Kitguru : Nvidia to release three GeForce GTX 800 graphics cards this October

[bakalu]

sorry my english.

i think that maxwell 880 can't have 2560 SP.

gtx 750 ti have 640 SP and 1.87B transistors.

if maxwell 880 has 2560 SP, so maxwell 880 must have 7.48B transistors. It is too much transistors for 28nm.

 

[tollingalong]

This thread is full of "real" questions.

We have one fact: a rumour states that Nvidia will release something during October.

Therefore, there is only one real question: Is the rumour true?

This isn't a rumor. It's been confirmed via interviews.

 

[AdamK47]

How about this:

This 880GTX won't be the replacement for the 780ti, but more something between 780 non-ti and 770. (in terms of position, not performance !)

A future GeForce Titan (Maxwell v3?) would be the fastes card and even a possible 880ti would stay behind it. The Titan would also have special coolers to keep it's "luxury" position.

This would shift the prices to the lower end and in this scenario, a 880GTX for 450$ makes total sense!

Source? My creative mind

I'm thinking GTX 880 will indeed perform better than the 780 Ti.

 

[tollingalong]

Aren't all that much faster. This is not accurate. The GK107 configuration that matches the GM107 used in the 750ti is the GT640, neither of these cards have a power plug requirement. GT640 = 65W TDP, GTX 750ti = 60W TDP. The GM107 in the 750ti is three times faster at a lower TDP than the GT640. Furthermore, the GTX 650 has a power plug requirement for the reference baseline model (750ti doesn't for the reference baseline) and the GM107 is two times faster than the beefed up GK107.

The GT640 is a 65W TDP. This is the Kepler version of the GTX 750ti. Again...the 750ti is THREE TIMES FASTER.

The GTX 650 is a scaled up version of the GK107 with a 110W TDP. The 750ti is two times faster than the scaled up GK107. This is double the TDP of the GM107 in the 750ti. This is, again, not an apples to apples comparison because the GTX 650 is beefed up with a power plug requirement. The GM107 is neither.

If your definition of not being faster is actually three times faster, well, I guess i'll throw that out there. The performance per watt doubled from kepler to maxwell, and that will certainly manifest itself in big performance gains at the same TDP levels. 225W TDP Maxwell will outperform a 225W TDP Kepler. How much so remains to be seen. If I had to guess i'd guesstimate 15-30% faster just depending. Really, there's no point to an 800 series if it were the same speed as the prior generation. At least that's my line of thinking.

Not a good comparison at all. 750 Ti's memory clock is over 200% greater to start and that's with the better cache of Maxwell.

Ignore power for the time being. We both agree Maxwell is more power efficient.

765M has 768 cores. 768 @ 850MHz 128 Bit bus with 4000MHz
850M has 640 cores. 640 @ 875MHz 128 Bit bus with 5000MHz

The Maxwell memory is about 10-15% faster than the last generation even with a 20% increase in memory clock but 15% less cores. I know cores aren't equal but my point stands.

Gen 2 Maxwell might be better but as of right now Maxwell's not showing a whole heck of a lot more more performance.

 

[dacostafilipe]

I'm thinking GTX 880 will indeed perform better than the 780 Ti.

I was not talking about the performance against the 780ti but more about the performance inside the new 800 Series line. For example:

Titan -10%-> 880ti -15%-> 880 -20%> 870 ...

Today it's more like:

780ti -8%-> Titan -7%-> 780 -15%-> 770 ...

 

[witeken]

Will maxwell 880 have integrated arm cores ?

Yes.

 

[Mand]

What's the benefit of having integrated ARM cores?

 

[tviceman]

Yes.

You are spouting out hopeful rumors which no one really knows nothing about. It is entirely unknown, and likely made up, that GM204 will have ARM cores integrated.

What's the benefit of having integrated ARM cores?

No one knows.

 

[FatherMurphy]

Like tviceman said, the ARM-core-in-Maxwell rumor is really unsubstantiated. I think it was 2011 when Nvidia said there would be ARM cores in its Maxwell architecture, but the road map has changed a lot since then. Indeed, at that time, Maxwell was intended to be on 22/20nm, but the rumors now suggest otherwise. Also, the ARM core that Nvidia said it was going to integrate into Maxwell was the Denver core, which hasn't made an appearance even in Tegra yet. EDIT: Quick Google search: http://wccftech.com/nvidia-roadmap-confirms-20nm-maxwell-gpus-2014-kepler-refresh-arrives-1h-2013/

In all likelihood (just me speculating), the only Maxwell chip that will have an ARM Core (Denver) will be GM200/210, and that won't be out until next year on the 20 nm process.

 

[Ajay]

You are spouting out hopeful rumors which no one really knows nothing about. It is entirely unknown, and likely made up, that GM204 will have ARM cores integrated.



No one knows.

The purpose of 'Denver' being integrated into the GPU was so that didn't need a host CPU and could be run independently (it would run some small custom linux distro), TTBOMK. This would only be for Compute purposes, so would only be in Big M; not in NV's consumer grade GPUs.

I thought that it turned out Denver was a separate ARM project and wouldn't be included in Maxwell. I'm not sure about this, since Nvidia changed their roadmap (there is no Volta now, but Pascal instead and NV has obviously rethought things).

The real bummer is that HBM won't be used till Pascal (probably on 16FF+) - and again may only be used for NV's high-end compute AIBs due to cost (perhaps NV will kindly sell these HBM equipped cards for high-end enthusiast's). Sadly, I think Pascal is 2-3 years off.

*Edit* Oops, Father Murphy beat me to it, guess I didn't refresh my browser

 

[witeken]

Nvidia will disclose Denver's architecture at Hot Chips in 3 days. TK1 with Denver was also planned for H2, so it could be possible.

 

[Ajay]

Nvidia will disclose Denver's architecture at Hot Chips in 3 days. TK1 with Denver was also planned for H2, so it could be possible.

Thanks, I thought I read something about Denver last week. So the Denver/Maxwell link will probably ultimately point towards an ARM SoC w/Maxwell GFX SMM.

 

[RussianSensation]

If Maxwell is at least as good as Kepler in terms of generational leaps, we can use the jump from GTX560Ti to 680 as guidance:

GTX680 at launch was 90% faster than GTX560Ti.
http://www.computerbase.de/2012-03/test-nvidia-geforce-gtx-680/12/

GTX770 is about 5-6% faster than GTX680:
http://www.computerbase.de/2013-05/nvidia-geforce-gtx-770-test/4/

Adjusting for 680 on this chart, we get GTX770 = 100% * (0.95 coefficient for 680's performance) = 95% of 770. Let's be a bit conservative and assume instead of getting 2x the performance/watt over GK104, GM204 gets 90%. Applying 90% performance/watt boost from GK104 (680) to GM204 at 195W TDP we get:

95% (770 adjusted to 680's performance) x 1.90 Maxwell vs. Kepler perf/watt increase = 181% on this chart.

GTX780Ti relative standing to 770 is 135% on the same chart.

Extrapolated GM204 880 is 181% of GTX770. This leads to 181/135 = 34% faster than 780Ti assuming GTX880 has 195W TDP and 90% performance/watt increase over GTX680.

Conclusion: 90% performance/watt increase from Kepler GTX680 to Maxwell GTX880, an efficiency gain which is actually slightly less than NV's own 2x performance/watt claim, would produce a GTX880 that's 90% faster than GTX680 and in turn 34% faster than GTX780Ti. As a result, GTX880 would produce a similar jump over 780Ti as 680 achieved over 580.

Any estimates below this imply that NV is either going to have a TDP lower than 195W and/or Maxwell's performance/watt increase over Kepler is less than 90% which contradicts their own perf/watt estimates and 750Ti's performance/watt increases over Kepler's 60W chips.

The 430mm2 die is also sufficiently large enough over 294mm2 GTX680 to give NV the necessary room to enlarge L2 cache and the extra transistor count that Maxwell requires over Kepler to achieve the 2x performance/watt. So why are all of us so conservative all of a sudden and projecting 880 to be only 15-25% faster?

 

[FatherMurphy]

Conclusion: 90% performance/watt increase from Kepler GTX680 to Maxwell GTX880, an efficiency gain which is actually slightly less than NV's own 2x performance/watt claim, would produce a GTX880 that's 90% faster than GTX680 and in turn 34% faster than GTX780Ti. As a result, GTX880 would produce a similar jump over 780Ti as 680 achieved over 580.

All without utilizing a new lithography process. That would be amazing.

The 430mm2 die is also sufficiently large enough over 294mm2 GTX680 to give NV the necessary room to enlarge L2 cache and the extra transistor count that Maxwell requires over Kepler to achieve the 2x performance/watt. So why are all of us so conservative all of a sudden and projecting 880 to be only 15-25% faster?

If GM107 has 2 MB (correct me if I am wrong) of L2 cache, what do you suppose the cache size will be in GM204?

 

[blackened23]

Cache basically serves to eliminate memory bottlenecks. Not just for GPUs, but for CPUs and a wide variety of silicon. This is why cache exists. My theory is that the 780ti is slightly lower at 4k resolution (relative to it's 1080p-1600p performance) because Kepler only uses 128/256kb of cache (can't recall exactly) while Hawaii uses a whopping 1MB of cache. You should also consider that the 780ti actually has higher memory bandwidth than the 290X (780ti = 336GBs , Hawaii = 320GBs) but the Hawaii has greater levels of L2 cache.

So if I had to guess....Nv will use whatever amount of cache is necessary to make their card shine at all situations and all resolutions. What I don't think is that NV is stupid, they know the relative weaknesses and strengths of Kepler (4k wasn't the absolute strong point, although the 780ti was clearly the best card at all other resolutions) so i'd imagine that the 880 will have a fairly massive L2 cache. How much? Who knows. Whatever it takes is my guess. Even though nobody really cares about 4k at this point, but the point remains, i'm sure NV is aware of it.

Also, to add, there are about 25,000 factors that goes into the final performance of silicon. Anyone who focuses in on one number on a specification sheet such as memory bandwidth or L2 cache, and declares winners on that and that alone, really does not know what the hell they're talking about. Silicon is a pretty complex beast.

 

[FatherMurphy]

Agreed.... though GK110 has 1.5 MB of L2 cache. http://www.nvidia.com/content/PDF/kepler/NVIDIA-Kepler-GK110-Architecture-Whitepaper.pdf

I'm no expert, but I tend to think that the 4k resolution deficit in the 780ti is an ROP issue, not solely a memory bandwidth/cache issue.

My understanding is that the larger cache in Maxwell serves two purposes 1) lowers energy usage by lessening memory transfers between the chip, on-board memory, and system memory and 2) helps Maxwell compete with lower memory bandwidth.

I would think that GM204 is going to have a massive L2 Cache as well. If GM204 is 4x GM107, like GK104 was 4x GK107, then perhaps GM204 will have 8MB of L2 cache, which would be unprecedented.

 

[tviceman]

If Maxwell is at least as good as Kepler in terms of generational leaps, we can use the jump from GTX560Ti to 680 as guidance:

GTX680 at launch was 90% faster than GTX560Ti.
http://www.computerbase.de/2012-03/test-nvidia-geforce-gtx-680/12/

GTX770 is about 5-6% faster than GTX680:
http://www.computerbase.de/2013-05/nvidia-geforce-gtx-770-test/4/

Adjusting for 680 on this chart, we get GTX770 = 100% * (0.95 coefficient for 680's performance) = 95% of 770. Let's be a bit conservative and assume instead of getting 2x the performance/watt over GK104, GM204 gets 90%. Applying 90% performance/watt boost from GK104 (680) to GM204 at 195W TDP we get:

95% (770 adjusted to 680's performance) x 1.90 Maxwell vs. Kepler perf/watt increase = 181% on this chart.

GTX780Ti relative standing to 770 is 135% on the same chart.

Extrapolated GM204 880 is 181% of GTX770. This leads to 181/135 = 34% faster than 780Ti assuming GTX880 has 195W TDP and 90% performance/watt increase over GTX680.

Conclusion: 90% performance/watt increase from Kepler GTX680 to Maxwell GTX880, an efficiency gain which is actually slightly less than NV's own 2x performance/watt claim, would produce a GTX880 that's 90% faster than GTX680 and in turn 34% faster than GTX780Ti. As a result, GTX880 would produce a similar jump over 780Ti as 680 achieved over 580.

Any estimates below this imply that NV is either going to have a TDP lower than 195W and/or Maxwell's performance/watt increase over Kepler is less than 90% which contradicts their own perf/watt estimates and 750Ti's performance/watt increases over Kepler's 60W chips.

The 430mm2 die is also sufficiently large enough over 294mm2 GTX680 to give NV the necessary room to enlarge L2 cache and the extra transistor count that Maxwell requires over Kepler to achieve the 2x performance/watt. So why are all of us so conservative all of a sudden and projecting 880 to be only 15-25% faster?

I'm basing my projections off two things:

1) The performance difference GK104 has over GK107 (GTX 770 vs. GTX 650) - which is exactly a 3x performance improvement.
2) GM107's performance improvement over GK107 (gtx 750 TI vs. GTX 650) - which is 70-75%.

When extrapolating both of those factors to guesstimate GM204's performance, it comes out to be 15-20% faster. Maxwell on 28nm isn't going to be the same leap in performance that Kepler on 28nm was over Fermi on 40nm. Had Maxwell been on 20nm, it likely would have been a bigger leap over Kepler than Kepler did over Fermi.

Three unknown variables in this situation though could skew my analysis.

1) If GM200/210 does NOT exist on 28nm, then GM204 is intended to be Nvidia's graphics flagship for the next year (at least), in which case it could very well have been engineered from the start to have a higher TDP than GK104 but with the same efficiency profile as GM107. This would allow a greater performance jump than what was demonstrated with GM107 over GK107.
2) Nvidia's labeling of GM107 as being "first generation maxwell." Does first generation mean 28m? Does it mean GM107? Does it not mean anything of significance?

I personally think GM200 is going to be released on 28nm and it'll be 50-60% faster than GK110. I don't think Nvidia wants to wait until finfets with a new flagship compute chip because Knights Landing is coming next year and not having a solid answer to that in a reasonable time frame would be potentially very disruptive to Nvidia's stranglehold on the HPC market. Therefore, I think given that GM204 will slot in 15-20% faster than GK110, allowing GM200/210 to exist as 2-3 Geforce skus (880 TI, 885, Titan M).

 

[blackened23]

Agreed.... though GK110 has 1.5 MB of L2 cache. http://www.nvidia.com/content/PDF/kepler/NVIDIA-Kepler-GK110-Architecture-Whitepaper.pdf

I'm no expert, but I tend to think that the 4k resolution deficit in the 780ti is an ROP issue, not solely a memory bandwidth/cache issue.

My understanding is that the larger cache in Maxwell serves two purposes 1) lowers energy usage by lessening memory transfers between the chip, on-board memory, and system memory and 2) helps Maxwell compete with lower memory bandwidth.

I would think that GM204 is going to have a massive L2 Cache as well. If GM204 is 4x GM107, like GK104 was 4x GK107, then perhaps GM204 will have 8MB of L2 cache, which would be unprecedented.

Hmm. 1.5MB, I was not aware of that. (and I obviously didn't double check, heh) Thanks for the correction. Maybe GK104 was 256k? I didn't look it up, just going by rough memory. It could very well be a ROP issue. But, again, it's a very complex situations and is kinda like asking questions with no answers. There's so much going on at the silicon level it's really hard to say. But I do think NV will do what it takes to make it strong in all of those situations. This also doesn't change that focusing in on any one parameter on a specification sheet to declare a winner (such as bus width or L2 cache) is stupid. I mean, from my memory the 680 maintained a 30%+ lead over the 580 at all resolutions, surround or not depsite the lower bus.
And then you had the ATI HD2900 series which had a 512 bit bus but struggled in terms of performance despite the bus width.

This is hardly scientific. Just goes to show you can't judge anything from a spec sheet, it's more complex and probably involves a lot of things that the majority of us (including me) won't understand at a transistor level. That said, you could very well be correct on the ROP issue. We'll see I guess.

 

[tviceman]

My theory is that the 780ti is slightly lower at 4k resolution (relative to it's 1080p-1600p performance) because Kepler only uses 128/256kb of cache

I'm no expert, but I tend to think that the 4k resolution deficit in the 780ti is an ROP issue, not solely a memory bandwidth/cache issue.

I am no expert either, but I think pinning the issue entirely on ROP's isn't accurate, and I don't think memory bandwidth and/or cache is entirely accurate either. I think it's tied more to fillrate (which is tied to ROP's and clock speed), but a Kepler ROP does not equal a GCN ROP, nor does it equal a Maxwell ROP.

GM204 could have beefed up ROP's over Kepler (although I concede that G107's ROP's appear to be very similar to Kepler's), doing more work per clock cycle than Kepler's ROPs.

EDIT: Like blackend said, you can't judge from a spec sheet alone.

 

[FatherMurphy]

Like blackend said, you can't judge from a spec sheet alone.

Fo' sho'.

Blackened, out of curiosity, I looked up the L2 cache in GK104. It has 128kb of L2 cache attached to each 64-bit memory controller (4 of them) for a total of 512kb. http://www.geforce.com/Active/en_US/en_US/pdf/GeForce-GTX-680-Whitepaper-FINAL.pdf

 

[RussianSensation]

I personally think GM200 is going to be released on 28nm and it'll be 50-60% faster than GK110.

But if NV themselves claims 2x the performance/watt for 1st generation (28nm Maxwell 750Ti) and this statement applied to the entire architecture, why would GM200 be only 50-60% faster than GK110, not 90-100%? For example, if GM200 is a 250W TDP card like 780Ti is, then why shouldn't be 90-100% faster? Simply looking at NV's claimed performance/watt increase, per NV a 100W Maxwell ~ 200W Kepler in performance. We see 750Ti delivering 90-100% more performance than a Kepler card with similar power usage. If anything, 16/20nm GM200 would have > 2x the performance/watt over 28nm Kepler.

 

[FatherMurphy]

But if NV themselves claims 2x the performance/watt for 1st generation (28nm Maxwell 750Ti) and this statement applied to the entire architecture, why would GM200 be only 50-60% faster than GK110, not 90-100%? For example, if GM200 is a 250W TDP card like 780Ti is, then why shouldn't be 90-100% faster? Simply looking at NV's claimed performance/watt increase, per NV a 100W Maxwell ~ 200W Kepler in performance. We see 750Ti delivering 90-100% more performance than a Kepler card with similar power usage.

Doesn't Nvidia need additional die space, though, to obtain those power savings on 28nm (just look at GM107 over GK107)? If so, I don't know that it is technically feasible to make a GM210 any larger than the current GK110 on 28nm, at least one large enough to accompolish that 90-100% gain. At the same die size and TDP, that would result in a GM210 that is more power efficient and faster than a GK110, but not 90-100% faster.

 

[blackened23]

Doesn't Nvidia need additional die space, though, to obtain those power savings on 28nm (just look at GM107 over GK107)? If so, I don't know that it is feasibly possible to make a GM210 any larger than the current GK110 on 28nm. At the same die size and TDP, that would result in a GM210 that is more power efficient and faster than a GK110, but not 90-100% faster.

The GK110 has a significant amount of die space for HPC floating point performance, which is disabled for consumer (ie geforce) products. While it's disabled for consumer products, the die space is still there. So i'd agree that there are limits on die space, but i'd imagine GM204 would not have any die space dedicated to HPC as the GK110 does.

That said, it's hard to predict what type of performance gain we're looking at. Although, I do feel that there will be an appreciable performance gain over the 780ti. I said it before, but there's really no reason for a "next gen" if it isn't better than prior gen in every way. I don't see NV doing that, it would be quite silly. That's also aside from the fact that the performance per watt of the Maxwell architecture will allow for sizable performance gains over kepler even at the same TDP levels, as the GT640 versus GTX 750ti shows (same TDP levels)

 

[FatherMurphy]

Sure thing, but RS was talking about the gains from GK110 to GM210, not GM204 over GK110/Gk104. It just isn't possible, I don't believe, to make GM210 on 28nm and get that 90-100% gain RS was discussing, due to die size limitations.

Obviously, the Gk104 to GM204 transition doesn't have that problem at 28nm, and Nvidia can increase the die size by ~150mm^2 to achieve 90-100% over GK104.

 

[Ajay]

Sure thing, but RS was talking about the gains from GK110 to GM210, not GM204 over GK110/Gk104. It just isn't possible, I don't believe, to make GM210 on 28nm and get that 90-100% gain RS was discussing, due to die size limitations.

Obviously, the Gk104 to GM204 transition doesn't have that problem at 28nm, and Nvidia can increase the die size by ~150mm^2 to achieve 90-100% over GK104.

^This^