The NVIDIA GeForce GTX 1080 & GTX 1070 Founders Editions Review: Kicking Off the FinFET Generation
by Ryan Smith on July 20, 2016 8:45 AM ESTDesigning 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.
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.
Facebook
Twitter
RSS
200 Comments
View All Comments
Ranger1065 - Thursday, July 21, 2016 - link
Your unwavering support for Anandtech is impressive.I too have a job that keeps me busy, yet oddly enough I find the time to browse (I prefer that word to "trawl") a number of sites.
I find it helps to form objective opinions.
I don't believe in early adoption, but I do believe in getting the job done on time, however if you are comfortable with a 2 month delay, so be it :)
Interesting to note that architectural deep dives concern your art and media departments so closely in their purchasing decisions. Who would have guessed?
It's true (God knows it's been stated here often enough) that
Anandtech goes into detail like no other, I don't dispute that.
But is it worth the wait? A significant number seem to think not.
Allow me to leave one last issue for you to ponder (assuming you have the time in your extremely busy schedule).
Is it good for Anandtech?
catavalon21 - Thursday, July 21, 2016 - link
Impatient as I was at the first for benchmarks, yes, I'm a numbers junkie, since it's evident precious few of us will have had a chance to buy one of these cards yet (or the 480), I doubt the delay has caused anyone to buy the wrong card. Can't speak for the smart phone review folks are complaining about being absent, but as it turns out, what I'm initially looking for is usually done early on in Bench. The rest of this, yeah, it can wait.mkaibear - Saturday, July 23, 2016 - link
Job, house, kids, church... more than enough to keep me sufficiently busy that I don't have the time to browse more than a few sites. I pick them quite carefully.Given the lifespan of a typical system is >5 years I think that a 2 month delay is perfectly reasonable. It can often take that long to get purchasing signoff once I've decided what they need to purchase anyway (one of the many reasons that architectural deep dives are useful - so I can explain why the purchase is worthwhile). Do you actually spend someone else's money at any point or are you just having to justify it to yourself?
Whether or not it's worth the wait to you is one thing - but it's clearly worth the wait to both Anandtech and to Purch.
razvan.uruc@gmail.com - Thursday, July 21, 2016 - link
Excellent article, well deserved the wait!giggs - Thursday, July 21, 2016 - link
While this is a very thorough and well written review, it makes me wonder about sponsored content and product placement.The PG279Q is the only monitor mentionned, making sure the brand appears, and nothing about competing products. It felt unnecessary.
I hope it's just a coincidence, but considering there has been quite a lot of coverage about Asus in the last few months, I'm starting to doubt some of the stuff I read here.
Ryan Smith - Thursday, July 21, 2016 - link
"The PG279Q is the only monitor mentionned, making sure the brand appears, and nothing about competing products."There's no product placement or the like (and if there was, it would be disclosed). I just wanted to name a popular 1440p G-Sync monitor to give some real-world connection to the results. We've had cards for a bit that can drive 1440p monitors at around 60fps, but GTX 1080 is really the first card that is going to make good use of higher refresh rate monitors.
giggs - Thursday, July 21, 2016 - link
Fair enough, thank you for responding promptly. Keep up the good work!arh2o - Thursday, July 21, 2016 - link
This is really the gold standard of reviews. More in-depth than any site on the internet. Great job Ryan, keep up the good work.Ranger1065 - Thursday, July 21, 2016 - link
This is a quality article.timchen - Thursday, July 21, 2016 - link
Great article. It is pleasant to read more about technology instead of testing results. Some questions though:1. higher frequency: I am kind of skeptical that the overall higher frequency is mostly enabled by FinFET. Maybe it is the case, but for example when Intel moved to FinFET we did not see such improvement. RX480 is not showing that either. It seems pretty evident the situation is different from 8800GTX where we first get frequency doubling/tripling only in the shader domain though. (Wow DX10 is 10 years ago... and computation throughput is improved by 20x)
2. The fastsync comparison graph looks pretty suspicious. How can Vsync have such high latency? The most latency I can see in a double buffer scenario with vsync is that the screen refresh just happens a tiny bit earlier than the completion of a buffer. That will give a delay of two frame time which is like 33 ms (Remember we are talking about a case where GPU fps>60). This is unless, of course, if they are testing vsync at 20hz or something.