So you may be wondering what Nvidia is cooking up next on the GPU front and the answer is, not susprisingly, Maxwell. But it’s not Maxwell as you might know it. Spyshots of an engineering sample of the GTX880 have leaked thanks to Chinese website Gamerspy and a few things can be learned about what we can expect from Maxwell. Bare in mind that this is an engineering sample you’re going to see in the article below – the final GM204 GPU won’t look like this at all, nor will it necessarily be called the GTX880.
The first shot is of the board itself and things are quite different. The board itself is much higher than standard ATX requirements, the SLI fingers have moved to the top and sideways and it looks like there are extra VRM phases and chokes to keep power draw in check. The memory configuration changes significantly with 2Gb SK Hynix chips arranged in the same manner as the Geforce GTX680. The eight chips contribute 4GB of GDDR5 memory on the front and 4GB on the rear for a total of 8GB.
This is a change from previous leaked engineering sample boards that shipped with 16GB of RAM, but without stacked VRAM it would require that Nvidia use a 512-bit bus. Engineering samples typically ship with large amounts of VRAM so that manufacturers like Nvidia can test the memory bus thoroughly and identify the optimal configurations for the different products they plan to sell. Sony’s Playstation 4 development kit, for example, ships with 12GB of GDDR5 memory.
The back of the board doesn’t reveal much, just another 4GB of VRAM. The sample on show here has two DVI-D ports as seen from the wire traces running off the board on the right-hand side of the picture.
This is most definitely a Geforce Maxwell chip, though, of that there is no doubt. The first leaks of the GM204 chips was in an Indian import-export database, which listed a GM204-A01 GPU with the model number TA995. This chip is already a new revision and a new hardware version and it wasn’t made that long ago – the 21st week of 2014 was 19-24 May, only six weeks ago.
Physically the chip is huge, measuring in at 430mm, which is almost as large as the GK110 chip found in the GTX Titan and the GTX780 Ti. GK104, found in the GTX680, GTX770, is just 294mm. If you cast your memory back to the Geforce GTX750 Ti launch, Maxwell was made on the same 28nm process as Kepler, but increased transistor counts by 43% while only increasing die size by 25% to reduce heat and improve power consumption.
That seems to be a similar arrangement here. The die size increases by 31% and it’s anyone’s guess how many more transistors or functional units are packed in there. GM204 might be as powerful as a single GTX Titan, given the die size and the fact that the GTX750 Ti was able to match and occasionally beat the bigger Geforce GTX660 once overclocked.
And this may be how Nvidia is going to get around power supply limitations. If you discount the power draw that a GPU with two 8-pin PEG connectors takes from the motherboard, you’re left with a 300W hard limit according to the ATX specification (you can draw more, but it’s tricky). By including two 6-pin and one 8-pin PEG connector here, Nvidia has added in another 75W of power draw. 375 Watts of power is nothing to sniff at and in fact that’s the TDP of the Geforce GTX Titan Z. Scary stuff.
If these samples are out in the wild already and working, it’s quite reasonable to expect Nvidia to be aiming for a Q4 2014 launch because TSMC’s 20–nanometer process isn’t ready for the big time yet.
So, who’s buying one?