some point they will get going on GP104, but it takes time to build the kind of volume needed for a real launch. This is just a paper launch to mess with Polaris.

Yes, I do. Your reasoning for why it will be a soft launch while AMD's Polaris will be a hard launch makes absolutely zero sense.

 

[Bacon1]

If GCN is so great, why is it such a power hog? Engineering is all about trade offs, and it looks like the ones NVIDIA made allowed it to sweep the gaming notebook market as well as the most lucrative portions of the high-end desktop market.

Yeah, I guess you can say "oh NVIDIA buyers are sheep/idiots/stupid/more money than brains" but that doesn't really hold up in my book.

I'm sorry GCN is bad perf/watt?

https://www.techpowerup.com/reviews/AMD/R9_Nano/31.html

Nano is same as 980 TI, and thats with really old drivers.

And thats done without having to cut out hardware DX12 support

I know you own SLI 980 TI, but you have to admit that GCN was better architecture, considering cheaper 290's are still keeping up with latest more expensive Nvidia cards in DX12 games.

 

[Arachnotronic]

I'm sorry GCN is bad perf/watt?

https://www.techpowerup.com/reviews/AMD/R9_Nano/31.html

Nano is same as 980 TI, and thats with really old drivers.

And thats done without having to cut out hardware DX12 support

I know you own SLI 980 TI, but you have to admit that GCN was better architecture, considering cheaper 290's are still keeping up with latest more expensive Nvidia cards in DX12 games.

980 Ti looks like it's winning and doing so at a higher performance point while using significantly less efficient GDDR5 memory versus the expensive and boutique HBM1. Proper comp is Fury X and it looks like the 980 Ti is ripping that one to shreds.

 

[IllogicalGlory]

Maxwell wins in performance per watt because it was designed with power efficiency in mind and is much newer than GCN, even GCN 1.2.

No one is denying NV the win in performance per watt, we're just wondering why that's the only metric that matters. It's a little bit troubling when a $300 card that's 2.5 years old rivals the performance of a 1 year old $600 card, even though said $600 card is designed only for the task the $300 is matching it in. That's what's going on in Quantum Break, Hitman, Ashes of the Singularity and Killer Instinct, remember? And you can only used the Gaming Evolved defense in two of those games.