Closing the Performance Gap with Desktops

If we look back at the past several generations of GPUs from NVIDIA, the GTX 480 launched in March 2010 and had 480 CUDA cores clocked at 700 MHz with a 384-bit memory interface and 3.7GHz GDDR5 (177.4 GB/s). The mobile counterpart GTX 480M officially launched just a couple months later (though it wasn't really available for purchase for at least another month), but it was a rather different beast. It used the same core chip (GF100) but with a cut-down configuration of 352 cores clocked at 425 MHz and a 256-bit memory interface clocked at 3.0GHz. In terms of performance, it was about 40-45% as fast as the desktop chip. GTX 580 came out in November 2010, with 512 cores now clocked at 772 MHz and 4GHz GDDR5; GTX 580M appeared seven months later in June 2011 with 384 cores at 620 MHz and 3GHz GDDR5, and it used a different chip this time (GF114 vs. GF110). Performance was now around 45-55% of the desktop part.

The story was similar though improved in some ways with GTX 680 and GTX 680M. 680M had 1344 cores at 720 MHz with 3.6GHz GDDR5 while GTX 680 had 1536 cores at up to 1058 MHz with 6GHz GDDR5. They were three months apart and now the mobile chip was around 55-65% of the desktop GPU. GTX 780/780M were basically announced at the same time (though mobile hardware showed up about a month later, in June 2013), and as with 580/580M the notebook part used a smaller chip than the desktop (GK104 vs. GK110). The performance offered was again around 55-65% of the desktop part. Then of course there's GTX 880M, which is sort of the counterpart to GTX 780 Ti. It uses a full GK104 (1536 cores) while 780 Ti uses a full GK110 (2880 cores), and the delay between the 780 Ti and the 880M launches was four months, and while the desktop GPUs never saw the 800 series, GTX 880M is down to around 50-60% of the top desktop GPU, the GTX 780 Ti.

That brings us to today's launch of the GTX 980M/970M. You might say that there have been patterns emerging over the past few years that hint where NVIDIA is going – e.g. Kepler GK107 first launched on laptops back in March 2012, with desktop GPUs coming a month later – but the higher performance parts have almost always been desktop first and mobile several months later, with at best 50-65% of the performance. Now just one month after NVIDIA launched the GTX 980 and 970, they're bringing out the mobile counterparts. What's more, while the mobile chips are yet again cut-down versions of the desktop GPUs, clocks are still pretty aggressive and NVIDIA claims the 980M will deliver around 75% of the performance of the GTX 980. Here's a look at the specifications of the new mobile GPUs.

NVIDIA GeForce GTX 900M Specifications
  GTX 980M GTX 970M
CUDA Cores 1536 1280
GPU Clock (MHz) 1038 + Boost 924 + Boost
GDDR5 Clock 5GHz 5GHz
Memory Interface 256-bit 192-bit
Memory Configuration 4GB or 8GB 3GB or 6GB
eDP 1.2 Up to 3840x2160
LVDS Up to 1920x1200
VGA Up to 2048x1536
DisplayPort Multimode Up to 3840x2160

The specifications are actually a bit of a surprise, as the core clocks on the 980M are right there with the desktop parts (though it may or may not boost as high). The 980M ends up with 75% of the CUDA cores of the GTX 980 while the memory clock is 29% lower. In terms of pure theoretical compute power, the 980M on paper is going to be 70-75% of the GTX 980. Of course that's only on paper, and actual gaming performance depends on several factors: GPU shader performance and GPU memory bandwidth are obviously important, but the CPU performance, resolution, settings, and choice of game are just as critical. In some games at some settings, the 980M is very likely to deliver more than 75% of the GTX 980's performance; other games and settings may end up closer to 70% or less of the desktop. Regardless, this is as close as NVIDIA has ever come to having their top notebook GPU match their top desktop GPU.

A big part of this is the focus on efficiency with Maxwell GM204. NVIDIA doesn't disclose TDP for their mobile parts, but the top mobile GPUs usually target 100W. NVIDIA went after efficiency in a big way with Maxwell 2, dropping TDP from 250W with GTX 780 Ti down to 165W with GTX 980, all while delivering a similar (often slightly better) level of performance. With further binning and refinements to help create a notebook GPU, the TDP target would be 60% of the GTX 980 and power requirements tend to scale quite a bit near the maximum stable clocks for any particular microprocessor. Reduce the memory clocks a bit and disable some of the SMM units and getting 75% of the performance with 60% of the power requirement shouldn't be too difficult to pull off.

Moving on to the GTX 970M, NVIDIA is still using GM204 but it has even more SMM units disabled leaving it with 1280 CUDA cores. The memory bus has also been dropped to a 192-bit interface, but with a slightly lower core clock and fewer cores to feed, the GTX 970M should do well with a 192-bit bus. The smaller memory bus also translates into less total memory this round, so NVIDIA isn't doing any asymmetrical memory interface on the 970M; it will have 3GB GDDR5 standard, with an option to go with 6GB. It's good to see the potential to get more than 3GB RAM, as we're already seeing a few games that are moving past that target.

In terms of theoretical compute performance (cores * clock speed), the GTX 980M will be about 30-35% faster than the GTX 970M in GPU-bound situations. If you're curious, the GTX 970M will also offer around 55-65% of the performance of the desktop GTX 970, so the second tier GPU ends up being closer to what we've seen with previous generations of NVIDIA mobile GPUs.

With the launch of the new GTX 970M and GTX 980M, it's also worth mentioning that NVIDIA is officially discontinuing some of the existing mobile parts. The current lineup of mobile GPUs from NVIDIA now consists of GeForce 820M, 830M, and 840M for the casual/less demanding market. The 820M is actually a Fermi-derived part, while 830M and 840M use GM108 with 256 and 384 cores, respectively. At the top of the product stack, the GTX 980M and 970M replace the GTX 880M and 870M, while GTX 860M and 850M continue as the "mainstream gaming" notebook GPUs; 860M also continues to be offered in two variants, a Maxwell GM107 version and a Kepler GK104 version, though the latter hasn't been widely used.

Introducing Mobile Maxwell: GM204 for Notebooks GTX 980M and 970M Notebooks and Conclusion
POST A COMMENT

68 Comments

View All Comments

  • chizow - Tuesday, October 7, 2014 - link

    Except most professionals don't want to be part of an ongoing beta project, they want things to just work. Have you followed the Adobe CS/Premiere developments and OpenCL support fiasco with the Mac Pros, and how much of a PITA they have been for end-users? People in the real world, especially in these pro industries that are heavily iterative, time sensitive, and time intensive cannot afford to lose days, weeks or months waiting for Apple, AMD and Adobe to sort out their problems. Reply
  • JlHADJOE - Tuesday, October 14, 2014 - link

    This. As cool as open standards are, it's also important to not stuck. The industry has shown that it will embrace open when it works. Majority of web servers use open source, and Linux has effectively displaced UNIX out of all sectors except for extreme big-iron.

    But given a choice between open and "actually working", people will choose "working" every time. IE6 became the standard for so long because all of the "open" and "standards-compliant" browsers sucked for a very long time.
    Reply
  • mrrvlad - Thursday, October 9, 2014 - link

    I have to work with both CUDA and openCL (for amd GPU) for compute workloads. The main advantage of CUDA is thier toolset - AMD compiler is ages behind and does not allow developer sufficient control of the code being generated. It's more of "randomizing" compiler, not optimizing... I would never even think about using openCL for GPU compute if I will start a new project now. Reply
  • Ninjawithagun - Sunday, June 28, 2015 - link

    The problem is that what you are stating is only half the story. Unfortunately, each company does have a superior solution. Going with OpenGL is a compromise at best because the coding does not maximize or optimize automatically for hardware specific architectures. Maybe in a perfect world we would have an open non-proprietary standard across all programming schemes, but it's just not possible. Competition and more importantly profit is the key to making money and neither AMD or Nvidia will budge. Both parties are just as guilty as the other in this aspect. Reply
  • atlantico - Wednesday, October 15, 2014 - link

    Apple will *never* embrace CUDA. OpenCL is an important part of the future vision and strategy of Apple, whatever Nvidia is pushing, Apple is not buying. Reply
  • RussianSensation - Tuesday, October 7, 2014 - link

    If all Apple cares about was performance/watt, the MacPro would not feature AMD's Tahiti cores. There is a paragraph even dedicated to explaining the importance of OpenCL for Apple:

    "GPU computing with OpenCL.
    OpenCL lets you tap into the parallel computing power of modern GPUs and multicore CPUs to accelerate compute-intensive tasks in your Mac apps. Use OpenCL to incorporate advanced numerical and data analytics features, perform cutting-edge image and media processing, and deliver accurate physics simulations."
    https://www.apple.com/mac-pro/performance/

    Apple is known to switch between NV and AMD. Stating that AMD is not in good graces with Apple is ridiculous considering the MacPro has the less power efficient Tahiti vs. GK104s. And that is for a reason -- because HD7990 beats 690 at 4K, and destroys it in compute tasks -- which is proof performance/watt is no the only factor Apple looks at for their GPU selection.
    Reply
  • Omoronovo - Wednesday, October 8, 2014 - link

    I didn't mean to imply that it was *literally* the only factors taken into account; they clearly wouldn't use a GPU that cost $3,000 if a competing one with similar (but worse) performance/watt was $300.

    I was trying to emphasize that, all factors being equal - ie, standards compliance, compatibility, supply, etc, then performance/watt is the prime metric used to determine hardware choices. The tahiti vs GK104 comparison is a great one - AMD has extremely heavily pushed OpenCL and their support for it was basically unanimous - nVidia was slow on the uptake of OpenCL support as they were pushing for CUDA.
    Reply
  • bischofs - Tuesday, October 7, 2014 - link

    I may be wrong, but it seems like the only reason the mobile chips are catching up to the desktop is that they haven't really improved PC cards in 5+ years. Instead of pushing the limits on the PC, building architectures that are based on pure performance and not efficiency, and scaling it down they are doing the opposite, thus the performance difference is getting smaller. It is strange that they are marketing this as a good thing being that there is a rather large difference in power and cooling availability on a tower, thus there should be a large performance gap. Reply
  • Razyre - Tuesday, October 7, 2014 - link

    Not at all. Haiwaii shows this if anything, the 290X is balls to the walls in OpenCL, while Nvidia's cards are more conservative and gaming optimised they still pack an as good and usually better punch in frame rates.

    Cards are getting too hot at 2-300W, you need silly cooling solutions which are either expensive, make your card larger or louder.

    The Maxwell series is phenomenal; it drastically improves frame rates while halving the power consumption of the same series chips from 2 years ago.

    GPUs have come on SO far since 2009 when you are touting they've barely improved. Let's say you pit a 5870 against a 290X. The 7970 is about twice as powerful as the 5870 (slightly less in places), a 2012 GPU and the current 290X is about 30% better than a 7970. So you're effectively seeing there a theoretical 130% improvement in performance over 4 years (I say this because the 290X is now a year old), so that's an average of 30%ish improvement per year.

    Considering the massive R&D costs and costs associated with moving to smaller dies to fit more transistors on a chip (which increases heat, hence why Nvidia's Maxwell is a great leap since they can now jam way more transistors on there for the GK110 replacement) GPUs have come on leaps and bounds.

    The only reason it might look like they haven't is because instead of jumping from a standard let's say 1680x1050 to 1920x1080, we jumped to 3840x2160 a FOUR TIMES increase in resolution.

    Mobile GPUs are even more impressive in progress really. That chart showing the performance closing between mobile and desktop GPUs isn't too untrue.
    Reply
  • bischofs - Tuesday, October 7, 2014 - link

    I don't know much about the AMD stuff you are talking about, but I have probably more anecdotal evidence. Software and more importantly games for PCs have been pretty stagnant as far as resources go, I used a GTX 260 card for about 5 years and never had problems running anything until recently. Seeing as Games are the largest driver of innovation most games are built for consoles, with a large percentage also being built on mobile devices. Ive been playing games that look pretty much the same at 1080p for 5+ years on my pc, the only thing that has been added is more graphical features. Further support of my argument is processors, I remember the jump to Nehalem from the Core 2 was astounding, but from then on ( still running my i7-920 from 2008 ) its been lower power consumption and more cores with small changes in architecture. So you might throw some percentages around but I just don't see it. Reply

Log in

Don't have an account? Sign up now