So, we recently had the GTX660 Ti launch which cemented the card as the go-to option if you wanted better performance on average over the competing Radeon HD7870, while also allowing the card to kick the HD7950 in the nads with a bit of overclocking. I mused on the existence of the GTX660 and even made a prediction based on what I knew from the GTX660 Ti analysis.
Nvidia does its OEM releases very quietly and doesn’t want the retail sector to know that the cards exist, because these are only meant for pre-built systems from the likes of IBM, Dell and HP. But judging from the pattern we’ve seen, chopping down shader modules, cache and RAM from the GK104 architecture seen in the GTX680, it looks like my predictions hit the nail on the head. Woohoo!
To be honest, I didn’t expect the final specifications to come exactly up to my prediction when I made it. From my GTX660 Ti analysis:
“The upcoming GTX650, on the other hand, should be a little bit lower than that. On good days, it needs to go toe-to-toe with the HD7870 and mostly mingle with the HD7850. You can expect bandwidth for that card to be lowered to 96GB/s with another shader disabled, bringing the card to 96 texture units, 24 ROPs and 1152 CUDA cores, with performance to around 20% slower. For the vanilla GTX660, I reckon the card’s RAM and core clocks will be lowered and physically limited, dropping performance by 10%, keeping everything else the same.”
Just like that, I’m right on the money with an extra shader module disabled, the same number of texture units (so the texture fillrate is unchanged) and the same number of ROPs. Clock speeds, as predicted, have fallen to 823MHz Core and 888MHz boosted, with RAM clocks falling down to 5.8Ghz. The bus width is still 192-bit in size, but Nvidia has chosen to forgo interleaving of the RAM on the lower-end model, sticking 1.5GB RAM on the lower-end model and interleaving RAM for the 3GB version. I call the OEM versions GK104-C, seeing as its a further extension of the GK104 base architecture in the GTX680.
Because these are OEM-only, there’s less of a focus on gaming performance and more on application performance, particularly since these cards will be a great fit for workstations that need to run CAD or other CUDA-compatible applications. Chopping off an interleaved set of 2Gb modules leaves the card with six chips on a 192-bit bus. Using the forumala to calculate bandwidth (Bus width * frequency in Ghz /number of chips) we land up at 185.6GB/s, not far from the GTX660 Ti’s 144.2GB/s but boosting performance a fair bit for memory-intensive applications – again, the target market for the OEM versions use apps that usually require a lot of very fast RAM. However, interleaving the extra 1.5GB of RAM would mean an extra six chips on board, dropping performance down to 92.8GB/s. If Nvidia clocks up the RAM on the 3GB OEM version that would bring it up to 96GB/s, not much of an improvement.
What’s going to be fun is finding this card in the retail sector and overclocking the RAM. Just by boosting up the RAM to 6GHz, you’d land up at 192GB/s, giving you the edge in bandwidth-favouring games like Crysis 2 with a large amount of textures that need to be loaded constantly.
What isn’t going to be fun is overclocking the core frequencies. The card is rated for 130W at boost speeds and only comes with a single PCI-Express 6-pin connector (meaning that its actually a better match-up for the Radeon HD7850). Adding that to the 75W available from the motherboard, your maximum power delivery is 150W. I’d be surprised if anyone gets these cards to anything near 1GHz, let alone 950MHz in a boosted state. With those limited overclocking capabilities, it would mean that the card is destined to dupe it out with the $240 market when it lands in the retail channels, competing with the price-cut Radeon HD7870 which boasts better numbers and will outperform it in games that need more RAM.
The rest of the card looks unchanged, using the same PCB as the GTX660 Ti and the GTX670 along with the same display outputs (two DVI dual-link ports, one HDMI-out and one full-size Displayport-out). The reference cooler is again the same one lifted from the GTX670 meaning that the vanilla GTX660 should stay much cooler than its bigger brothers.
Still, I’m stoked that I’m right (again) and can’t wait to see how the vanilla version will perform. One thing though that bothers me is the amount of RAM available. Nvidia may well interleave things like on the GTX660 Ti, keeping the card at 2GB of RAM so that it doesn’t lose out too badly in games when AA is applied. That would bring performance down further, which might prompt some buyers to source out the OEM versions instead.