NVIDIA Tweaks Kepler; Launches GTX 700M Series

It's Computex time again in Taipei, which means it's time for a whole lot of companies to unveil/launch a whole lot of products, all at once. With Haswell waiting in the wings, NVIDIA has jumped all over that bandwagon -- the GTX 770 launched earlier today, the 780 launched last week, and the company is taking the wraps off its new 700M series of mobile GPUs right now. Like the previous GTX 600M family, the GTX 700M cards are all based on Kepler and are built on a 28nm process.

What's new about the 700M family is that some of the relative positioning has changed. This is relatively unusual for NVIDIA or AMD -- while the companies make no promises about cross-series comparisons, both manufacturers typically maintain a degree of cohesion between series and part numbers. With the 700M series, NVIDIA is shaking things up a bit.

First, a note:  The 2GHz data clock NVIDIA has listed is actually meant to make it easier to compare against GPUs with DDR3 memory. Normally, you quadruple a GPU's listed memory clock if its using GDDR5 to get the actual data rate -- but in this case, you simply double it.

The old GTX 660M was a 384-core part clocked at 835MHz. The new 760M doubles the number of cores to 768 and trims the clock speed lower, to 675MHz (with Boost enabled to give back part of the difference in certain scenarios). That's going to be a net win for the GPU -- a 22% clock speed reduction with twice the cores still significantly boosts performance. The GTX 765M has no direct analogue, but the 770M closely maps the 670MX. Both GPUs have 960 cores, 80 TMUs, and 24 ROP partitions. The difference, in this case, is the clock speed. The 670MX topped out 600MHz; the 770M hits 811MHz with Boost offered on top of that.

The top-end card, the GTX 780M, is a 1536-core chip that corresponds to the GTX 680MX. Again, the basic stats are identical, and again, it's clock speeds that've been increased. Where the GTX 680MX ran at a flat 720MHz, the 780M is an 823MHz core with Boost available on top of it.

The bottom line is this:  The 760M should be anywhere from 30-50% faster than the 660M, the 770M is clocked 35% higher than the 670M (most likely translating to a 20-25% performance increase) and the 780M will be ~10% faster than the GTX 680MX, again thanks to clock increases. While it's fair to argue that new nomenclatures should be driven by new architectures as opposed to better binning and smaller improvements, NVIDIA's high-end GTX mobile parts have added a significant amount of performance compared to the older 600-series. They're not just rebranded 600-series parts, even if the feature gap is relatively small.

Sales Drop? What Sales Drop?

To hear NVIDIA tell it, times have never been better for GPU sales in mobile. The company is claiming to have won the vast majority of all Haswell-equipped gaming notebook launches, and believes that despite notable improvements, Intel's integrated graphics are still fundamentally incapable of driving modern titles, and as such, total sales are up sharply. NVIDIA's own sales and product margins suggest that thus far, this has held true -- the company's GPU revenue was flat last quarter, as opposed to experiencing the normal seasonal decline. Yearly GPU sales were actually up eight percent for Q1 of FY2014 as compared to FY2013.

This suggests that NVIDIA has captured a significant share of the mobile GPU market -- enough to insulate it from falling overall shipments, at least to date. In the long run, this trend may or may not be sustainable, but the company has had an enormous hit with its Kepler architecture.