Even More GeForce GTX 980 and GM204 Specs Tumble Out

[btarunr]

Ahead of its launch later this week, even more details of NVIDIA's upcoming GeForce GTX 980, and the 28 nm "GM204" silicon it's based on, tumbled out. To begin with, the GM204 silicon is confirmed to be built on the 28 nm silicon fab process. The chip bigger than that of the GK104, with a die area of 398 mm², yet smaller than the GK110, which measures 581 mm². Its transistor count is 5.2 billion, about 2 billion more than the GK104.

The component hierarchy of GM204 is similar to that of the GM107 silicon, on which the GTX 750 Ti is based. The GPU features a 256-bit wide GDDR5 memory interface, and PCI-Express 3.0 x16 bus. The GigaThread Engine dispatches workload between four graphics processing clusters (GPCs), the basic subunit. Each GPC has a common raster engine shared between four streaming multiprocessors Maxwell (SMMs), which each hold 128 CUDA cores. The total CUDA core count is hence 2,048. The L2 cache has been quadrupled over GK104. The chip features 2 MB of it, compared to 512 KB on its predecessor. The GM204 features 64 ROPs, double that of the GK104, and should hence come with a strong geometry processing muscle. The chip features a revolutionary new 3-bit delta color compression technology that makes the most of the limited memory bus width of this chip.



Here are the final specifications of the GTX 980 and GTX 970, carved out of this chip.

GeForce GTX 980

• 2,048 CUDA cores
• 128 TMUs, 64 ROPs
• 256-bit wide GDDR5 memory interface
• 4 GB standard memory amount
• Core clock speeds of 1126 MHz, with 1216 MHz GPU Boost, and 7012 MHz memory
• 165W TDP

GeForce GTX 970

• 1,664 CUDA cores
• 112 TMUs, 64 ROPs
• 256-bit wide GDDR5 memory interface
• 4 GB standard memory amount
• Core clock speeds of 1051 MHz, with 1178 MHz GPU Boost, and 7012 MHz memory

View at TechPowerUp Main Site

 

[GhostRyder]

That is one heck of a boost frequency and ram speed!

 

[AlienIsGOD]

looking great spec wise

 

[inthedark1980]

So is the higher end be called the GTX 990 and the only assuming they could call the titan killer then GTX1000. Sounds a bit like a motorbike. But it will most likely have a name like Saturn or some shit like that

 

[MxPhenom 216]

64 ROPs on GM204 node? I felt like 32 or 48 made a lot more sense for the chip, the latter more so. What will GM200 have?

 

[Ralfies]

Very happy to see more than 32 ROP's. IMO that's what holds back smaller memory bus designs, since bandwidth can be compensated for through other means(memory speed, cache, compression).

 

[W1zzard]

But it will most likely have a name like Saturn or some shit like that

and then you wonder how Apple can make more money than God, sticking with the same name and a super simple number scheme ...

 

[RHoudini]

165W TDP for 780-like performance is awesome.

 

[romeg]

I've not closely followed news of these new GPUs. The pricing isn't half bad, nor the specs, so I'm curious how the 980 will stack up against my Asus Direct CU II GTX 780Ti. All things considered, the 780Ti is a tough act to follow.

 

[BorisDG]

165W TDP for 780-like performance is awesome.

I think that 980 is comparing even with 780Ti or?

 

[Ferrum Master]

3-bit delta color compression technology

YAY more washed out textures... exVoodoo team at the helm lol...

Seriously, are the idiots? They are badly famous for their in inchip hardware crap dithering since Riva TNT times, in smaller scale they still cheat on that... I hate my nvidia cards for that. Noone says it is lossless... sigma delta modulations seldom are...

ROP are for 4K, that's natural as AMD has a advantage there, but weak bus will introduce serious problems with traditional anti aliasing, it may spank badly... they try to compensate it with high frequency specs, but you never can outmatch the hardware pipe wideness...

But still it will benchmark fast... but... real time scenarios... khem khem... we do like SSAA, do we? If using triple monitors it is still needed despite 4K render resolution...

 

[MxPhenom 216]

I think that 980 is comparing even with 780Ti or?

its been said it'll sit in between the 780 and 780Ti, but that was before THESE specs.

 

[RHoudini]

3-bit delta color compression technology

YAY more washed out textures... exVoodoo team at the helm lol...

Probably similar to the technique used by the AMD R9 285, a loss-less compression algorithm without any impact on the image quality.

 

[Suka]

Looking like it will be a strong GPU same for the 970. Compression everywhere

 

[dj-electric]

To be honest, i'm quite disappointed if the specs and prices are right.

I hoped for them to make a true step ahead in GTX 900 series. There were talkings about a 3200 core part, witch at the time seemed too good to be true.

I did not expect the GTX 970 part to contain a mare 1664 shader count, or the GTX 980 part to have 2048. This leaves plenty of room for them to milk 2304 and 2560 parts as-well. Especially considering the low TDP.

 

[Silas Woodruff]

That performance with just 165 watts, that is quite impressive, I like it.

 

[Ferrum Master]

Probably similar to the technique used by the AMD R9 285, a loss-less compression algorithm without any impact on the image quality.

Won't happen, less significant bit will be omitted for sure, I dig up the patent, it has a destructive mode that offers this magic thus making compensation the lack of a third of the bus, the seconds... those both cards are rubbish... especially considering that his faster 7970 predecessor is a aging 2 year grandpa...

And we all know... that any kind of algorithm in between introduces more latency...

We are being fed for the same for too many years...

 

[BorisDG]

its been said it'll sit in between the 780 and 780Ti, but that was before THESE specs.

So now with THESE specs? Where is positioned?

 

[64K]

So now with THESE specs? Where is positioned?

According to the source the OP cited the GTX 980 will be launched this Friday. I suspect we will see some reviews very soon and then we will know. It's just a few days.

 

[Sihastru]

3-bit delta color compression technology YAY more washed out textures... exVoodoo team at the helm lol... Seriously, are the idiots?

Delta compression is lossless. AMD uses a similar method for the 285. Most cloud backup services use a similar method.

 

[Yellow&Nerdy?]

I have to say, I'm really impressed with how efficient Maxwell really is. Just imagine the type of performance per watt numbers Nvidia will be able to deliver when they shrink it down to 16nm FINFET... I think getting a decent 4K gaming experience on a single GPU might be closer than we think.

 

[W1zzard]

Most cloud backup services use a similar method.

source? delta compression is not very efficient (but fast) compared to real compression algorithms. i'd expect most backup services to use deduplication techniques to store identical files or disk blocks just once, and some might use compression not unlike zip

 

[RejZoR]

Everyone wanking on the efficiency. If you want efficiency, stay with stupid integrated crap. I don't care what the TDP is for as long as performance justifyes the insane prices we are expected to pay every time for tiny small bumps in performance. It's idiotic. Every new card brings cost of 500 EUR, yet it gives what, 10-15% more performance. Which is pretty much useless difference.

 

[MikeMurphy]

To be honest, i'm quite disappointed if the specs and prices are right.

I hoped for them to make a true step ahead in GTX 900 series. There were talkings about a 3200 core part, witch at the time seemed too good to be true.

I did not expect the GTX 970 part to contain a mare 1664 shader count, or the GTX 980 part to have 2048. This leaves plenty of room for them to milk 2304 and 2560 parts as-well. Especially considering the low TDP.

Not at 28nm they won't. Expect an increase with a die shrink.

 

[Sihastru]

source? delta compression is not very efficient (but fast) compared to real compression algorithms. i'd expect most backup services to use deduplication techniques to store identical files or disk blocks just once, and some might use compression not unlike zip

dropbox is one example that comes to mind. They use a "delta" api mostly to keep track of changes to files, but also to cut down on space and bandwidth usage. The api is even exposed to developers as a service.

Of course backup services use every trick in the book to save on space usage.

It could also be just jargon...