Graphics card rumor roundup

[tviceman]

Tviceman, I think your estimate that flagship Maxwell will only be 25-30% faster than 980 is wayyy too conservative. NV has a lot of headroom to play with 250-275W power usage that 780Ti hit. There is no reason they couldn't theoretically push GM200 to that power usage. By the time this card launches, 28nm will be even more mature. GM200 may be Maxwell 3.0 with slightly higher IPC (v1 had 35%, v2 is 40% over Kepler). I think it will be around 50% faster than a 980, maybe more. 980 reference uses less power than 680 reference so that means Big Maxwell has at least the same headroom from 980 as 780Ti had over 680!

You need to readjust your expectations. GM200 taped out a month (or so) after GM204. There is no Maxwell 3.0 on 28nm. GM200 is going to get all of GM204's improvements + the extra compute capabilities. There won't be any new tricks. Also, GM204 averages less power than GK104, but peaks higher, all in all they're both very similar in power use.

1. GK110 was limited by clock headroom with respect to it's TDP. GM200 won't be limited by clock headroom, instead it will be limited by die space. GK110 had 87.5% more functional cores than GK104, but there is no way GM200 can come close to that much more on-die processing powre than GM204.

2. GM204 was able to outperform GK104 (50% faster than 770, 58% faster than 680) but came with a 35% larger die size. GM200 doesn't have the luxury of getting significantly more die size than GK110 so it will have to rely on an increased clock speed to realize it's advantage. I think GM200 is going to have 20 SMM's and a 384-bit bus / 96 ROP's, and similar clock speeds as GM204. If my predictions are right, that means 25% more shader performance, 50% more throughput, and up to 50% more bandwidth (Nvidia may go with 6ghz though as 7ghz maybe overkill given Maxwell's bandwidth saving characteristics).

I think GM200 would pull away in cases of high AA / 4k use, but otherwise I think 30% is spot on.

 

[KaRLiToS]

 

[Wild Thing]

Tonga offered NOTHING over Tahiti unless you are frothing for Mantle and true audio. GM204 offered EVERYTHING over GK104 and also took out the bigger, more power hungry chips (GK110, Hawaii) being built along side it. But I get it. You don't care. You're not impressed. You want a $250 Titan Black killer. You want everyone to ignore reality. I get it. Keep waiting. Let me know how impressed you are when Fiji comes on 20nm and can only beat the craptastic GM204 by 20%.

 

[KaRLiToS]

Haters gonna hate.

 

[werepossum]

Russian you are throwing everything regarding today's reality out of the window. GPU's double in performance because NODES. We are still on the same node. GM204 is built on the same exact node as Tahiti, and obliterates it. Destroys it. In every way possible. GM204 is built on the same node as GK110 and Hawaii, and destroy those cards in all metrics. No question, no contest. And it's on the same node.

Everyone is stuck on 28nm. We aren't getting double performance on the same node. It isn't happening. It's not Nvidia's fault they can't defy the laws of physics. Nvidia is doing more given their situation than AMD. Case in point, there was a shorter time frame between Tahiti and Tonga than there was between GK104 and GM204. Tonga offered NOTHING over Tahiti unless you are frothing for Mantle and true audio. GM204 offered EVERYTHING over GK104 and also took out the bigger, more power hungry chips (GK110, Hawaii) being built along side it. But I get it. You don't care. You're not impressed. You want a $250 Titan Black killer. You want everyone to ignore reality. I get it. Keep waiting. Let me know how impressed you are when Fiji comes on 20nm and can only beat the craptastic GM204 by 20%.

I tend to agree. AMD came in with better performance at an unheard-of price almost across the board, forcing nvidia to drop prices trying to compete. Now nvidia is doing the same thing to AMD at the top and near-top performance tiers, which drives down prices through the mid-range. To those of us who aren't interested in "investing" the cost of a decent used car into video cards but are concerned with getting top notch performance at a price we won't be embarrassed for people to know we spent, it's been a great year.

 

[Silverforce11]

GM200 won't be limited by clock headroom, instead it will be limited by die space.

Russian is correct to say if NV wanted to push GM200 with higher clocks (because they can) to use up the TDP headroom (300W if they wanted to), it would be a major perf improvement over the 980.

It's not just die space, but as long as the process allows for higher clocks (with higher vcore), then pumping clock speed to utilize the TDP available will allow NV to have a 50% leap in general performance easily.

Imagine a large die GM200 with extra vcore boosting to 1.6ghz at stock settings.

Efficiency is a winner in all aspects as long as it allows tuning for performance at the expense of efficiency. This is where AMD will suffer if they do not improve efficiency on GCN2.0, because NV has plenty of room in terms of untapped power to utilize for raw performance.

 

[raghu78]

Russian you are throwing everything regarding today's reality out of the window. GPU's double in performance because NODES. We are still on the same node. GM204 is built on the same exact node as Tahiti, and obliterates it. Destroys it. In every way possible. GM204 is built on the same node as GK110 and Hawaii, and destroy those cards in all metrics. No question, no contest. And it's on the same node.

GM204 arrives 33 months after Tahiti on a mature 28nm process. GM204 and GM200 will have to be benchmarked against R9 390X/R9 390/ R9 380X and R9 380. So lets not get too ahead of ourselves. Nvidia does have a significant time to market advantage which they are milking for all its worth like when AMD brought HD 5870 to market six months before GTX 480.

You want everyone to ignore reality. I get it. Keep waiting. Let me know how impressed you are when Fiji comes on 20nm and can only beat the craptastic GM204 by 20%.

How do you know Fiji is on 20nm. All indications are 28nm massive die with HBM. btw you are making a performance prediction without knowing anything about what architectural changes AMD has made to R9 3xx series.

 

[raghu78]

Russian is correct to say if NV wanted to push GM200 with higher clocks (because they can) to use up the TDP headroom (300W if they wanted to), it would be a major perf improvement over the 980.

gm200 will be a significant performance upgrade but do not expect perfect 50% improvement. Nothing scales perfectly.

It's not just die space, but as long as the process allows for higher clocks (with higher vcore), then pumping clock speed to utilize the TDP available will allow NV to have a 50% leap in general performance easily.

There is never a case where we see a perfect 1:1 scaling in perf with shader count. Nvidia will not push for higher than 250W TDP unless they are forced to due to competition (which is unlikely to happen).

Imagine a large die GM200 with extra vcore boosting to 1.6ghz at stock settings.

Do not expect GM200 to clock on par with GM204 both out of the box and max overclocked. The bigger the die the more difficult it becomes to keep pushing voltage and expects clocks to keep scaling without running into thermal problems. Water cooling will be needed for people who want to hit 1.6 Ghz with voltage overclocking.

 

[Silverforce11]

Do not expect GM200 to clock on par with GM204 both out of the box and max overclocked. The bigger the die the more difficult it becomes to keep pushing voltage and expects clocks to keep scaling without running into thermal problems. Water cooling will be needed for people who want to hit 1.6 Ghz with voltage overclocking.

GK110 scaled very well compared to GK104. Did GK104 clock better than GK110? Nope.

If anything, bios unlocked and vcore modded 780ti clocked very high.

How well it clocks isn't limited by die size but by the transistor of that particular node as well as density. If NV keeps density similar with GM200, on the same 28nm node from TSMC, theres no reason why it cannot clock as well or better due to maturity of the node.

But how high they push it will depend on where the competition is at. Certainly they have plenty of power left to push clocks high.

 

[raghu78]

GK110 scaled very well compared to GK104. Did GK104 clock better than GK110? Nope.

GTX 770 consistently clocked higher than GTX 780 Ti in reference and factory overclocked versions. The 780 Ti classifieds clocked higher than other 780 Ti cards though they had a custom BIOS and higher voltages. Moreover to hit 1400+ Mhz on GTX 770 on air was easier than GTX 780 Ti.

How well it clocks isn't limited by die size but by the transistor of that particular node as well as density. If NV keeps density similar with GM200, on the same 28nm node from TSMC, theres no reason why it cannot clock as well or better due to maturity of the node.

Keeping density similar means Nvidia would be hitting close to 650+ sq mm. GM200 is rumoured to be 3072 cc, 384 bit memory and will incorporate hardware for Quadro / Tesla specific features and full double precision performance. I am guessing Nvidia will try and increase density for a slight tradeoff in max clocks.

http://hexus.net/tech/reviews/graphics/55857-nvidia-geforce-gtx-770/

GK104 - 3540 million transistors at 294 sq mm - 12.04 million transistors / sq mm
GK110 - 7100 million transistors at 551 sq mm - 12.88 million transistors / sq mm

 

[SunburstLP]

Russian you are throwing everything regarding today's reality out of the window. GPU's double in performance because NODES. We are still on the same node. GM204 is built on the same exact node as Tahiti, and obliterates it. Destroys it. In every way possible. GM204 is built on the same node as GK110 and Hawaii, and destroy those cards in all metrics. No question, no contest. And it's on the same node.

Everyone is stuck on 28nm. We aren't getting double performance on the same node. It isn't happening. It's not Nvidia's fault they can't defy the laws of physics. Nvidia is doing more given their situation than AMD. Case in point, there was a shorter time frame between Tahiti and Tonga than there was between GK104 and GM204. Tonga offered NOTHING over Tahiti unless you are frothing for Mantle and true audio. GM204 offered EVERYTHING over GK104 and also took out the bigger, more power hungry chips (GK110, Hawaii) being built along side it. But I get it. You don't care. You're not impressed. You want a $250 Titan Black killer. You want everyone to ignore reality. I get it. Keep waiting. Let me know how impressed you are when Fiji comes on 20nm and can only beat the craptastic GM204 by 20%.

There are some really good points being raised here and I'd like to say thanks for everyone for keeping everything pretty civil so far.

I'd like to point out that being stuck on a node for so long, as frustrating as it is for us enthusiasts, has a real silver lining. Other members are far more historically inclined than I am and could possibly double check my thought, but have we ever seen so much progress on a single node performance wise? Isn't it a bit of a blessing in disguise to force both NV and AMD to think so hard about making decisions based on finesse and experience? Rather than going the historic route of counting on the node migrations to bring tons of performance gains in seemingly generational jumps they have to become better at providing true innovation in their designs. It both sucks and has to be admired at the same time, IMO.

TL;DR - NV and AMD realized that they have more tools than just a hammer and this problem wasn't a nail.

 

[tviceman]

My whole point was GM204 is disappointing in name only. Its an extremely impressive feat of engineering that is likely a first in all metrics while being stuck on an aging manufacturing process. Neither AMD nor Nvidia can work magic and until better nodes are viable we are stuck wishing and wanting.