I think Nvidia messed up. They made a card that will cannabalize their future high end sales while being so far above their low end that they'll have to cut cards for market segmentation, not just yields. In essence they made GP104 too big.
Traditionally Nvidia "4" midrange chips have been pretty close in performance to their previous high end "0" designs, but the 4 GPC GP104 has exactly one more GPC than is required to match GM200, as shown by the 3 GPC cut down 1070 matching GM200 cleanly. I think if they had done a 3 GPC design from the start it'd have been much more competitive with AMD and frankly would make more profit.
#1: They have no direct RX 480 competitor, it'll take a 1536-1664 shader part to match. GP106 only has 1280, 1070 has 1920. They're going to have to make a 1060 Ti to compete. The problem is yields are such that chips that don't meet 1070, yet meet 1060 Ti, are not going to be common. Meaning they'll have to cut down otherwise perfectly good 1080 and 1070 dies to meet market demand, which loses money; what's the point of making a big die if you need to cut 40% off to sell it?
#2: With a 3 GPC design they could have fit 200~ vs the 160~ per wafer they can fit now, yields would be far better as it's exponential with die size. They also could have gone with GDDR5 memory as a 256 bit bus is sufficient for 3 GPC's, as shown by the 1070. Not only that, but they could have omitted the GDDR5X memory controller entirely, further saving on die size, and likely power consumption. Comparatively Polaris 10 fits 260~ per wafer, and probably ends up a quarter the price to produce assuming wafer costs between GloFlo and TSMC are comparable, wouldn't be surprised if GloFlo is cheaper.
#3: When GP100 comes out eventually, it won't be a compelling upgrade over GP104. We know GP100 has 6 GPC, 50% more than GP104, but remember GM200 only scaled around 20% above GM204 despite 50% more hardware by the numbers. Meanwhile GK110 had a clean 40% lead above GK104. Parallelization only goes so far, and increasing core counts isn't a linear gain. They're going to end up in another situation where there'll be a Titan and a 1080 Ti that is barely cut down, with a whole bunch of thrown out dies. Rather than having something akin to the 780 Ti and 780 over the 680/770. This loses money.
Honestly, with the design they went for it seems like a high end stopgap card, not a viable product that'll fit in with their later product stack. Its a 25% improvement, with GP100 being another 25%. The full chip is too expensive for most looking to upgrade, the performance gain is too small for those with a GM200 to upgrade (they want GP100), and GP100 won't be a compelling upgrade for those who are buying GP104. A 3 GPC chip could have easily been $350 launch due to cheaper manufacturing, not $450 as the 1070 is. All they benefit from the 4 GPC design is being able to claim to have the two top cards on the market.
tl;dr:
1080 too big; should have been a smaller design
Nvidia will sell simply because it is Nvidia. Fanboys will cling to it no matter what.
Consumers aren't getting GP100. They're behemoth expensive HBM2 monsters. I'd be a miracle if they released a cut down version this year.
>>55099141
GP102 is bretty much GP100 sans FP64 and HBM
>>55098821
Bullshit consumers, and especially OEM's, will go for the best priced option. If AMD can supply an equal performance card for half the price and with more supply behind it, OEM's will jump on it.
GP100 is for scientific compute. We'll probably see a GP102 for teh gaymen.
>>55099236
We're trying to have a conversation. At least know what the fuck you're talking about. Have a (You) to hold (You) out.
>>55098508
In that case, couldn't they just rebadge the 980 as the 1060?
>>55099272
16nm wafers are much cheaper because of wafer surface area.
The 1060 will be slightly better then the 480 but be $50 more like how they usually do there mid range game, the 480 is overrated until actual in game fps numbers are released
>>55099236
>Muh cores
2816 / 2048 = 1.375
2304 / 1.375 = 1675 which is close to 1664 m8
>Muh frequency
1050 / 1216 = 0.86
1266 / .86 = 1472
1070 only boosts 14% higher than that, and 86% of 1675 is 1440
1060 can't match it.
OP was right
>>55099310
>GTX 980 4.6 TFLOPS matches or exceeds AYYMD's 2816 HOUSEFIRES 5.6-5.9 TFLOPS in real world benchmarks
Barely and at low resolution, and at 2x the power draw.
>>55099318
>the 480 is overrated
It's currently overhyped. I'm not expecting anything amazing, but I'm expecting at least the performance of the low high tiers from last generation. I would be surprised if it actually had 980 performance. It would be a nice win for all of us, but I don't think it will happen. Two weeks seem so far away.
>>55099307
How difficult would it be to shrink it down, compared to a new gpu design?
Nvidia will have to pull some bullshit out of their ass for the 1060
How do shitposters have so much dedication?
970 was a sensation. 1080 and 1070 are sold out everywhere. The $320-$350 price point seems to be the new $200.
Nvidia doesn't need anything that directly competes with every AMD chip. Some generations they match, some they interleave.
Nvidia makes a massive profit on these. Remember, the chip that was in the $500 GTX 680 was eventually selling at $200 in the 760, and they were still making a profit.
Nvidia has another winner on their hands here. Great performance for great power consumption. Telling them not to make this chip is retarded. They can cut it down and still make a profit, while a lower capability chip wouldn't be able to warrant $700.
The last few generations have shown that after setting a high 104 bar, people will pay a premium for a mediocre upgrade to a monster just to have "the best."
This is working well for Nvidia: Don't market the (1)X80 as the premium, top of the stack, just leave it there for a while and release a Ti/workstation card later for even more. So now they've set simple 256bit cards all the way up at $350/$550 as "everyman" cards" and can milk even more from their expensive workhorses.
>>55099255
Another die would increase costs exponentially
GP102 will be a cut down GP100, unless they secured enough HPC deals to be able to make a die purely for them, and leaving enough to make a gaming only GP102
>>55099357
Less difficult than a new design, but the validation and tape out costs make it too expensive to be worth it, specially when they will have to launch a newer card later
The 980 can't into SIP HEVC decoding, which is essential in cards at those price ranges, there's a reason the 960 was the first card with HEVC decoding
Though they might go at it and rebadge the 980 as the 1060Ti, but that's a last ditch strategy, and they will probably lose money with something like that
>>55099391
>>55099366
>>55099353
>>55099236
>>55099310
OP was so right he triggered a shill into overdrive
>>55099217
FP64 would be easy to remove, but HBM wouldn't
They would need to ship a HBM and a GDDR5X memory controller in the same die, which would increase die size substantially
GP102 will be GP100 sans FP64 and with less memory, still HBM though
>>55099448
>1080 and 1070 are sold out everywhere
That's because no one can buy them. For recent examples, think the Fury X.
First off I want to say nice post OP. You're actually not a faggot this time.
Secondly, GP100 is missing key hardware to make it into a consumer gpu. There are no ROPs or TMUs. It's also an enormous die with basically no yield and it uses HBM2 which also has garbage for yields.
Consumer part will be GP102, with 50% extra shaders.
>>55099453
The only shill is you
Do you seriously think Nvidia doesn't model their performance before sending it to the fab?
https://forum.beyond3d.com/threads/nvidia-pascal-announcement.57763/page-13#post-1905723
>GP100 is fully graphics capable. They aren't drawn on the diagrams, but it has display controllers, ROPs, etc. And I agree a Quadro is a good bet at some point.
GP100 has ROPs & TMUs, stop lying if you guys don't know SHIT
>>55099448
>970 was a sensation
That's because the 960 was complete shit, and AMD made nothing but rebrands
>1080 and 1070 are sold out everywhere.
This is because their availability is non existant
>The $320-$350 price point seems to be the new $200.
Neither of this cards is anywhere near that price point right now
> They can cut it down and still make a profit, while a lower capability chip wouldn't be able to warrant $700.
At the RX480 price point they might still pull a bit of profit, but AMD is going to be pulling much more than them
>>55099484
>Do you seriously think Nvidia doesn't model their performance before sending it to the fab?
Do you seriously think Nvidia knows AMD's performance of unreleased products years in advance?
>>55099475
>Secondly, GP100 is missing key hardware to make it into a consumer gpu. There are no ROPs or TMUs
Source? GP100 as a HPC oriented chip makes sense, but not so much as a purely compute chip
>Consumer part will be GP102, with 50% extra shaders.
This would involve making a whole new die just for two halo product lines, it's almost impossible for it to be economically feasible
>>55099475
>Nvidia is going to fab two separate wafers with ROP & TMUs for Geforce and another one without ROP & TMU for workstations instead of just doing it in a single one and laser cutting what's not needed.
NO.
>>55098542
>We know GP100 has 6 GPC, 50% more than GP104,
GP100 also has a 1/3rd fp64 ratio, GP104 has a 1/32nd ratio like maxwell had. NVIDIA could easily 2-3x the performance over GP104 with a GP100 based card with the fp64 stripped out (which is useless for gaymen anyway).
>>55099622
You seem like you really want to buy a card for thousands of shekels.
>>55099530
>At the RX480 price point they might still pull a bit of profit, but AMD is going to be pulling much more than them
Nvidia is in the black. AMD is deep in the red.
The "Nvidia tax" that Nvidia warrants just for having working drivers is huge. You didn't think a GPU in a $500 card costs $500 to make, did you? They're more like $50. Being able to pull in another $15 per chip is a big deal.
>>55099732
>You didn't think a GPU in a $500 card costs $500 to make, did you?
You know the R&D of this chips is in the billions of dollars, right?
>>55099530
>That's because the 960 was complete shit, and AMD made nothing but rebrands
Not the same poster but the 970 was recognized as a sensation before the 960 was launched or AMD even announced it was mostly doing rebrands. Maxwell was noted for its increases in performance and decreases in power requirements on the same node. If only Nvidia had been open about the memory issue instead of being found out by the media.
>>55099708
the 980ti, 780 and 780ti were all $650-$700. would never spend more than that on a single GPU
>>55099758
Yes, but that doesn't mean they're selling the cards at a loss. The R&D is over.
Nvidia has been making more of a profit than their competitors while using a bigger die since the beginning of time. You can't just calculate it as being a loss to need a bigger chip at a certain price point, you need to factor in what it would cost to design an entirely new chip.
Nvidia using one chip for $700, $350, and $200 while making a profit on all three is not a worse situation than AMD only selling at $200.
>>55099732
16nm wafers cost around 6k each
Nvidia can only fit 160, thus $37.5 each assuming everything yields vs $23 for Polaris. In reality it's probably yielding 40% while Polaris yields 65%. Ie $94 vs $36~
>>55099866
>Yes, but that doesn't mean they're selling the cards at a loss. The R&D is over.
The R&D costs are amortized throughout the card generations, the money spent on it doesn't appear out of thin air
if the buyers of the 1080 are 3% of the total gpu market, and most of them are out of stock, just imagine how much terrible the yields are.
>>55098821
this
>>55099254
you're wrong, i literally bought a 1070 because it's Nvidia, for no other reason. i can confirm that Nvidia will win
>>55098542
I wondered about 2560 cores, but its OK. Similar to GM200, 50 % raw performance.
>>55099236
Kepler needs 2880 cores to compete a 2048 shader Maxwell. Thats the future.
>>55099255
This.
>>55099272
No, Pascal is Maxwell 3.0. Far more efficient. AMD already delivers 980 performance around 100W. We'll see a 1280 core Pascal GP106 and another even more cut down GP104 this autumn. Maybe 1400-600 cores, but only 192-bit 1060Ti. The 1060 will compete with Polaris and the 980.
>>55099622
GP102 was designed for that reason, it doesn't make sense to give gamers full FP64 support. But they will unlock all shaders instead. The Tesla P100 runs at 1.3 GHz, the gaming variant will probably run around 1.6, which will deliver 50-60 % performance increase over a 1080.
>>55099871
Are you talking about GP104?
>>55100071
>I wondered about 2560 cores, but its OK. Similar to GM200, 50 % raw performance.
By the numbers it should have 47% more performance.
25% more performance in reality, at stock clocks, and it OC's worse proportionally. Clock speed didn't scale mate. Very disappointing.
>Kepler needs 2880 cores to compete a 2048 shader Maxwell. Thats the future.
Again by the numbers the 980 would have been slower by 7%, but core count and frequency isn't linear. It instead got 7% more at release.
>No, Pascal is Maxwell 3.0. Far more efficient.
Around 40%, about in line with what the electro characteristics of the 16nm process vs 28nm.
>which will deliver 50-60 % performance increase over a 1080.
You're delusional. GP100 boosts at 1480. If GP102 keeps to anywhere near a similar clock speed you'll see a 30%~ theoretical. The 25% improvement OP stated was generous and likely assumed a much higher clock speed. Bigger chips clock slower, check GM200 vs GM204 as an example.
>Are you talking about GP104?
Yes
>>55100446
Why do you know that? I don't think GP102 will be clocked lower than 100.
No, the shader count does only apply within a generation, thats right. GCN needs 4096 shaders to compete 3072 Maxwell 2.0 cores around 1000 MHz.
I don't know if they changed a lot of the architecture itself. It seems more like a Maxwell shrink, allowing more transistors per mm2 and more clock speeds.
And yes, on stock. You can overclock GM200s to 1500 or 1550 without much problems. Even the shitty Titan I had reached 1450 around 1.2V. GM204 and 200 were similar in terms of OC, they never did more than 1500 or 1550 at normal voltage levels. I expect the Gp102 to reach similar OC levels like the 104 does, around 2 GHz. But the increse will be even higher if they clock it lower another time like the GM200.
>>55099898
Selling for a profit *3 is not a loss.
>Kia has a car they sell for $20k. They net $500 profit off it.
>Toyota made a more expensive platform. They sell a car based on it at $20k, but only make a $300 profit.
>That platform is the basis of a car they sell for $30k and make $6k profit from.
>And the basis of a car they sell for $40k and make $12k profit from.
>* and Toyota has a cheaper platform where they can make a profit off <$18k cars.
>"Kia is in better shape because they net more profit at $20k"
Somehow I don't think the financials are going to show that.
>>55100540
>Why do you know that? I don't think GP102 will be clocked lower than 100.
Because Nvidia released the full specs for GP100. It's not going to hit 1700mhz stock clocks mate.
>No, the shader count does only apply within a generation, thats right. GCN needs 4096 shaders to compete 3072 Maxwell 2.0 cores around 1000 MHz.
Those 4096 shaders only scaled 20% above 2816 shaders, at equal clocks, vs 45% theoretical.
Also GCN3 is meant to bring IPC increases
>I don't know if they changed a lot of the architecture itself. It seems more like a Maxwell shrink, allowing more transistors per mm2 and more clock speeds.
Not really on the consumer side; deeper pipeline, groups of 640 vs 512, a few ROP improvements, etc.
>And yes, on stock. You can overclock GM200s to 1500 or 1550 without much problems. Even the shitty Titan I had reached 1450 around 1.2V.
You actually got a good one. Average OC for GM200 is 1400mhz. GM204 typically hits 1500mhz. Scaled up vs the average 2050mhz GP104; GP102 Might hit 1900mhz; which I think is high.
>>55101054
For GP100 yes. For GP102 its not sure.
Even 1400 is a 40 % OC! Mine wasn't really good, 68 ASIC. I saw a lot of people with higher speeds, 1500 or near 1500, sometimes 1550, but thats already rare, even for GM204.
They are still working on a power target workaround for the GP104 cards, that is what is currently limiting max clock speeds. But I don't expect the large one to reach more than 2 GHz. Bigchips are always a bit slower.
>>55098508
until we see GP106, we won't really know if Nvidia completely fucked up, but the 1080/1070 seem kind of foolish once the release reality distortion bubble starts to fade.
I don't know how they intend to sell $500-$800 cards that everyone knows will be outperformed by 40%-$50% early next year.
The 1070 needs to be $300-$350, max.
>>55098508
>#1: They have no direct RX 480 competitor, it'll take a 1536-1664 shader part to match. GP106 only has 1280, 1070 has 1920. They're going to have to make a 1060 Ti to compete.
I laughed. 1070 is like 50% faster than a 480 who gives a shit if it looks better on paper if there are no drivers.
>>55101575
>The 1070 needs to be $300-$350, max.
Depending on how the Rx 480 performs, Nvidia could very well knock $30+ off the price
>>55101597
>Card stock is shit
>4x cost to manufacture
>Thinking the 1070 won't be $450 for months and months
Oh you sweet summer child.
>>55101582
Daily Reminder:
He's a poorfag who used a Core 2 Duo laptop with 9500 GT (9600M GT) until 2012 and only could afford an old Thinkpad with a workstation Fermi GT 430 (5400M) as a replacement that he uses to this day. He unironically uses Gentoo as a daily operating system.
He hasn't used an ATI card since they were bought by AMD.
Yes, he actually was stupid enough to buy a workstation Fermi when Kepler had already released.
He even admitted the drivers for the card broke frequently, especially with Optimus.
>>55098821
Or maybe it has something to do with drivers that actually work.
>>55099871
>yields that low
No? Do you not know that Intel doesn't even begin shipping parts unless the yields are over 60%?
>>55102235
Intel has direct control over process and design; they also fabricate far smaller parts for consumers than the 314mm2 GP104; quad core Skylake is 122.4mm2, and quad core Haswell is 177mm2 for example.
The last time Intel quad cores were around 300mm2 was on the 45nm process, with 296mm2 Lynnfield.
That includes the integrated GPU btw
>>55098508
>I think Nvidia messed up
Their profit margins say otherwise
/thread
>>55102304
Fabricating CPU and GPU isn't remotely the same thing, a modern CPU with all the bells and whistles with a diesize of a GPU will be harder to fab as the CPU doesn't have nearly as much as duplicated parts.
>>55102320
Size offsets that easily
>>55100057
>t. a retard
>>55101575
GP106 is gonna be shit
