[guru3d rumor] GTX 880- 8GB-3200 SP's

[TeknoBug]

So lower memory bandwidth isn't a bad thing after all?

 

[Lepton87]

So lower memory bandwidth isn't a bad thing after all?

Of course it is. I lowered the bandwidth by half to simulate a 256bit bus and it really affects performance quite significantly.

 

[BrightCandle]

This node has taken too long to come. We have already had a refresh of cards over a year ago and we should right now be looking at lovely shiny 20nm GPUs with 2x the transistors and near 2x the performance (at least in compute). Having to wait nearly a year or see an architecture designed for 20nm on 28nm is a huge disappointment and not one which looks like it will bring much performance. Are people really going to ditch their cards for an extra 20% and a price premium as well? Not in volume they aren't.

 

[TeknoBug]

This node has taken too long to come. We have already had a refresh of cards over a year ago and we should right now be looking at lovely shiny 20nm GPUs with 2x the transistors and near 2x the performance (at least in compute). Having to wait nearly a year or see an architecture designed for 20nm on 28nm is a huge disappointment and not one which looks like it will bring much performance. Are people really going to ditch their cards for an extra 20% and a price premium as well? Not in volume they aren't.

Hence why I'm holding off until I see what they're like. I have a 750Ti and it's a pretty surprisingly good card for what it offers, I recorded a shadowplay clip playing Planetside 2 last night using it and was getting 60-110fps with closely the same settings as I use with the R9 280X or GTX770- but the R7 265 performs a little better than the 750Ti at the cost of more power consumption.

I somewhat doubt Nvidia will go with the performance to power consumption with the higher end 800's, it's one of the things a lot of us look for now due to increased power bills (my last bill was $1247).

 

[Ajay]

Oh, and while we're at it, if this chip scales similarly in performance over GM107 like GK104 did over GK107, then it should be significantly faster than a gtx 780 TI. I don't think it will be significantly faster, not with 32 ROP's and ~240gb/s bandwidth, but I do think with the given specs it will be somewhat faster.

If the 3200 CC number is factual, then it is Big Maxwell. Looking at RS' numbers, that's equivalent to ~4300 Kepler CCs. 32 ROPs and ~240 GB/s of bandwidth will crippling such a GPU. So, something isn't right, IMHO.

 

[Silverforce11]

20nm has been "Just around the corner" for two years, and is just now becoming available for SoC's. You seriously think that an HP 16nm process will be available in 2016?

Probably not on time from TSMC, but others may just be on track.

 

[Lepton87]

I did some tests to see how crippling 230GB/s of memory bandwidth would be. The test results don't look promising, I have a hard time believing that such bandwidth would suffice to overtake 780ti by 20%.
Uniengine Tropics

1440 MSAA 8X 876(1GHz)/1500MHz 288.4GB/s 74FPS
1440 MSAA 8X 1001MHz(1.125GHz)/1500MHz 288.4GB/s 81.6FPS
1440 MSAA 8X 1001MHz(1.125GHz)/1650MHz 317.8GB/s 84.7FPS
1440 MSAA 8X 1001MHz(1.125GHz)/1200MHz 230.8GB/s 71.6FPS


BW +37%
FPS +16.5%

It seems that even a Titan with just 2688 shaders is already hurting for bandwidth maybe not by much but it certainly doesn't have bandwidth to spare unlike past NV flagships. (GTX280/285, GTX580)

ps. To be honest with 3200SP that are more efficient to boot I would think that even keeping the bandwidth intact from 780ti would still be limiting this card. GTX580 was insensitive to memory overclocking that's why 680 with the same bandwidth could beat it soundly. Titan/780ti are sensitive to memory overclocking, they could certainly use more.

 

[Cookie Monster]

I did some tests to see how crippling 230GB/s of memory bandwidth would be. The test results don't look promising, I have a hard time believing that such bandwidth would suffice to overtake 780ti by 20%.
Uniengine Tropics

1440 MSAA 8X 876(1GHz)/1500MHz 288.4GB/s 74FPS
1440 MSAA 8X 1001MHz(1.125GHz)/1500MHz 288.4GB/s 81.6FPS
1440 MSAA 8X 1001MHz(1.125GHz)/1650MHz 317.8GB/s 84.7FPS
1440 MSAA 8X 1001MHz(1.125GHz)/1200MHz 230.8GB/s 71.6FPS


BW +37%
FPS +16.5%

It seems that even a Titan with just 2688 shaders is already hurting for bandwidth maybe not by much but it certainly doesn't have bandwidth to spare unlike past NV flagships. (GTX280/285, GTX580)

ps. To be honest with 3200SP that are more efficient to boot I would think that even keeping the bandwidth intact from 780ti would still be limiting this card. GTX580 was insensitive to memory overclocking that's why 680 with the same bandwidth could beat it soundly. Titan/780ti are sensitive to memory overclocking, they could certainly use more.

The test result you have presented are quite interesting but this is assuming that the same memory controller from GK110 (Kelper uArch) is used in the GM204. I highly suspect this is not the case.

 

[Mondozei]

20nm has been "Just around the corner" for two years, and is just now becoming available for SoC's. You seriously think that an HP 16nm process will be available in 2016?

Samsung/GloFo are starting volume production of 14 nm this year.
We will see AMD GPUs on 14 nm in 2016.

P.S I find it funny how everyone thinks AMD/NVIDIA will skip 20 nm. They can't. 14/16 nm - ex Intel - won't be here until 2016. And 20 nm is coming this year for SoCs, the rest are next year.

 

[Mondozei]

Picking up on the whole "Nvidia are skipping 20 nm!" meme which never made any sense at all:

20 nm Maxwell to offer native H.265 support.

Sure, it is just a rumor, but so is everything at this point, including the basis for this entire thread. I keep asking people to give a good, well-grounded argument for no 20 nm Maxwell next year. So far there hasn't been a single one.

 

[ShintaiDK]

Samsung/GloFo are starting volume production of 14 nm this year.
We will see AMD GPUs on 14 nm in 2016.

P.S I find it funny how everyone thinks AMD/NVIDIA will skip 20 nm. They can't. 14/16 nm - ex Intel - won't be here until 2016. And 20 nm is coming this year for SoCs, the rest are next year.

I think when people saw 20SOC they somehow instantly rejected it. The 16FF and 14FF is also SOC targetted nodes. But you can still make highend GPUs on it.

 

[Keysplayr]

Cant really argue a point that nobody knows anything about.

 

[dangerman1337]

I think when people saw 20SOC they somehow instantly rejected it. The 16FF and 14FF is also SOC targetted nodes. But you can still make highend GPUs on it.

28nmHP to 20nmSOC is not much of a jump for high performance GPUs (around 15% performance increase IIRC?) but TSMC's 16nmFF+SOC (http://www.cadence.com/Community/bl...-ahead-for-16nm-finfet-plus-10nm-and-7nm.aspx) is a considerable jump over 28nmHP. I wonder if Nvidia can create a Pascal on TSMC's 16nmFF+SOC in the early segment of 2H 2016.

You can make high end GPUs on 20nm but it is not economically feasible or provide much performance returns (though a TSMC 20nmHP wouldn't be much different) and you'll likely end up having customers to pay for more expensive GPUs in their relative segments.

 

[Lepton87]

The test result you have presented are quite interesting but this is assuming that the same memory controller from GK110 (Kelper uArch) is used in the GM204. I highly suspect this is not the case.

Nv has been making memory controllers for a long time, if they are going to improve bandwidth utilization it's going to be a rather small improvement.

 

[UaVaj]

given this is a 28nm part. this is what the spec needs to be.

GPU: GM204
CUDA: 3200 +
TUM: 240 +
ROP: 64 +
MEM: 8 GB
Interface: 512

GPU Speed: 1 GHz+
Mem Speed: 7400 MHz
Mem Bandwidth: 474 GB/s

Texture: 240 GT/s
Pixel: 64 GP/s

if so. will take 4 for 1440p surround (60hz absolute minimum with ALL the eye candy).

 

[boxleitnerb]

Nope. GM204 is GK104 successor, it doesn't need such crazy specs. Be realistic - the big jump comes with GM200 in late 2015.

 

[Cookie Monster]

Nv has been making memory controllers for a long time, if they are going to improve bandwidth utilization it's going to be a rather small improvement.

A GTX480 has twice more bandwidth than the GTX 750Ti, yet the latter performs about the same.

A GTX680 has the same bandwidth as the GTX580 yet the former performs 30~50% more than the latter.

Bandwidth is rarely the deciding factor, and this really depends on the underlying architecture. If everything is purely based on the quantity than the GTX660 would have soundly thrashed the GM107 based cards yet it doesn't (specs are 50% higher across the board yet it does not outperform in by 50% or even close to it).