Kitguru : Nvidia to release three GeForce GTX 800 graphics cards this October

[MagickMan]

Not necessarily. 680 launched at $499 and stomped the more expensive 7970.

As was said, I don't consider a 6% advantage a "stomping". You can't even discern a difference that small in gameplay.

Anyway, my point is that nvidia isn't going to leave money on the table, that would be stupid, and that's what they'd be doing if the new 880 is substantially quicker than a 290X. $499 means it's a small, incremental performance boost, likely with improvements in thermals and features.

 

[Keysplayr]

As was said, I don't consider a 6% advantage a "stomping". You can't even discern a difference that small in gameplay.

Anyway, my point is that nvidia isn't going to leave money on the table, that would be stupid, and that's what they'd be doing if the new 880 is substantially quicker than a 290X. $499 means it's a small, incremental performance boost, likely with improvements in thermals and features.

We've moved on from this a while ago.

 

[Lepton87]

2x R290 will stomp all over it for less money. MEH.
UPDATE: Maybe I overestimated its price thinking it will be the same as 780ti which has terrible price for its performance, only my card is worse in this metric.

Do you think it is worth waiting for it over 2x r290 Sapphire Vapor X? Will it offer better or comparable performance per dollar? I'm very close to pulling the trigger for those two cards.

 

[tviceman]

GM204 vs Tahiti (R7 290X). I bet on this one.

GM204 will be faster at 1080p and 1440p, Tahiti and GK110 will be faster at 1600p and up.

GM204 (GTX880 ??) will be released at $499, R9 290X will fall to $379 and Tonga to be released at $249. GTX-780 will be EOL and GTX-780Ti will be at $649.

Nostradamus is back :biggrin:

Why would AMD need to drop pricing if GM204 is slower than Hawaii at high resolutions? Your prediction is terribad.

 

[Lepton87]

So the price in unknown as of yet? The cheapest 780Ti costs 2500PLN here, r290 Vapor-X costs 1700 and the cheapest aftermarket R290 is about 1300PLN. Is it worth the wait? What I'm afraid is that it's just going to replace 780ti pricing-wise while offering 20% performance boost, so I can have a better value now by buying two radeons. I'm slowlying giving up on trying to buy a second titan, there's one on sale and it has no warranty and the guy wants 2500pln the same as a new 780Ti. Crazy price. I offered him 2000PLN but after having thought about it it's still too much considering it has no warranty.

 

[3DVagabond]

Why would AMD need to drop pricing if GM204 is slower than Hawaii at high resolutions? Your prediction is terribad.

AMD has to be cheaper. End of story. Maybe someday they can turn that around, but it hasn't happened yet.

If the 880 simply matches the 780 non ti in performance and is as efficient as many are hoping @ $500, the 290X will have to be $400 or they'll lose 1/2 their sales. Look at how much cheaper the 7970GHz had to be compared to the 680 and the 680 didn't beat it in anything.

 

[MagickMan]

We've moved on from this a while ago.

I hadn't. You're the one who first replied to me, remember?

 

[Keysplayr]

I hadn't. You're the one who first replied to me, remember?

Oh, well then you can browse the debate 3DVagabond and I had. Just pretend we had one to because it would be a clone of that one. Save us both some time.

 

[Keysplayr]

I think many here are focusing on the 256 bit bus for a prediction of GM204's performance at higher resolutions. Remember, this is a new architecture with a MUCH larger cache design and may not really need a tremendous amount of bandwidth to compete at higher resolutions. This is just spec of course but look at what GM107 can keep up with at 128 bit bus and at MUCH lower power consumption. Maxwell is very impressive so far. I can't wait to see these new GPUs.

 

[CakeMonster]

Well to what extent the bus matters depends on what its meant to compete with. If it is supposed to perform better than 780* and 290*, then the chip will have to have unprecedented improvements to still be able to pull ahead with a 256bit buts, especially at 1600p/4K. I have a hard time believing we will see that this late in 28nm's lifespan. And if its going to be in that performance tier, why not go with 384 or 512bit since the price will be in the upper tier anyway?

If I was to guess, I'd rather assume NV will need to go 512bit this time to pull ahead of 290*/780* at higher resolutions. If they don't it will either be a midrange chip, or they have pulled some magic and are saving the 384/512 bus for yet another release in 3+months (not likely IMO).

 

[Keysplayr]

Well to what extent the bus matters depends on what its meant to compete with. If it is supposed to perform better than 780* and 290*, then the chip will have to have unprecedented improvements to still be able to pull ahead with a 256bit buts, especially at 1600p/4K. I have a hard time believing we will see that this late in 28nm's lifespan. And if its going to be in that performance tier, why not go with 384 or 512bit since the price will be in the upper tier anyway?

If I was to guess, I'd rather assume NV will need to go 512bit this time to pull ahead of 290*/780* at higher resolutions. If they don't it will either be a midrange chip, or they have pulled some magic and are saving the 384/512 bus for yet another release in 3+months (not likely IMO).

Yeah, it's sure a hard sell to think that a 28nm 256 bit GPU could outperform a 780Ti. I'll agree there. The unprecedented improvements did manifest a little in GM107. That, I think, is the only reason anyone is entertaining the idea of a mid-range GM204 outperforming a 780Ti at a smaller power envelope on 28nm. A lot to process, but fun at the same time.

And I think 384 bit bus will be reserved for GTX980 or GK200/210 on 20nm later on in 2015, IMHO.

 

[CakeMonster]

And I think 384 bit bus will be reserved for GTX980 or GK200/210 on 20nm later on in 2015, IMHO.

In that case, wouldn't you want the widest possible bus right now if you figure this is the flagship? There is a risk this might be the flagship for a long time too, you want to stay ahead of possible competition in that period AND you want to convince people its worth upgrading to. Just too risky to not go with more than 256 bit IMO.

So I'm still guessing mid range unless I'm missing something.

 

[Keysplayr]

In that case, wouldn't you want the widest possible bus right now if you figure this is the flagship? There is a risk this might be the flagship for a long time too, you want to stay ahead of possible competition in that period AND you want to convince people its worth upgrading to. Just too risky to not go with more than 256 bit IMO.

So I'm still guessing mid range unless I'm missing something.

Well, it seems to have worked great with Kepler. So I think Nvidia will attempt this model for as long as they possibly can.

 

[blackened23]

Well to what extent the bus matters depends on what its meant to compete with. If it is supposed to perform better than 780* and 290*, then the chip will have to have unprecedented improvements to still be able to pull ahead with a 256bit buts, especially at 1600p/4K. I have a hard time believing we will see that this late in 28nm's lifespan. And if its going to be in that performance tier, why not go with 384 or 512bit since the price will be in the upper tier anyway?

If I was to guess, I'd rather assume NV will need to go 512bit this time to pull ahead of 290*/780* at higher resolutions. If they don't it will either be a midrange chip, or they have pulled some magic and are saving the 384/512 bus for yet another release in 3+months (not likely IMO).

Silicon isn't that simple. There are about 25000 variables that go into final performance of a product and bus width is only 1/25000 of the equation. If bus width it were the end all be all, well then, the 680 would not have outperformed the 580. This is just one example of course, but you can't blindly look at GPU specs and come to a sweeping conclusion. And i've also brought up the HD2900 series which was the first GPU (IIRC) which had a 512 bit bus and it also really underperformed.

Bus width matters, but there are other factors in the transistor and silicon that also matter too. Of course, it is confusing that the 880 may have a 256 bit bus, but I still feel like it will outperform the prior generation by a decent margin at all resolutions. I could be wrong, but that's just my guess. I just can't see NV releasing next gen GPUs that aren't better in every way than prior generation - I mean, thinking logically, what's the point of that? They wouldn't win any mindshare if that were the case and it would be a complete waste of time really. If I remember right, the 680 maintained a 30%+ lead over the 580 that widened or maintained at high resolutions. But when the rumors of the 680 hit (before launch) it was the same thing. Nobody thought it would beat the 580 which had better specifications and a wider bus. We'll see I guess.

The one rumor that sounds kinda iffy to me is the 400$ price floating around. If it's 20% faster than the 780ti I don't think it would be 400$. 550$-600$ sounds more likely depending (depending on whether it's 15 or 30% faster, who knows), but who knows. Of course if it is 400$, that sounds good to me. That would also have the side effect of lowering prices of all similar products across the board indirectly so everyone wins regardless of what brand/whatever they prefer. But I kinda doubt that 400-450$ scenario will happen. I hope to be pleasantly surprised though.

 

[CrazyElf]

Bus width matters, but there are other factors in the transistor and silicon that also matter too. Of course, it is confusing that the 880 may have a 256 bit bus, but I still feel like it will outperform the prior generation by a decent margin at all resolutions. I could be wrong, but that's just my guess. I just can't see NV releasing next gen GPUs that aren't better in every way than prior generation - I mean, thinking logically, what's the point of that? They wouldn't win any mindshare if that were the case and it would be a complete waste of time really. If I remember right, the 680 maintained a 30%+ lead over the 580 that widened or maintained at high resolutions. But when the rumors of the 680 hit (before launch) it was the same thing. Nobody thought it would beat the 580 which had better specifications and a wider bus. We'll see I guess.

Supposedly the larger L2 cache would offset that narrower bus. Well, it's plausible based on the 750Ti's performance. It's higher resolutions though that we see the 290X close the gap so to speak with the 780Ti. At 2560x1600/1440, the 780Ti dominates the 290X at most titles. At 4k, it starts to trade blows. Similar stories with 3x 2560 monitors. It's also I think the point where the 4Gb of VRAM probably helps the 290X more. So perhaps this card will dominate at 2560x1600 or under? Hypothetically though, there's nothing stopping Nvidia from coming out with a ~550mm^2 "big Maxwell". 28nm is a mature process and Maxwell too appears to be completed. Price-wise, Nvidia rarely competes on price, so don't be surprised if this is not a cheap card, so to speak. We'll have to wait and see what AMD has to offer before there's any price cuts.

 

[tviceman]

Supposedly the larger L2 cache would offset that narrower bus. Well, it's plausible based on the 750Ti's performance. It's higher resolutions though that we see the 290X close the gap so to speak with the 780Ti. At 2560x1600/1440, the 780Ti dominates the 290X at most titles. At 4k, it starts to trade blows. Similar stories with 3x 2560 monitors.

AMD has scaled better with higher resolutions since Cayman (HD6970), despite the fact that Cayman only had a 256-bit bus. This may have to do with fill-rate, and it may (or may not) also have to do with how the graphics pipeline for each respective company is optimized. Nvidia may purposefully be concentrating on 1080p because it is the overwhelming most common resolution. Steam survey shows that 1440p and up represents less than 2% of their user base. OR, it may be that AMD's driver overhead is less than Nvidia's, which becomes more apparent as the bottleneck moves more and more onto the GPU (i.e. higher resolutions).

 

[BrightCandle]

AMD has scaled better with higher resolutions since Cayman (HD6970), despite the fact that Cayman only had a 256-bit bus. This may have to do with fill-rate, and it may (or may not) also have to do with how the graphics pipeline for each respective company is optimized. Nvidia may purposefully be concentrating on 1080p because it is the overwhelming most common resolution. Steam survey shows that 1440p and up represents less than 2% of their user base. OR, it may be that AMD's driver overhead is less than Nvidia's, which becomes more apparent as the bottleneck moves more and more onto the GPU (i.e. higher resolutions).

I personally think its just their multithreaded drivers. When I trace Arma on AMD and Nvidia cards I see a clear difference in the amount of CPU time spent in DirectX, and Nvidia shows a lot more parallelism and a lot shorter times in quite a lot of cases. I presume this follows across to other games but I have only done such detailed traces in Arma since its so heavily CPU dominated and I was trying to help the developers fix the problem.

 

[3DVagabond]

I personally think its just their multithreaded drivers. When I trace Arma on AMD and Nvidia cards I see a clear difference in the amount of CPU time spent in DirectX, and Nvidia shows a lot more parallelism and a lot shorter times in quite a lot of cases. I presume this follows across to other games but I have only done such detailed traces in Arma since its so heavily CPU dominated and I was trying to help the developers fix the problem.

That would explain why as the scenario becomes more GPU bottlenecked the AMD cards catch up. Without further investigation I'm not saying this is definitely the case, but it makes sense.

 

[BrightCandle]

That would explain why as the scenario becomes more GPU bottlenecked the AMD cards catch up. Without further investigation I'm not saying this is definitely the case, but it makes sense.

AMD also seems to have much faster hardware at a raw specs level. From more memory throughput to compute performance it would be a shock to me if it didn't outperform Nvidia. Another way to phrase this is that I have never understood really why AMD's cards appear to under perform at lower resolutions. DirectX overhead is certainly a possible answer, as is my observation of multi threading in the drivers, but I also think its too simple an answer. AMD ought to kick Nvidia's cards around the block and back with a big margin based on the raw numbers, yet even at 4k it doesn't look like its really fulfilling its promise.

I have asked the question a lot of times and never been convinced anyone has a handle on the answer, it always boils down to "drivers". One thing I have been thinking recently is maybe AMD's hardware is actually buggy, rather than it being a software issue within the company its a hardware design issue and that actually a lot of time is spent working out software workarounds that rob the cards of performance. All speculation of course, based on an observation that, in my opinion, AMDs hardware should be capable of more than it currently achieves even at 4k where it currently has a noticeable advantage.

 

[bakalu]

Sorry my english

Ex: gtx 750 has 512 SP
gtx 650 Ti has 768 SP

gtx 750 is still 9% faster than gtx 650 Ti.

So 1 SP of maxwell is stronger 1.5 times than 1 SP of kepler.

If gtx 880 has 2560 SP (equivalent 2560*1.5= 3840 SP of kepler), and gtx 880 is certainly faster than gtx 780 ti.

 

[Pinstripe]

Did it ever occur to anyone that these photos of that ~430mm^2 die were actually from a GM200 sample?

 

[tviceman]

Yes it did occur, but was then shot down immediately because Nvidia is still on 28nm.

 

[FatherMurphy]

Also, a 430mm^2 GM200, even at 20nm, makes no sense. Nvidia's big dies are always above 500mm^2 (see GT200, GF110, GK110).

 

[RussianSensation]

Sorry my english

Ex: gtx 750 has 512 SP
gtx 650 Ti has 768 SP

gtx 750 is still 9% faster than gtx 650 Ti.

So 1 SP of maxwell is stronger 1.5 times than 1 SP of kepler.

If gtx 880 has 2560 SP (equivalent 2560*1.5= 3840 SP of kepler), and gtx 880 is certainly faster than gtx 780 ti.

You can't compare CUDA cores without looking at clock speed. The increase in IPC from Kepler to Maxwell 1st gen is 35% not 1.5x. If it was 50% faster, NV would have undoubtedly advertised such a figure. Secondly, even to assume a perfectly linear 35% increase in IPC per CUDA core could be a stretch since it would imply the rest of the chip has to be scaled in valance to prevent pixel/texture fill rate or memory bandwidth bottlenecks. For instance 780Ti OC does not scale linearly against 780 OC based in the difference in CUDA cores.

 

[Deders]

I think they calculated the cores as 20% stronger, I could be mistaken.