NVIDIA Unveils Dual-GPU Powered GeForce GTX 690 93
MojoKid writes "Today at the GeForce LAN taking place in Shanghai, NVIDIA's CEO Jen Hsun Huang unveiled the company's upcoming dual-GPU powered, flagship graphics card, the GeForce GTX 690. The GeForce GTX 690 will feature a pair of fully-functional GK104 "Kepler" GPUs. If you recall, the GK104 is the chip powering the GeForce GTX 680, which debuted just last month. On the upcoming GeForce GTX 690, each of the GK104 GPUs will also be paired to its own 2GB of memory (4GB total) via a 256-bit interface, resulting in what is essentially GeForce GTX 680 SLI on a single card. The GPUs on the GTX 690 will be linked to each other via a PCI Express 3.0 switch from PLX, with a full 16 lanes of electrical connectivity between each GPU and the PEG slot. Previous dual-GPU powered cards from NVIDIA relied on the company's own NF200, but that chip lacks support for PCI Express 3.0, so NVIDIA opted for a third party solution this time around."
Re:Sure... (Score:3, Interesting)
Re:Wake me up when GK110 hits. (Score:5, Interesting)
I do have to hand it to Nvidia. The power requirements for the current 680 are very low and performance is quite impressive but GK 110 is going to be a monster...
Re:CUDA Double Precision? (Score:5, Interesting)
They are. However, their relative FP64 performance has dropped compared to the previous generation. If I remember correctly, there is now separate silicon to do FP64, rather than just a modified path in the FP32 cores. In the previous architecture, we were down to 1/12 of FP32 performance, only a third of some of the Fermi chip cores could do FP64, and at half speed. In the new chip, the FP64 cores can do full-speed calculations, but there are only 8 such cores, versus 192 conventional cores, giving a 1/24 performance ration.
However, Ryan Smith at Anandtech [anandtech.com] speculated that the existence of dedicated FP64 cores means that a future Fermi based on Kepler will be a mean beast, if they do a tape-out with exclusively FP64 cores. The only thing holding back double-precision then will be memory bandwidth (which would be a large enough deterrent in many cases).
As a game developer... (Score:2, Interesting)
As a game developer, I can tell you that the only thing that significantly affects frame rate in a GPU-bound game is GFLOPs. And as the owner of a 3-year old PC with a stock power supply, I'm most interested in the "x40" cards, because those are the highest card you can install in a machine with a stock 350W power supply.
According to what I see on Wikipedia [wikipedia.org], NVIDIA apparently pulled a fast one this generation and re-branded some 500 series cards as the PCIe 2.0 x16 versions, while all the cards with impressive performance are PCIe 3.0 x16. The impressive ones get ~2x higher GFLOPs/W.
Old PCs like mine can't use PCIe 3.0, so this means the GF116-based GT 640 that gives 415 GFLOPs at 75W is still the fastest card that you're likely to find in a 2-3 year old PC with a game enthusiast who updates his card every generation. Compare that to the GT215-based GT 240 from 2009 which gets 386 GFLOPs at 69W, and you can see that there is ZERO reason to upgrade this generation, unless you also plan on upgrading your motherboard.
So yes, you can get a GK107-based GT 640 with 730 GFLOPs at 75W, but you have to upgrade your machine and get a PCIe 3.0 x16 motherboard. Boo, NVIDIA. BOO.