Anybody else unimpressed with new midrange Nvidia GPUs, and much higher MSRP?

[renderstate]

https://www.khronos.org/registry/gles/extensions/OVR/multiview.txt

https://www.khronos.org/registry/gles/extensions/OVR/multiview2.txt



On current GPUs you can do geometry instancing, rendering instancing as well, and have different projection matrix and even different modelview matrix for each render target and also do this in a single-pass, but, this is getting way out of topic for this forum section.



I still say that until we actually see the SDK, and the actual code (will it be a new extension?), that this is nothing more than smoke & mirrors (speculation), and I am betting it can also be done on all current generation hardware as well.

First of all that's an OpenGL ES extension, not exactly the reference market for top VR apps. Second, the extension doesn't guarantee that the geometry is not going to be sent and transformed twice, in fact it's explicitly stated you don't have to support that. No current GPU on the market can do it (Pascal excluded). That extension does allow to avoid writing commands for the GPU twice, saving CPU cycles. Pascal is probably the first GPU that can support that extension AND implement it in a fully efficient way (i.e. save CPU and GPU cycles, not just the former).

 

[Azix]

How is the founders edition worse for consumers? A Nvidia is charging as high price as possible for new gpus since they have NO COMPETITION ?

They have every right to charge these prices and if people continue to buy then Nvidia clearly has the right move.

AMD is insanely idiotic to cede the high end to Nvidia. They deserve whatever Nvidia has in store for them this then as they brought it on them selves. The 1070 will repeat the 970 success and Nvidia has not only increased prices but hyper segmented again with founders editions and leaving their 1080ti til later to again get consumers to upgrade.

Nvidia just spoiled the amd parade. If you want to play blind go ahead.

Just want to put some fairness into this. They are ceding the high-end in price. They are still free to squeeze out whatever performance they can out of the price point they have chosen. They won't put a cap on performance just because they aren't trying for $600.

The deciding factor will be clock speed and what power consumption they choose to target. If they can't get high clocks then maybe trouble. If they can get similarly high clock speeds then, with the rumored core counts, its going to be significantly faster than 390x. It could match or beat the 1070 for a lower price. remember 390x is 5.9 Tflops, 1070 is just 6.5tflops. A good guess would be that polaris 10 beating 1070 is a given. Its a huge difference between 1070 and 1080. Under dx12 anyway. dx11 performance might also be improved so they can get closer to their theoretical peak. a 1070 could end up just matching a fury or 390x in dx12. In maxwell there was only a 0.8 tflops difference between 980 and 970. This is a 2.5 tflops diffference between 1080 and 1070. A difference which could suggest yield issues resulting in using much lower spec chips for the 1070.

Not going for that $600+ price point is not going to hurt the margins as much as they will gain targeting the larger market with really good performance. If they reach 980ti perf or close to it, at $300, its a done deal.

and I suspect they are building up inventory for a solid launch, not this rushed launch nvidia seems to be planning (which could be a reaction to polaris). At a good price and performance they will need good stock to not piss people off and lose sales to nvidia.

 

[tential]

Just want to put some fairness into this. They are ceding the high-end in price. They are still free to squeeze out whatever performance they can out of the price point they have chosen. They won't put a cap on performance just because they aren't trying for $600.

The deciding factor will be clock speed and what power consumption they choose to target. If they can't get high clocks then maybe trouble. If they can get similarly high clock speeds then, with the rumored core counts, its going to be significantly faster than 390x. It could match or beat the 1070 for a lower price. remember 390x is 5.9 Tflops, 1070 is just 6.5tflops.

A good guess would be that polaris 10 beating 1070 is a given. Its a huge difference between 1070 and 1080.

Not going for that $600+ price point is not going to hurt the margins as much as they will gain targeting the larger market with really good performance. If they reach 980ti perf at $300 its a done deal.

and I suspect they are building up inventory for a solid launch, not this rushed launch nvidia seems to have (which could be a reaction to polaris). At a good price and performance they will need good stock to not piss people off and lose sales to nvidia.

Polaris 10 beating 1070 puts Polaris 10 at faster than the 980ti/390x/Fury X, which is a completely different story.... if it's also cheaper. But from all reports, it's supposed to be a mainstream offering, and Fiji isn't discontinued.....

 

[escrow4]

If the 970 was gimped will the 1070 have any special surprises in stock? Its already heavily cut back. And the top shelf 1080 (and Pascal in general) that you spent billion and billions on is only 25% faster? That's it? This is next gen? It seems like Nvidia has paper launched a pair of mid range cards that are overpriced.

 

[Ajay]

And the top shelf 1080 (and Pascal in general) that you spent billion and billions on is only 25% faster? That's it? This is next gen? It seems like Nvidia has paper launched a pair of mid range cards that are overpriced.

New process tech nodes don't reduce costs (and hence prices) anymore. Nvidia complained about that a few years ago - there's really nothing that foundry's, like TSMC, can do about it. Right now yields probably aren't awesome on 16FF+ for 300+ mm^2 dice, so that increases cost as well. Add to that, the 1080 comes with GDDR5X, cutting edge RAM that isn't cheap either. All this boils down to an expensive AIB. So I'm not surprise. The only thing that annoys me is that I can't afford one right now.

Does NV make a great profit on these? I'm sure they will - but they are a 'for profit' company. Their job is to maximize profits for their shareholders. Sucks, but that's life.

 

[crisium]

Do we know how many ROPs these have?

 

[Azix]

Polaris 10 beating 1070 puts Polaris 10 at faster than the 980ti/390x/Fury X, which is a completely different story.... if it's also cheaper. But from all reports, it's supposed to be a mainstream offering, and Fiji isn't discontinued.....

mainstream concerns price. And these things are kind of tricky. because in dx12 a 390x would come close to or beat a 980ti. Changes to the arch could actually result in a 6.5 tflop polaris being faster than a 980ti. Not so easy to guess where things will end up till we see the actual benches. That performance seems very much in reach though.

that they they haven't stopped fiji might be a good point, though we don't know if they will or what the state of the shipments/manufacturing is.

By launching early they may avoid direct launch comparisons to polaris and possibly dampen the reaction when direct comparisons emerge.

 

[DaveSimmons]

If the 970 was gimped will the 1070 have any special surprises in stock? Its already heavily cut back. And the top shelf 1080 (and Pascal in general) that you spent billion and billions on is only 25% faster? That's it? This is next gen? It seems like Nvidia has paper launched a pair of mid range cards that are overpriced.

Faster and cheaper is still faster and cheaper, but (like intel CPUs) it shows the problem with lack of competition at the high end.

Hopefully AMD will at least match the 980ti and come in with a low enough price that nvidia is forced to drop the 1070 and 1080 prices and release the 1080ti.

 

[crisium]

My impressions overall are favorable, other than the price. It's another $50 hike over the 980 which was a $50 hike over the 680. Mid range creeping up. But with no new products from AMD quite yet, it's not surprising.

I really how Nvidia gets the best performance out of limitations. The 1070 will have the fastest clocked GDDR5 memory yet at 8GBs, and the 1080 will have GDDR5X which is even faster. 10GBs Impressive bandwidth for 256-bit. Excellent for mid range to push barriers like this, and the last hurrah of non-HMB2 cards having the performance crown. In the face of this, if the rumours of Polaris having only 6GBs DDR5, which is actually slower than nearly 2 year old Maxwell's 7GBPs, then AMD would be willingly choosing lower performance.

 

[Glo.]

Do we know how many ROPs these have?

Yes, 64, because of 256 Bit memory bus.

 

[moonbogg]

I think if someone who is price conscious pays $700 for 1080, they may regret it if 1080ti comes out 9 months later for the same price and is twice as fast, lol. I expect the big dies to absolutely haul ass.

 

[DooKey]

I think if someone who is price conscious pays $700 for 1080, they may regret it if 1080ti comes out 9 months later for the same price and is twice as fast, lol. I expect the big dies to absolutely haul ass.

How do you figure? Price conscious people don't spend $700 on a video card. For others who aren't price conscious they should buy 1080, sell before Ti comes out for 75-80% of the new cost. Apply towards Ti. Let the good times roll. Enjoy new tech every release that way.

 

[alcoholbob]

Polaris 10 beating 1070 puts Polaris 10 at faster than the 980ti/390x/Fury X, which is a completely different story.... if it's also cheaper. But from all reports, it's supposed to be a mainstream offering, and Fiji isn't discontinued.....

I don't know how AMD is going to beat a 596mm² die with a 232mm² die, especially when you consider the addition of FinFETs take die space, so even if you double transistor density on a lower manufacturing process, you don't have the double the density on the die itself when you take into consideration the space you have to use for FinFets. It can clearly be seen with Pascal. In theory they should be able to fit in 8.6+ billion transistors on the GTX 1080 die if they were simply straight porting Maxwell. The problem is FinFets take die space, so they only managed to fit in 7.2 billion transistors.

I highly doubt that Polaris 10 will come anywhere close to a GTX 1070.

 

[xpea]

Not a chance. Q1 2017 is when GP100 launches for HPC buyers. No way they'd sell a $1000 Titan that early, much less a $750 1080Ti.

GP100 will only be for HPC, no doubt about it.
But a GP102 with 12GB GDDR5X on 384bit bus (480GB/s) will fill nicely 1080Ti name and $700 bracket to compete with Vega...

 

[maddie]

I don't know how AMD is going to beat a 596mm² die with a 232mm² die, especially when you consider the addition of FinFETs take die space, so even if you double transistor density on a lower manufacturing process, you don't have the double the density on the die itself when you take into consideration the space you have to use for FinFets. It can clearly be seen with Pascal. In theory they should be able to fit in 8.6+ billion transistors on the GTX 1080 die if they were simply straight porting Maxwell. The problem is FinFets take die space, so they only managed to fit in 7.2 billion transistors.

I highly doubt that Polaris 10 will come anywhere close to a GTX 1070.

What sort of reasoning is this? 14nm finfet has a 20nm backend, so you can fit 2X and in Samsung's case maybe 2.2X.

232mm^2 on this process roughly equals at least 464mm^2 on 28nm TSMC, maybe even 500mm^2. More than 390X and if there are any improvements to GCN shader performance, which we have been told has occurred, plus higher clocks, then you're in 980Ti territory easily.

Why is this so hard to imagine?

Nvidia not packing 2X can be from many design decisions. Power density factors, etc. Dissipating 180+ watts from 310mm^2.

A 1070 looks like 75-80% of a 1080. Right around a 980Ti.

 

[lixlax]

I don't know how AMD is going to beat a 596mm² die with a 232mm² die, especially when you consider the addition of FinFETs take die space, so even if you double transistor density on a lower manufacturing process, you don't have the double the density on the die itself when you take into consideration the space you have to use for FinFets. It can clearly be seen with Pascal. In theory they should be able to fit in 8.6+ billion transistors on the GTX 1080 die if they were simply straight porting Maxwell. The problem is FinFets take die space, so they only managed to fit in 7.2 billion transistors.

I highly doubt that Polaris 10 will come anywhere close to a GTX 1070.

It all depends on how much they can rise the clocks and improve the architecture(ipc). Seeing that Pascal can easily run 1500MHZ+ shows that Nvidias architecture is still high clocking as is TSMC's process node. I really hope that those 800-1000MHz rumours for Polaris aren't accurate.

 

[AtenRa]

Can anyone tell me what is the source for the 330-340mm2 die size ??

 

[escrow4]

I don't know how AMD is going to beat a 596mm² die with a 232mm² die, especially when you consider the addition of FinFETs take die space, so even if you double transistor density on a lower manufacturing process, you don't have the double the density on the die itself when you take into consideration the space you have to use for FinFets. It can clearly be seen with Pascal. In theory they should be able to fit in 8.6+ billion transistors on the GTX 1080 die if they were simply straight porting Maxwell. The problem is FinFets take die space, so they only managed to fit in 7.2 billion transistors.

I highly doubt that Polaris 10 will come anywhere close to a GTX 1070.

Nowhere near except in price. Which will sell more - Polaris 10 (the mid/upper mid range chips, say 380/390 equivalent, not the higher end ones) at $249 or GTX 1070 at $379?

 

[master_shake_]

you know what's really disappointing about this gpu announcement?

it's been less than 24 hours and there is not one video of this guy

and the founders edition cards.

that's the real tragedy here.

 

[Silverforce11]

Polaris 10 beating 1070 puts Polaris 10 at faster than the 980ti/390x/Fury X, which is a completely different story.... if it's also cheaper. But from all reports, it's supposed to be a mainstream offering, and Fiji isn't discontinued.....

Polaris 11 and 10 is replacing AMD's entire lineup.

HBM1 stacking and interposer integration for AMD happens at eSilicon and Amkor (remember these guys?).

They have moved on to a 14nm FF + HBM2 process, currently due for a ramp by September. I posted their conference video awhile back where they mention all the stuff they are working on.

Fiji SKUs are obsolete, AMD would not keep wasting the production capability of HBM integration on Fiji when they need to ramp up for Vega.

 

[Silverforce11]

Can anyone tell me what is the source for the 330-340mm2 die size ??

Die shot, compared to GDDR5 modules. It's smaller than GM204. I'd say ~330mm2 is a good estimate.

 

[Silverforce11]

I think if someone who is price conscious pays $700 for 1080, they may regret it if 1080ti comes out 9 months later for the same price and is twice as fast, lol. I expect the big dies to absolutely haul ass.

GP100 in a GTX form, should be ~25% faster than the 1080.

1080 has 2/3 of the GP100's FP32 CC, and none of it's hindrances such as 1:2 FP64 & NVLink. P100 Tesla already hit the 300W barrier, as a cut-down part with 1480mhz boost clocks. Don't expect the GTX variant to be running 1.86ghz.

This is why some people think there's a gaming all FP32 GP102 chip somewhere.

 

[Stuka87]

I think the bigger issue for 980Ti owners is the 1070. Once after-market 1070 cards start coming out in July for $380, 980Ti won't be worth more than $325-350 on the used market. So instead of purchasing a 1080, it may be possible to 'downgrade' to a 1070 and have similar performance to a 980Ti OC with all the other benefits such as features, power usage, more VRAM, etc. But ya, if someone is water cooling, then all of this is out the window now with Founders Editions tacking on huge premiums on reference blower cards.

I highly doubt we will see after market cards for 380. 380 will be for the base reference cards. You can add 20-50 dollars for after market cards.

But even that destroys the resale value of any previous gen card.

 

[el etro]

VR World rumor says that Polaris 10 will fit in both 480 and 490 series, with Tweaktown saying non-x 480 packs already 390x performance. Benchlife says it is a ~115W ACP card, Wccf says it performs like a TitanX on FStrike at far less than the 170W card TDP. And the rumor mill says it packs perf/mm2 and perf/w lead.


So is not too much optimism to expect Polaris 10 beatimg GTX 1070.

 

[boozzer]

Does anyone know when we will get to see some benchmarks?