Just a few short days left until we’ve got Nvidia’s official confirmation regarding the hardware specs of their top of the line Maxwell parts (the GTX 980 and the GTX 970), but the latest word puts the CUDA count at 2048 for the higher end part. Previously, we’d heard the GeForce Maxwell contains 1920 CUDA cores for the GTX 980.
This means the 980 features 16 SMM (Streaming Multiprocessor Maxwell), while the GTX 970 contains 13 SMM’s (for grand total of 1664 CUDA cores). That’s a gap of 384 CUDA cores between the two Maxwell parts – which isn’t too bad. If the cards overclock well, there’s plenty of scope for the 970 to overclock to GTX 980 levels.
GK104 is all to be effectively replaced by GM204, and while the performance is rumored to not be a great increase over Kepler, the efficiency has made significant strides. To illustrate the point, the GTX 770 has a TDP of 230W (1x 6pin and 1x 8pin power connector), while Maxwell’s GTX 970 is touted to only be 148W (2x 6 pin connectors).
From what we’re hearing, both parts while make use of 4GB og GDDR5 RAM, on a 256-Bit memory bus. The figure of 224 GB/s of memory bandwidth is still there. Some viewers and readers have asked my why the figure is so ‘low’. There are a few reasons. One, Maxwell is likely sporting a much larger cache, while puts less lean on the local memory bandwidth of the card. We’re also likely seeing the efficiency and improvements in the GPU’s architecture at play too. This could be similar to how AMD lowered the actual bandwidth of Tonga, but gained effective bandwidth by the usage of compression in the frame buffer. For more info on Tonga’s architecture check our comprehensive R9 285 review.
As things stand, it’s still not confirmed – we’ve a few more days for that. But don’t be surprised if it isn’t a million miles off. I’ve been hearing figures that Maxwell will be around 10 – 15 percent faster than the Kepler part they replace. So in other words, the GM204-400 will be about 10 percent quicker than the GTX 780. In theory, this means adding a moderate overclock to the core (and possibly memory) will produce similar results to the GTX 780 Ti – not too bad. Of course, this really does depend on pricing of the card.
My own concerns (and that of many others interested in the GPU market) haven’t changed. This is a great part for those who’ve not invested in either Kepler or say Hawaii, but for those which have it’s debatable if the so called Maxwell A revisions offer a compelling reason to part with your cash. As we’ve discussed just a few days ago (which also say the GTX 860 will be released in October), rumors do persist we’ll be seeing both a GM204-A1, A2 and finally a B version of Maxwell. Supposedly the B revision shall be the die shrink – but who can say.
We do know AMD are readying their cooler for the R9 300 series, but since it’s liquid cooling it could well be a power hungry beast. It’s possible as consumers we’ll be left with a choice – Nvidia’s more power efficient architecture and slower performance, or AMD’s hotter and higher TDP part – but which produces higher frame rates.