These last few Nvidia GPU generations one has become accustomed to having Nvidia rejig the mobile GPU chips from one generation to another, rebranding its old parts as low-end in the next generation, and marketing the shiny new stuff as the high-end kit, given the difference in performance.
With Kepler, however, things seem to be a bit upside down. As the new parts are far more energy-efficient than their predecessors, Nvidia seems to have thought it OK to stick Kepler in the middle range and leave the high-end 600M parts for Fermi. That’s right. The high-end Geforce 675M and 670M are based on Fermi, while the mid-range are Keplers. The Geforce GT 640M LE may or may not be a Kepler part, so Nvidia suggests you look under the hood before doing it. Fermis re-occur in the Geforce GT 635M, 630M and 620M parts, which throw another spanner in the works.
Brian Choi, Product Marketing Manager at Nvidia, responding to users’ comments, said: “Our GeForce GT lineup has new chips up and down the stack, and Kepler can be found in GeForce GTX 660M, GeForce GT 650M, GeForce GT 640M and most GeForce GT 640M/LE. The performance range is guaranteed within a branding range but if you're interested in the particular architecture, we encourage you to check it out before buying.”
While this doesn’t sound too bad, it’s the benchmarketing you can find on this page – trying to emphasise the superiority of the Geforce 600M series – that raises some eyebrows. It’s true that Nvidia has not been claiming wildly superior performance with its Kepler design, then again, review sites have done that job for it.
The company has simply stated the cards to be more energy efficient while outperforming their GF1xx ancestry, however we can’t help but question why does a GTX 660M (based on Kepler, according to Nvidia) with 384 CUDA cores, pushing pixels at the rate of 835MHz get hammered by a GTX 670M running 336 CUDA cores at 598MHz?
You may consider memory bandwidth to be the issue here as the GTX 660M carries a 128-bit interface and the GTX 670M a 192-bit one, but the Kepler units carry just 16 Raster Output Pipelines, aka ROPs. That being the case, Kepler is seriously underpowered in its mobile version. Unhindered by this, a GTX 660M should, clock for clock, CUDA core for CUDA core, wipe the floor with the 675M.
You can see the differences below.
We had already raised this concern when we discussed the Acer Timeline M3 Ultrabook powered by the Kepler-based GT 640M, earlier this month, as the Kepler-based card was only mildly superior to its ancestor.
Another point that sends alarm bells ringing is that throughout the entire benchmarketing exercise, Nvidia does not mention power consumption a single time - something vital to notebooks - while claiming :“The GeForce 600M series represents a huge step forward, continuing to optimize power usage, heat output, and game performance, making ultrabooks with dedicated graphics a reality.” This is only partly true if 40nm-based GF1xx parts continue to be present.
In the end, what Nvidia promises is performance within a given price range, and that’s what it is delivering. If you specifically want Kepler, you should definitely be looking under the hood and not jump at the opportunity to buy the first 600M-based notebook that shows up.