While Nvidia’s Kepler architecture permeates every single price point in the high-end market, the rest of the old GTX500 lineup slowly depletes its stock while waiting for a refresh. Anxious gamers eager to see how things will turn out will have to wait no longer – in less than eight days’ time, the GTX660 and GTX650 cards will be unleashed and Nvidia’s Geforce 600 series refresh will be complete.
The introduction of the GTX650 also paves the way for a Ti version thanks to the expected performance gap that trimming down the Kepler architecture will bring – and that will be one elegant budget performer. More info after the jump.
The card above was pictured by Mydrivers.com and all indications are that its exactly like the GT640, with the same fan and chip arrangement and the pin-outs filled with an extra 6-pin PCI-Express connector. That’s a bit weird considering Nvidia puts the TDP at 65W (the GT640 is pegged at 75W). At launch it’ll probably be an overclocked GT640. The GTX650 packs 384 CUDA cores, 16 ROPs and 1GB of DDR5 RAM on a 128-bit bus. Clock speeds haven’t been confirmed, but the highest-SKU GT640 started off with 950MHz for the core, so this will likely be around 1GHz to combat the inevitable GHz editions of the HD7750.
Its not going to set pants on fire, lets face it, but it may deliver some playable performance, especially if the memory is clocked at 5GHz. That PCB is also tiny, allowing you to stick that card into a lot of small ITX-sized chassis. Portablte 720p LAN powerhouse? You’d bet that something like this will find its way in there.
If you’re disappointed, I urge you to hold on for the launch of the inevitable GTX650 Ti. Considering how the Kepler architecture has scaled thus far, I predict it’ll boast 768 CUDA cores, 16 ROPs, 48 Texture Units and 1-2GB 5GHz DDR5 RAM on a 128-bit interface.
The Geforce GTX660, meanwhile, has yet to be photographed in full, but chip shots have been seen online and various sites have tried to measure them from pictures taken of the GPU. Visually, its a smaller chip, a departure from the cut-down GK104 dies that are used in the GTX660 Ti, GTX670, GTX680 and GTX690, with various parts disabled or cut away. GK106 is around 54mm² smaller and is even smaller than the OEM GTX660, currently on the way to OEM partners around the world. I previously predicted that the retail GTX660 would ship with 1152 CUDA cores, 96 texture units and 24 ROPs. Turns out, according to a new table that I’ve had to revise, this is what the GTX660’s core should look like:
The GTX660 vanilla will ship with 960 CUDA cores, 80 texture units and 24 ROPs, with cut-down cache on a 192-bit bus. Performance will be around 15-20% trailing behind the GTX660 Ti and the card will use a single PCI-Express 6-pin connector, pegging the power draw under 150W maximum. Will it feature technology like GPU Boost? That’s debateable, considering the nature of the mid-range market that doesn’t always overclock their cards for extra performance. However, if this ends up at around 800MHz for the core, there might not be a lot of headroom to exploit for overclockers, leaving things to GPU Boost to rather do its job.
Size-wise it should end up a lot shorter than the PCB of the GTX660 Ti and will fit in nicely in any chassis, even a few ITX ones. Coolers will be limited to dual-slot versions, but don’t be surprised of OC versions don’t deviate too far from the original specifications. Overclocked it should run close to the OEM GTX660, but staying firmly away from the Ti version thanks to the expected L2 cache cut.
Well, we’ll just have to see what’s cooking on the 12th of September, won’t we?