Upcoming Maxwell GPU pictured!!!!

[RussianSensation]

I don't see anything wrong with NV and AMD sticking to 28nm node for another 2 years assuming there is a large improvement in piece/performance and absolute performance. I would rather have a 28nm card 50% faster than 780Ti for $500 with a 275W power use than a card 60% faster on 20nm with 200W but for $750. Once 20nm makes financial sense, I am sure NV and AMD will with go that route or even skip it entirely and shoot straight for 14/16nm designs. Having said that I think we will see products move beyond 28nm before end of 2016.

If NV can replicate 750Ti's 2x the performance/watt with 550mm2 Maxwell flagship, it will be a great product.

 

[Kippa]

Is there any chance that the next Nvidia card could handle 4K gaming reasonably well?

 

[Qwertilot]

Good question That's actually another potentially good reason for delaying with process jumps.

To be really attractive the next set of high end cards more or less have to do 4k sensibly, which entails a huge performance jump. If 20nm wouldn't have got them there anyway might as go with a shortish lived cycle of cards on the cheaper process while they're waiting.

A tiny bit like Intel juggling Broadwell/skylake etc around.

 

[Flapdrol1337]

If NV can replicate 750Ti's 2x the performance/watt with 550mm2 Maxwell flagship, it will be a great product.

Especially if it's a near 300W card.

 

[RussianSensation]

Is there any chance that the next Nvidia card could handle 4K gaming reasonably well?

Nope. Even 780Ti SLI can't max out the most demanding current gen games at 4K. It is 4x as many pixels as 1080p and 780Ti can't even max Crysis 3 at 1080p. We will need to wait until Pascal before a single-GPU card can reasonably provide 60 fps at 4K. I still think 4K will require multiple GPUs for a while since I honestly expect PC games to get more demanding in the next 2-5 years or otherwise I will concede that PS4/XB1 have failed to get developers to target higher end baseline hardware.

 

[tviceman]

Russian, you keep aiming too low with the potential performance of the upcoming AMD/Nvidia cards. If the GM107 is anything to go by (and it should be), Nvidia should be able to deliver 65-70% more graphics performance with the same power consumption as gtx 780 ti.

AMD is the wildcard here, since so little is known about their upcoming parts. But AMD seems to go after more dense designs, so they may beat Nvidia on perf/mm2 while nvidia wins on perf/watt. Either way, if we get >50% gtx 780 TI for $549 I'd be pleasantly surprised. More than likely, I'm guessing if we see that kind of jump, $599 is the more realistic starting price.

 

[BFG10K]

Those opposing connectors aren't so strange; the GTX680 did the same (it only had two instead of three).

I didn't like them as it made things harder to plug in. Everything in a row on the top is the best.

 

[Cloudfire777]

The GPU is confirmed GTX 880 (GM204). Estimated to be about 430mm2.

https://translate.google.com/translate?sl=auto&tl=en&js=y&prev=_t&hl=no&ie=UTF-8&u=http%3A%2F%2Fwww.computerbase.de%2F2014-07%2Fgeforce-gtx-880-gm204-pca-bilder-nvidia%2F&edit-text=

 

[alcoholbob]

If the die size estimates are accurate then it probably wont be more than 5-10% faster than 780ti even if maxwell is 30-40% more efficient than kepler. Which sounds exactly like what the early leaks said, 10% faster than 780ti and cheaper.

 

[tviceman]

If the die size estimates are accurate then it probably wont be more than 5-10% faster than 780ti even if maxwell is 30-40% more efficient than kepler. Which sounds exactly like what the early leaks said, 10% faster than 780ti and cheaper.

Maxwell is 70% more efficient than Kepler in perf/watt (gm107 vs. gk107). If this card consumes about the same power as gtx770, then it should end up ~25% faster. If it consumes around 200 watts on average gaming loads, then it should end up around 50% faster.

All of this is assuming it's ROP's and bandwidth are balanced to Maxwell's characteristics.

 

[Cloudfire777]

GTX 750 Ti (GM107):
2x the performance of GTX 650 (GK107)
About the same power consumption as GTX 650

GTX 880 (GM204):
2x the performance of GTX 680 (GK104). Which means about 30-40% faster than GTX 780 Ti.
About the same power consumption as GTX 680.

Depends what route Nvidia plan for this card. They could increase TDP from 680 and increase performance more. Which may have happened depending on what power plugs they keep (6+6 or 6+8). GTX 680 is 6+6.

 

[Grooveriding]

Wild estimates and getting carried away with things here. Reminds me of GTX 680 madness pre-launch.

20-30% faster than 780ti. Normally I'd expect cheaper, but realistically it is going to be $600-$700 given 28nm pricing. This board pictured is almost certainly a Quadro card imo. 16GB VRAM ? No way, not from nvidia unless this is a $1000 card. 8GB is what we will see on a geforce variant.

This may not even release as geforce for a while. Maybe they will do like they did with GK110 and release the pro card first and leave 780ti in place. No real reason to replace it right now and they make more money off pro cards. Seems likely this is a professional card with the 16GB.

 

[Cloudfire777]

Wild estimates and getting carried away with things here. Reminds me of GTX 680 madness pre-launch.

20-30% faster than 780ti. Normally I'd expect cheaper, but realistically it is going to be $600-$700 given 28nm pricing. This board pictured is almost certainly a Quadro card imo. 16GB VRAM ? No way, not from nvidia unless this is a $1000 card. 8GB is what we will see on a geforce variant.

This may not even release as geforce for a while. Maybe they will do like they did with GK110 and release the pro card first and leave 780ti in place. No real reason to replace it right now and they make more money off pro cards. Seems likely this is a professional card with the 16GB.

The board have 8GB VRAM (0.5GB per chip, 16 of them) and they are rated at 7000MHz
6GB/8GB is becoming more and more the average on high end cards now. GTX 870M have 6GB, 880M have 8GB. Titan Black 6GB. And so forth.
With the unified structure CUDA now have, 8GB could most likely an indication of a consumer cards as well.

 

[Grooveriding]

The board have 8GB VRAM (0.5GB per chip, 16 of them) and they are rated at 7000MHz

Your post says 16GB, now it is 8GB ? Still my guess is pro card first and then geforce. Probably see the first geforce as a $1000 Titan as well

 

[Cloudfire777]

Your post says 16GB, now it is 8GB ? Still my guess is pro card first and then geforce. Probably see the first geforce as a $1000 Titan as well

Yeah sorry about that. I was quoting the source which said it was 16GB. But people googled the VRAM "H5GQ4H24MFR" and found out they are 512MB per chip, not 1024MB like source indicated.

 

[ocre]

I don't see anything wrong with NV and AMD sticking to 28nm node for another 2 years assuming there is a large improvement in piece/performance and absolute performance. I would rather have a 28nm card 50% faster than 780Ti for $500 with a 275W power use than a card 60% faster on 20nm with 200W but for $750. Once 20nm makes financial sense, I am sure NV and AMD will with go that route or even skip it entirely and shoot straight for 14/16nm designs. Having said that I think we will see products move beyond 28nm before end of 2016.

If NV can replicate 750Ti's 2x the performance/watt with 550mm2 Maxwell flagship, it will be a great product.

there is also capacity to consider. Are you sure TSMC can deliver an entire shift from Amd & nvidia to large die 20nm chips at this time?

 

[RussianSensation]

Russian, you keep aiming too low with the potential performance of the upcoming AMD/Nvidia cards. If the GM107 is anything to go by (and it should be), Nvidia should be able to deliver 65-70% more graphics performance with the same power consumption as gtx 780 ti.

Either way, if we get >50% gtx 780 TI for $549 I'd be pleasantly surprised. More than likely, I'm guessing if we see that kind of jump, $599 is the more realistic starting price.

It's because looking at GM107 vs. GK107, NV increased die size 25% in order to get 2x the performance/watt. Since GK110 is already 561mm2, there is no room to increase die size another 25% on 28nm. As a result, I estimated that the performance/watt will not scale 2x over 780Ti. That's why I went with a more conservative 50-60% increase assuming GM210 is on 28nm. Of course, NV has been able to work miracles when moving from Fermi to Kepler (580--> 780Ti) so I wouldn't count 65-75% out of the question either.

Regarding the leaked chip, the sizing looks bigger than GK104 but smaller than GK110, which to me implies a GM204 successor is the most likely option.

VC updated the article and confirmed that ComputerBase says the leaked chip is a GM204.

It would be a welcome surprise if NV went with 8GB right out of the gate. 3GB is now mid-range territory and can be found on a $200 R9 280. For 4K gaming, 6-8GB is probably necessary to account for grossly inefficient next generation ports like Watch Dogs hehehe. I think NV might offer 4GB as standard and 8GB for more money though given how they've been able to charge a premium between 780 3 vs. 6GB and 770 2 vs. 4gb. NV has been VRAM stingy during 570/580 era too. Maybe Maxwell fixes this.

 

[blackened23]

If the die size estimates are accurate then it probably wont be more than 5-10% faster than 780ti even if maxwell is 30-40% more efficient than kepler. Which sounds exactly like what the early leaks said, 10% faster than 780ti and cheaper.

Pretty sure it's going to be a lot more than 10% faster than the 780ti. There's no point to a new 880 unless it's quite a bit faster. Every x80 card in the past was significantly faster than the prior gen, or else there's no point to a new family of cards.

Then again, everyone downplayed the performance of the GTX 680 prior to launch until it actually launched and turned out to be the fastest single GPU and was quite a bit faster than the 580, and faster than the 7970 despite the 7970 having better paper specifications.

I'd bet money that it's at least 30% faster than the 780ti, probably more. Else there's no freakin' point to a new generation of cards. I'm sure it will be more than 10% in any case.

 

[RussianSensation]

With GM107, NV increased CUDA core count to 640 (+66.7%) and die size went up from 118 to 148 (+25%), or CUDA core per mm2 increased almost 33%. If we take GK104's CUDA core count 1536/294mm2 die size x 430mm2 estimate for leaked chip x1.33 CUDA efficiency increase per mm2, we get almost 3,000 CUDA cores. But Nv increased L2 cache and made other tweaks which makes 750Ti's CUDA core efficiency per mm2 lower than GM204's since more of a smaller 148mm2 die will be devoted to L2 cache, etc. I think a 430mm2 Maxwell could even fit 3200 CUDA cores.

Add 35% increase in IPC for Maxwell, and this chip should beat 780TI by 30%, even 40%. Of course I am just guessing in 3200 cores and we don't know if and how much will the 256-bit bus impact high Rez performance for Maxwell's 204 chip.

 

[tviceman]

With GM107, NV increased CUDA core count to 640 (+66.7%) and die size went up from 118 to 148 (+25%), or CUDA core per mm2 increased almost 33%. If we take GK104's CUDA core count 1536/294mm2 die size x 430mm2 estimate for leaked chip x1.33 CUDA efficiency increase per mm2, we get almost 3,000 CUDA cores. But Nv increased L2 cache and made other tweaks which makes 750Ti's CUDA core efficiency per mm2 lower than GM204's since more of a smaller 148mm2 die will be devoted to L2 cache, etc. I think a 430mm2 Maxwell could even fit 3200 CUDA cores.

Add 35% increase in IPC for Maxwell, and this chip should beat 780TI by 30%, even 40%. Of course I am just guessing in 3200 cores and we don't know if and how much will the 256-bit bus impact high Rez performance for Maxwell's 204 chip.

Exactly, it's just so hard to make guesses right now. How much the L2 cache alleviates ROP and bandwidth bottlenecks is a big question mark. And how much (if any???) L2 cache is scaled up in size on larger Maxwell cores is a big question as well.

For reference, GF114 was exactly 4x the cuda cores of GF117 and GK104 was the same 4x cores of GK107. In a similar comparison, GK104 (gtx770) is exactly 3x faster than GK107 (gtx650). If GM204 is 3x faster than GM107, that would only be about 15% faster than gtx780 TI, but at GK104 power consumption. While the efficiency is fantastic, the performance improvement over GK110 would be dismissive. This new die will need to be able to cope with a higher TDP in order to realize a very sizable gain over GTX780 TI.

 

[n0x1ous]

I think the key for this card is going to be if it overclocks like GM107. If its stock 15% faster than 780ti but has the huge headroom like GM107 it will be fantastic

But....either way im waiting for big daddy maxwell

 

[Grooveriding]

I think the key for this card is going to be if it overclocks like GM107. If its stock 15% faster than 780ti but has the huge headroom like GM107 it will be fantastic

But....either way im waiting for big daddy maxwell

Yup. This is my expectation, about 25% faster than 780ti. GTX 680 was a new architecture and a new process, but was only 25% faster than a 580 was. This is a bigger die than GK104 and Maxwell with the same process.

If by some miracle it was 50% faster than a 780ti I'd get two, but my expectation is about 25% and a price close to where a 780ti is, in which case it's wait for the real deal. The other question is how cut down will the launch chips be ? What if it has 3 or 4 clusters disabled and a plan to milk the chip for a year re-releasing it with another cluster 2 or 3 times.

From now on when buying nvidia I am getting the big die and the whole chip or close to it. Otherwise you'll get milked paying flagship prices for midrange cards or getting heavily cut down big dies. I'd also expect if this card were able to be 50% faster than a ti that we will get it as a Titan first. 150% of a ti I would probably get them anyways unfortunately, but my bet is on 125% and the price of the current ti.

 

[Cloudfire777]

I think I found traces of both notebook and desktop GM204 test boards being shipped.

 

[f1sherman]

I think I found traces of both notebook and desktop GM204 test boards being shipped.

HA! Nvidia thought they can ship it across the sea and we won't notice... riiight

 

[dangerman1337]

Wonder if the GM204 will be the absolute highest numbered 800 series GeForce GPU and GM200 will be professional/Titan only. Along with Eyrines at Beyond3D, apparently Charlie wrote in Semi-Accurate (only an article reply says this) that the GM200 will be on 28nm: https://semiaccurate.com/forums/showpost.php?p=216233&postcount=4, http://semiaccurate.com/2014/07/01/nvidia-updates-gm204-maxwell-roadmap/ (subscription)

Is there a possible chance that the GM204 will have a 384-bit or 512-bit memory bit bus? The 8GB configuration suggets 256-bit bus or maybe even 512-bit bus but I am not sure if that 8GB is the highest theortical number that can be put on but offically they'll be 3 and 6GB versions. I also think Nvidia could release the GM204 sooner (mid to late september?) if AMD tonga is a surprise.