Will These be the GTX 1070 Specs?

Yesterday, all eyes were on Nvidia as they launched their new Pascal GeForce GTX 1080 graphics card. As the current undisputed performance champion, it carries a hefty price tag. This is why there is the GTX 1070, with about 81.25% of the performance at a much lower price. Unfortunately, Nvidia has only officially confirmed the GTX 1080 specifications alone. Based on those specifications, we can make a number of assumptions about the GTX 1070 specs.

First off, both cards are based on the same GP104 die. The variant the GTX 1080 features has 2560 CUDA cores clocked at 1607/1733 MHz base/boost, offering 8TFLOPS of performance x Using the cut-down GP104-200, the GTX 1070 offers 6.5 TFLOPS, approximately 81.25% of the GTX 1080. As the TFLOPS is a theoretical calculation based on the number of CUDA cores x clock speed, we can use this as a basis to determine the GTX 1070 specifications.

Given past Nvidia cut chips, the GTX 1070 will either feature 2304 or 2048 CUDA cores, with either 256 or 512 cores cut out. if the card features the same clock speed, 2048 cores would be enough to offer 6.5TFLOPS. If Nvidia chooses the lower the clock speeds to say 1400-1500 MHz, they would need a few more CUDA cores at 2304. I suspect that Nvidia will choose the option better suited to the dynamic between pure yields and stable clock speeds.

The last part of the chip left is the memory controller. Nvidia has already revealed that the GTX 1070 will use 8GB of GDDR5. In order to not bottleneck the card, it would have to be 8Gbps and over a 256bit bus for 256GB/s of bandwidth. This would give us 80% of the GTX 1080 bandwidth, right where we want to be. Of course, Nvidia still might be playing tricks like they did with the GTX 970.

Looking at all this, Nvidia looks to have a real winner on their hands. The move to Pascal and 16nmFinFET has reaped a windfall with improved performance per clock, efficiency gains, and higher clock speeds to boot. So far Nvidia has claimed 2x performance per watt. With AMD claiming 2.5x performance per watt, the fight is on as AMD prepares to launch their own Polaris GPUs later this month.