So Big Maxwell came out. And it's $999. It's basically one and a half 980s. The 295X2 seems to keep up when you're measuring average FPS, but since no one does that anymore, all you should care about is frame time consistency. And the Titan X wins because it's a single GPU configuration. MultiGPU setups are just choppier - no way around that. Oh and power consumption, lel. The best part is that I think this is a fully enabled GM200 so there might not be a "980 Ti" that magically finds extra SMMs in a few months. Trickster gon get dat pocket book ready, mhmm.
Meh, I'd just wait a few months to see what 390X & Friends do to gpu prices. You might be able to snag a 290 for like $200, which oughta work well enough around the time of Star Citizen 2.0's multi crew stuff this summer/fall.
This is probably amazing to use to play high end games with high settings even if i am too poor and gaming companies will think we have hardware like this.
We'll probably see a 6 GB GM200 GPU that is a cut-down version of the TITAN X soon. But I'm more excited for the 390X.
Pascal is rumored to be much better according to pcgamer but the 10x increase they are saying sounds ridiculous
Yeah, it's "CEO math". I believe it's 10x on a particular use case which isn't gaming. And it's generally understood that the 10x is "rounded up." So I wouldn't expect Pascal to be unusually amazing for gaming. We get pretty much the same performance increases at pretty much the same rate.
10x is the marketing number for neural networks (I don't really know what they are) and the improvements for gamers will almost certainly be much lower. Also the numbers are theoreticals and actual performance may be lower. The "5x" total for the left half is an average of the compute and bandwidth improvements. The "2x" total for the right half is because the physical number of GPUs is doubled—Pascal's NVLINK interconnect allows 8 GPUs to be linked together instead of the previous maximum of 4. I don't know if that automatically means that 8-way SLI will arrive with Pascal. "Mixed precision" refers to FP16 (half precision, 16-bit floating point numbers) and FP32 (single precision, 32-bit floating point numbers). In Pascal (and the Tegra X1), twice as many FP16 operations as FP32 operations can be done per clock cycle. Combine that with the 2x FP32 increase from Maxwell to Pascal and you get the 4x number. The 6x number for 3D memory consists of 3x from the increase in memory bandwidth from High-Bandwidth Memory (HBM) and 2x from using FP16 instead of FP32. The 10x number on the right half refers to the performance of the interconnect that connects multiple GPUs, which consists of 5x from the upcoming NVLINK interconnect and 2x from using FP16 instead of FP32. If Pascal ends up as planned, it will be awesome for neural networks and presumably other compute tasks. But for gaming we are likely to get a different story. As far as I know, desktop games generally use FP32 (single precision). Also, I don't assume that all Pascal GPUs will have HBM, and I think any GDDR5 chips will have much more modest bandwidth improvements. If you have $10,000 to spend on electronics, you could probably get 8 Pascal-based TITAN-level GPUs instead of something else. If you use 1 GPU, NVLINK doesn't matter (AFAIK). So for most gamers, we're looking at around 2x compute, 1x–3x bandwidth (depending on the GPUs), and the same number of GPUs as before, for a total of 2x from Maxwell to Pascal.
Why do they even bother designing shit to work in eights? The only people who'd do that are nuts because they can and worshippers of Khorne
