Designing GP104: Running Up the Clocks

So if GP104’s per-unit throughput is identical to GM204, and the SM count has only been increased from 2048 to 2560 (25%), then what makes GTX 1080 60-70% faster than GTX 980? The answer there is that instead of vastly increasing the number of functional units for GP104 or increasing per-unit throughput, NVIDIA has instead opted to significantly raise the GPU clockspeed. And this in turn goes back to the earlier discussion on TSMC’s 16nm FinFET process.

With every advancement in fab technology, chip designers have been able to increase their clockspeeds thanks to the basic physics at play. However because TSMC’s 16nm node adds FinFETs for the first time, it’s extra special. What’s happening here is a confluence of multiple factors, but at the most basic level the introduction of FinFETs means that the entire voltage/frequency curve gets shifted. The reduced leakage and overall “stronger” FinFET transistors can run at higher clockspeeds at lower voltages, allowing for higher overall clockspeeds at the same (or similar) power consumption. We see this effect to some degree with every node shift, but it’s especially potent when making the shift from planar to FinFET, as has been the case for the jump from 28nm to 16nm.

Given the already significant one-off benefits of such a large jump in the voltage/frequency curve, for Pascal NVIDIA has decided to fully embrace the idea and run up the clocks as much as is reasonably possible. At an architectural level this meant going through the design to identify bottlenecks in the critical paths – logic sections that couldn’t run at as high a frequency as NVIDIA would have liked – and reworking them to operate at higher frequencies. As GPUs typically (and still are) relatively low clocked, there’s not as much of a need to optimize critical paths in this matter, but with NVIDIA’s loftier clockspeed goals for Pascal, this changed things.

From an implementation point of view this isn’t the first time that NVIDIA has pushed for high clockspeeds, as most recently the 40nm Fermi architecture incorporated a double-pumped shader clock. However this is the first time NVIDIA has attempted something similar since they reined in their power consumption with Kepler (and later Maxwell). Having learned their lesson the hard way with Fermi, I’m told a lot more care went into matters with Pascal in order to avoid the power penalties NVIDIA paid with Fermi, exemplified by things such as only adding flip-flops where truly necessary.

Meanwhile when it comes to the architectural impact of designing for high clockspeeds, the results seem minimal. While NVIDIA does not divulge full information on the pipeline of a CUDA core, all of the testing I’ve run indicates that the latency (in clock cycles) of the CUDA cores is identical to Maxwell. Which goes hand in hand with earlier observations about throughput. So although optimizations were made to the architecture to improve clockspeeds, it doesn’t look like NVIDIA has made any more extreme optimizations (e.g. pipeline lengthening) that detectably reduces Pascal’s per-clock performance.

Beyond3D Suite - Estimated MADD Latency

Finally, more broadly speaking, while this is essentially a one-time trick for NVIDIA, it’s an interesting route for them to go. By cranking up their clockspeeds in this fashion, they avoid any real scale-out issues, at least for the time being. Although graphics are the traditional embarrassingly parallel problem, even a graphical workload is subject to some degree of diminishing returns as GPUs scale farther out. A larger number of SMs is more difficult to fill, not every aspect of the rendering process is massively parallel (shadow maps being a good example), and ever-increasing pixel shader lengths compound the problem. Admittedly NVIDIA’s not seeing significant scale-out issues quite yet, but this is why GTX 980 isn’t quite twice as fast as GTX 960, for example.

Just increasing the clockspeed, comparatively speaking, means that the entire GPU gets proportionally faster without shifting the resource balance; the CUDA cores are 43% faster, the geometry frontends are 43% faster, the ROPs are 43% faster, etc. The only real limitation in this regard isn’t the GPU itself, but whether you can adequately feed it. And this is where GDDR5X comes into play.

FP16 Throughput on GP104: Good for Compatibility (and Not Much Else) Feeding Pascal: GDDR5X
POST A COMMENT

200 Comments

View All Comments

  • TestKing123 - Wednesday, July 20, 2016 - link

    Then you're woefully behind the times since other sites can do this better. If you're not able to re-run a benchmark for a game with a pretty significant patch like Tomb Raider, or a high profile game like Doom with a significant performance patch like Vulcan that's been out for over a week, then you're workflow is flawed and this site won't stand a chance against the other crop. I'm pretty sure you're seeing this already if you have any sort of metrics tracking in place. Reply
  • TheinsanegamerN - Wednesday, July 20, 2016 - link

    So question, if you started this article on may 14th, was their no time in the over 2 months to add one game to that benchmark list? Reply
  • nathanddrews - Wednesday, July 20, 2016 - link

    Seems like an official addendum is necessary at some point. Doom on Vulkan is amazing. Dota 2 on Vulkan is great, too (and would be useful in reviews of low end to mainstream GPUs especially). Talos... not so much. Reply
  • Eden-K121D - Thursday, July 21, 2016 - link

    Talos Principle was a proof of concept Reply
  • ajlueke - Friday, July 22, 2016 - link

    http://www.pcgamer.com/doom-benchmarks-return-vulk...

    Addendum complete.
    Reply
  • mczak - Wednesday, July 20, 2016 - link

    The table with the native FP throughput rates isn't correct on page 5. Either it's in terms of flops, then gp104 fp16 would be 1:64. Or it's in terms of hw instruction throughput - then gp100 would be 1:1. (Interestingly, the sandra numbers for half-float are indeed 1:128 - suggesting it didn't make any use of fp16 packing at all.) Reply
  • Ryan Smith - Wednesday, July 20, 2016 - link

    Ahh, right you are. I was going for the FLOPs rate, but wrote down the wrong value. Thanks!

    As for the Sandra numbers, they're not super precise. But it's an obvious indication of what's going on under the hood. When the same CUDA 7.5 code path gives you wildly different results on Pascal, then you know something has changed...
    Reply
  • BurntMyBacon - Thursday, July 21, 2016 - link

    Did nVidia somehow limit the ability to promote FP16 operations to FP32? If not, I don't see the point in creating such a slow performing FP16 mode in the first place. Why waste die space when an intelligent designer can just promote the commands to get normal speeds out of the chip anyways? Sure you miss out on speed doubling through packing, but that is still much better than the 1/128 (1/64) rate you get using the provided FP16 mode. Reply
  • Scali - Thursday, July 21, 2016 - link

    I think they can just do that in the shader compiler. Any FP16 operation gets replaced by an FP32 one.
    Only reading from buffers and writing to buffers with FP16 content should remain FP16. Then again, if their driver is smart enough, it can even promote all buffers to FP32 as well (as long as the GPU is the only one accessing the data, the actual representation doesn't matter. Only when the CPU also accesses the data, does it actually need to be FP16).
    Reply
  • owan - Wednesday, July 20, 2016 - link

    Only 2 months late and published the day after a different major GPU release. What happened to this place? Reply

Log in

Don't have an account? Sign up now