nVIDIA Maxwell

[Magic Carpet]

Minus all the power savings and what not. Is it going to be a performance leap (Voodoo Graphics style) or just another incremental upgrade?

What are your expectations?

I would like to hear your opinions. Thank you

 

[parvadomus]

It will be worse than going from fermi to kepler.

 

[Pottuvoi]

It should be similar to what we saw when jumping from geforce 8 based series (including 2xx) to Fermi, it is true architectural leap after all. (first in ~4-5 years.)

 

[brandon888]

worse then from fermi to kepler ? are you serious ? i mean .. soon people will play on 4K monitors and even 2 Titans will be weak for that ... come one .... and consoles are ready to launch to ! no more 30% performance jump please ! give at least 50% faster 880 then 780 !


Pottuvoi


hope that will happen .... 295 vs gtx 480 .... 480 has beaten or was similar to dual 295 GPU so it was like 60-70% faster then 285 !

 

[boxleitnerb]

As far as architecture goes, I heard that Maxwell will bring greater changes vs Kepler than Kepler did vs Fermi. Obviously that is true since Kepler is just Fermi refined and trimmed for efficiency.

 

[Sable]

I think it will be 8 better than Kepler.

 

[brandon888]

I think it will be 8 better than Kepler.

8 better ? do you mean 8X better performance ? no chance


Kepler has weak memory bandwidth but great performance with Texture filtrate cause of much Cuda cores .... hope Maxwell won't have any weakness and will be balanced with Pixel , Texture and bandwidth performance ...



I expect like 50-60% performance over 780 with 880 .... and hope midrage GTX 860 will be at least as fast as GTX 780 .... and that will be my next upgrade from GTX 660 ^_^

 

[f1sherman]

It will be worse than going from fermi to kepler.

so you mean bigger die, higher TDP, and only ~50-60% like 6970->7970GHz,
as opposed to lower TDP/smaller die and 80-100% like GTX 580 -> Titan / GTX 560 Ti -> GTX 680?

I doubt it because Nvidia is all power-consumption these days, and ultimately it's TDP that will be performance limiting factor.

80% across the board is what they've always been able to deliver. I see no reason why this would change

 

[notty22]

I'm going with this:

NVIDIA 20nm Maxwell GPUs would feature more than double the performance per watt over current generation Kepler architecture

 

[Elfear]

so you mean bigger die, higher TDP, and only ~50-60% like 6970->7970GHz,
as opposed to lower TDP/smaller die and 80-100% like GTX 580 -> Titan / GTX 560 Ti -> GTX 680?

I doubt it because Nvidia is all power-consumption these days, and ultimately it's TDP that will be performance limiting factor.

80% across the board is what they've always been able to deliver. I see no reason why this would change

It was 2.5yrs from 580 to Titan so does that mean Maxwell will be ready Q3/Q4 of 2015?

 

[boxleitnerb]

^ I think we have to get used to getting the bigger GPUs late during the lifetime of a process node. This "hiccup" with Titan being so late only happens once. Now Nvidia has changed their approach to releasing the small GPU first and the large one later and probably will stick with that for the foreseeable future. If TSMC manages a 2 year cadence with the shrinks, I see no problem with a 2 year interval between GK104 and GM104 or GK110 and GM110 respectively.

 

[Elfear]

^ I think we have get used to getting the bigger GPUs late during the lifetime of a process node. This "hiccup" with Titan being so late only happens once. Now Nvidia has changed their approach to releasing the small GPU first and the large one later and probably will stick with that for the foreseeable future. If TSMC manages a 2 year cadence with the shrinks, I see no problem with a 2 year interval between GK104 and GM104 or GK110 and GM110 respectively.

I agree. I was just trying to point out the flaw in his logic. If Maxwell is due out next year, I doubt it will be Titan + 80%. More than likely it will be 680 + 80%. It would be great if I'm wrong as moar powuh is always better but the performance leaps we used to get have slowed in the last couple years.

 

[brandon888]

I agree. I was just trying to point out the flaw in his logic. If Maxwell is due out next year, I doubt it will be Titan + 80%. More than likely it will be 680 + 80%. It would be great if I'm wrong as moar powuh is always better but the performance leaps we used to get have slowed in the last couple years.

Titan + 80% funny .... but why 780 + 50-60% is unreal ? it's like 40-45% more then titan ....

But if we look at the picture .... Maxwell will use 50% power then Kepler ... but no one said it will be 2X Faster ....

 

[Zanovar]

It will be worse than going from fermi to kepler.

Probably my next card,jesus chear me up

 

[boxleitnerb]

Judging by GTX580 vs GTX680, I would put GM104 (GTX 880?) at Titan +40%. Depends on how high Titan boosts, of course. But in that ballpark.

 

[tviceman]

I compiled the average increase in performance this generation vs. the last generation of chips here: http://alienbabeltech.com/abt/viewtopic.php?f=6&t=30552#p89541 using anandtech and techpowerup for my performance comparisons.

The 680 (GK104) is 90% faster than GF114 (gtx560 ti) at 1600p and 79% faster at 1200p.
The 770 (GK104) is 104% faster than GF114 at 1600p and 85% faster at 1200p.

GK104's jump over GF114 is quite extraordinary (even more amazingly, GK106 has an even larger performance jump over GF116). I don't know if Nvidia can increase Maxwell's performance over Kepler to this degree, but if they do then GM104's best first-run sku will be to Titan what gtx680 was to gtx580 (25-35% faster).

 

[BallaTheFeared]

Maxwell is suppose to be a bigger leap than Kepler was over Fermi, but I'll believe that when I see it.

 

[Sable]

8 better ? do you mean 8X better performance ? no chance


Kepler has weak memory bandwidth but great performance with Texture filtrate cause of much Cuda cores .... hope Maxwell won't have any weakness and will be balanced with Pixel , Texture and bandwidth performance ...



I expect like 50-60% performance over 780 with 880 .... and hope midrage GTX 860 will be at least as fast as GTX 780 .... and that will be my next upgrade from GTX 660 ^_^

Nope, just 8 better. It's a pure speculation thread where we have zero facts so that's my completely meaningless arbitrary number of better.

 

[Erenhardt]

Nope, just 8 better. It's a pure speculation thread where we have zero facts so that's my completely meaningless arbitrary number of better.

:thumbsup:

100% this

 

[f1sherman]

this has to be reposted:

I was glazing over Titan performance scores and noticed an interesting trend - Nvidia had a much bigger perf/mm^2 jump refining Fermi into Kepler than AMD did with Cayman and barts moving to GCN.

GK110 is about 5% larger than GF110, yet it's performance is 85% faster than GF110 according to techpowerup (at 2560x1600) and nearly 100% faster according to anandtech's review.
GK104 (as the gtx680) is about 20% smaller than GF114 (in reverse, GF114 is about 23% larger), yet Nvidia managed to increase perfomance by an average of 89% at 1600p and 79% at 1920x1200 (vs gtx 560 ti).
GK104 (as the gtx770) is about 20% smaller than GF114 (in reverse, GF114 is about 23% larger), yet Nvidia managed to increase perfomance by an average of 104% at 1600p and 85% at 1920x1200 (vs gtx 560 ti).
GK106 is about 10% smaller than GF116 (GF116 is about 11% larger), and Nvidia manage to increase performance by an average of 145% at 1600p and 121% at 1920x1200 (vs gtx550 ti).
GK107 is virtually the same size as GF118. Nvidia increased performance by 108% at 1920x1200 and 1680x1050.

Tahiti is 6% smaller than Cayman (Cayman is 6.5% larger). It's performance is 62% higher than Cayman at 1600p and 56% at 1200p.
Pitcairn is 17% smaller than Barts (Barts is 20% larger). It's performance is 46% higher than Barts at 1600p and 41% higher at 1200p.
Bonaire is 4% smaller than Juniper (Juniper is 4% larger). It's performance is 93% higher at 1200p and 100% higher than Juniper at 1050p.
Cape Verde is virtually the same size as Turks (~5mm difference). It's performance is 81% higher at 1200p and 82% at 1050p.

• Looking over the data, the trend has been to decrease die sizes (the exception being GK110).
• More noticeably, Nvidia increased performance per mm^2 across the board with every chip more so than AMD did by a substantial amount.
• While Pitcairn and GK106 are about equal across the board with respect to performance and power consumption, GK104 offers substantially more performance per mm^2 than Tahiti.
• AMD brought several more chips (8 completely different chips) to market than Nvidia (5 completely different chips) at 40nm. On 28nm, Nvidia has made 5 kepler derivatives while AMD has made 4.
• Considered to be GCN 1.1, Bonaire is significantly faster and more efficient than it's predecessor (Juniper) vs. any of the GCN 1.0 chips when compared to their 40nm predecessors.

It is very interesting data to examine. Despite AMD having the architectural advantage on 40nm with regards to power envelope and perf/mm^2, they used substantially more resources in making a large variety of chips, and simultaenously failed to recapture the fastest single performing GPU once Nvidia's Fermi was released. In other words, AMD had more chips covering less of a performance span at 40nm than Nvidia did. On 28nm up to this point, AMD has made 4 different chips at 28nm whereas Nvidia has made 5 different chips (although only 4 are sold to consumers). Also at 28nm, Nvidia has currently managed to completely close the gap and reverse in all measurable categories (perf/mm^2, die sizes, perf/watt). Furthermore, Nvidia's efforts aren't finished yet either as "refreshed" parts are have been released and more are coming.

No one knows what is going to happen at 20nm, but AMD has said future architectures will be based off their GCN design they are working with now (much like how Kepler is based off Fermi). Nvidia's CEO JHH has come out and said that "Maxwell will crush Kepler" (exact quote) in respects to performance per watt (which should be expected given a new node). A wildcard factor in AMD's favor is their intimate development of GDDR6 memory, which could lead to improved stability at higher vram speeds for AMD products (witnessed with Nvidia's Fermi designs with GDDR5 vs. AMD's at the time).

Sources:
http://en.wikipedia.org/wiki/Geforce_400_Series
http://en.wikipedia.org/wiki/GeForce_500_Series
http://en.wikipedia.org/wiki/Radeon_HD_5000_Series
http://en.wikipedia.org/wiki/Radeon_HD_6000_Series
http://en.wikipedia.org/wiki/Radeon_HD_7000_Series
http://www.techpowerup.com/reviews/NVIDIA/GeForce_GTX_Titan/27.html
http://www.anandtech.com/show/6774/nvidias-geforce-gtx-titan-part-2-titans-performance-unveiled

 

[SirPauly]

It's too early to gauge but if there are similar performance gains as past architectural generations -- would be pleased over-all.

What I subjectively like more about the GPU is how art and technology collide and would enjoy to break down their specific abilities, strengths and weaknesses, pros and cons, flexibility and software tools, understand the vision for their architecture.

Must be an incredible, daunting task to create these marvels of engineering.

 

[Magic Carpet]

Maxwell is suppose to be a bigger leap than Kepler was over Fermi, but I'll believe that when I see it.

My gut feeling is telling, that nVIDIA is going to follow the market and optimize Maxwell solely for power-efficiency and thus, making low-power designs their number 1 priority. I would like to be wrong, though.

 

[boxleitnerb]

I compiled the average increase in performance this generation vs. the last generation of chips here: http://alienbabeltech.com/abt/viewtopic.php?f=6&t=30552#p89541 using anandtech and techpowerup for my performance comparisons.

The 680 (GK104) is 90% faster than GF114 (gtx560 ti) at 1600p and 79% faster at 1200p.
The 770 (GK104) is 104% faster than GF114 at 1600p and 85% faster at 1200p.

GK104's jump over GF114 is quite extraordinary (even more amazingly, GK106 has an even larger performance jump over GF116). I don't know if Nvidia can increase Maxwell's performance over Kepler to this degree, but if they do then GM104's best first-run sku will be to Titan what gtx680 was to gtx580 (25-35% faster).

Both, the 680 and 770 use more power than a 560 Ti, though.

 

[SirPauly]

My gut feeling is telling, that nVIDIA is going to follow the market and optimize Maxwell solely for power-efficiency and thus, making low-power designs their number 1 priority. I would like to be wrong, though.

Performance per watt efficiency, imho!

Would be interesting to see how nVidia handles unified virtual memory and how this may benefit heterogeneous applications for the PC.

 

[brandon888]

The Maxwell architecture, the successor to Kepler, will for the first time feature an integrated ARM CPU of its own;[1] This will make Maxwell GPUs more independent from the main CPU according to Nvidia's CEO Jen Hsun Huang.[2]


That will work like inter virtu ? D: i mean add some extra performance to current cpus ? like 10-15% or even less ?