Finally!

My oh my how time flies. It's been 4yrs since GPUs benefitted from a die-shrink. Time for MOAR transistors!

Original Thread
False Start

Teasers & announcements:
http://www.anandtech.com/show/9886/amd-reveals-polaris-gpu-architecture
http://www.anandtech.com/show/10145/amd-unveils-gpu-architecture-roadmap-after-polaris-comes-vega
http://www.techpowerup.com/218898/amd-demonstrates-revolutionary-14-nm-finfet-polaris-gpu-architecture.html
http://www.techpowerup.com/218883/4th-generation-graphics-corenext-architecture-codenamed-polaris.html
and
http://www.techpowerup.com/220372/nvidia-to-unveil-pascal-at-the-2016-computex.html

MOAR FLOPs!!!
http://www.techpowerup.com/220135/nvidia-gp100-silicon-to-feature-4-tflops-dpfp-performance.html


MOAR memory bandwidth:
https://devblogs.nvidia.com/parallelforall/nvlink-pascal-stacked-memory-feeding-appetite-big-data/
https://www.techpowerup.com/219527/techpowerup-news-amd-polaris-and-the-memory-war-between-hbm2-and-gddr5x.html


MOAR on-screen pixels for GPUs to push because of VR:
https://blogs.nvidia.com/blog/2016/03/30/vr-startups/
and new drivers:
http://www.anandtech.com/show/10195/oculus-rift-launch-day-news
http://www.geforce.com/whats-new/articles/oculus-rift-now-available-and-game-ready
https://community.amd.com/community/gaming/blog/2016/03/28/asynchronous-shaders-evolved

And a 4yr-old technology that is once again trying to go mainstream:

Is Thunderbolt 3 an opportunity for eGPUs (external GPUs) to shine, or do we have to wait a few more years?
http://www.anandtech.com/show/10137/razer-core-thunderbolt-3-egfx-chassis-499399-amd-nvidia-shipping-in-april
http://www.anandtech.com/show/10188/the-razer-blade-stealth-ultrabook-review
http://www.anandtech.com/show/10133/amd-xconnect-external-radeons


Pascal in the spotlight April 5th so stay tuned! Hopefully we get an ETA too ;)
http://www.gputechconf.com/

----

Further reading:
https://devblogs.nvidia.com/parallelforall/
https://blogs.nvidia.com/blog/2014/03/25/gpu-roadmap-pascal/
https://blogs.nvidia.com/blog/2015/03/17/pascal/

Official:

AMD expects shipments of Polaris architecture-based GPUs to begin in mid-2016

Still on track for a summer release...

also:
http://www.amd.com/en-us/innovations/software-technologies/25x20
https://youtu.be/5g3eQejGJ_A


(Hopefully NVIDIA's Pascal comes soon after. We'll probably have an official ETA this Tuesday.)

If you're willing to sit through a couple of hours of "Deep Learning & Artificial Intelligence", "Self-Driving Cars", and whatever else NVIDIA might be up to in case there's a Pascal announcement then feed goes live in about 5 hours:

https://blogs.nvidia.com/blog/2016/04/04/gtc-live/

NVIDIA Announces Tesla P100 Accelerator - Pascal GP100 Power for HPC

"GP100 is a whale of a GPU, measuring 610mm2 in die size on TSMC's 16nm FinFET process and composed of 15.3B transistors."

"With a boost clock of 1.48GHz, altogether Tesla P100 will offer 10.6 TFLOPS of FP32 performance or 5.3 TFLOPS of FP64 performance, more than doubling and tripling Tesla K40's rated throughput on these metrics respectively."

"Paired with the GP100 GPU on Tesla P100 is 16GB of HBM2 VRAM, laid out in 4 stacks for a 4096-bit memory bus. NVIDIA quotes P100 as offering 720GB/sec of memory bandwidth, which works out to a memory clock of 1.4Gbps. As we've seen with other HBM products, this marks a significant increase in memory bandwidth, more than doubling NVIDIA's last generation of cards" ***


*** Us code/app-challenged folk would love to hear what the volunteer devs have to say about this new memory and what it may (or may not) mean for S@H ;)

With a boost clock of 1.48GHz...

That'll probably get "fixed" to "1480MHz" soon (for continuity) ...

But the point is: last time we saw frequencies like those was Fermi and the 400/500 series. I wonder if shader clocks are back? Hmmm... I'll see if I can dig up any info.*

*Nope, sorry.

Tons more info below, 99.9% over my head ;)

Inside Pascal: NVIDIA’s Newest Computing Platform

CUDA 8 Features Revealed

Maybe our code-whisperers can translate? :P

PS Rounding up a bunch of rumors: a 980 successor (aka GP104) might be announced in May/July at around 300mm2 and GDDR5. Oh, and there are a few rumors of a GP102 floating around too (whatever that may be, if true)...

*** Us code/app-challenged folk would love to hear what the volunteer devs have to say about this new memory and what it may (or may not) mean for S@H ;)

Faster, better, memory subsystem generally makes optimisation easier/simpler, though a lot will depend on the balance of cumpute horsepower to memory bandwidth. Looks like for our purposes, might amount to a near linear double of current generation (for the HBM2 variant), presumably with the GP104 without HBM somewhere in between.

---

"Living by the wisdom of computer science doesn't sound so bad after all. And unlike most advice, it's backed up by proofs." -- Algorithms to live by: The computer science of human decisions.

From this article: Nvidia drops the Pascal bomb as a Tesla P100

Like Fiji, Nvidia is tapping HBM (High-Bandwidth Memory) for the P100, only they're using HBM2 instead of HBM1. The net result is four layers of stacked memory running on a 4096-bit bus, only the memory this time is running at 1.4Gbps instead of 1.0Gbps, yielding a total memory bandwidth of 720GB/s. That's all well and good, but perhaps more important than simply providing tons of memory bandwidth, HBM2 significantly increases the amount of memory per HBM stack, with P100 sporting a total of 16GB of VRAM. This was obviously a critical factor for Tesla cards, considering the older Tesla K40 already had 12GB of memory, and M40 likewise supports 12GB—not to mention the newly released M40 that comes with 24GB of GDDR5. HBM2 also includes "free" ECC protection, which is a plus for professional applications where reliability and accuracy are paramount.

Thanks to the move to the 16nm FinFET process technology, Nvidia has also been able to substantially increase the number of transistors in the GPU core. Where GM200 in the M40 has 3072 CUDA cores and consists of eight billion transistors, P100 nearly doubles transistor counts to 15.3 billion. Nvidia also noted that this is their largest GPU ever, measuring 610mm2, but while that's impressive, GM200 also measured around 600mm2, so that aspect hasn't changed too much. That size does not include the silicon interposer, however, which has to cover the area of both the GPU as well as the HBM2 chips, so this definitely qualifies as a gargantuan chip. If you count all the transistors in the GPU, interposer, and HBM2 modules, Nvidia says there are 150 billion transistors all told.

What about core counts? Here's where things get a bit interesting. The Pascal architecture has once again evolved, changing the SM module size. In Kepler, a single SMX consisted of 192 CUDA cores, with the GK110 supporting up to 28 SMX units for 2880 CUDA cores total. Maxwell dropped the core count to 128 per SM, but the architecture was built to better utilize each core, leading to improved efficiency. In Pascal P100, Nvidia drops to just 64 CUDA cores per SM, and apparently there are further improvements to efficiency. What's interesting to note is that each SM in the P100 has 64 FP32 cores, along with 32 FP64 cores, and P100 also adds support for half-precision FP16, potentially doubling throughput in situations where raw performance takes priority over precision.

A fully enabled P100 has 60 SMs, giving a potential 3840 cores, but Tesla P100 disables four SMs to give 3584 total cores. That might sound like only a small step forward, considering the M40 has 3072 cores, but clock speeds have improved. Where M40 runs at 948-1114MHz, P100 can run at 1328-1480MHz. Raw compute power ends up being 21.2 half-precision FP16 TFLOPS, 10.6 single-precision FP32 TFLOPS, or 5.3 double-precision FP64 TFLOPS. M40 by comparison had half- and single-precision rates of 6.8 TFLOPS, but double precision rates of just 213 GFLOPS; that's because GM200 only included four FP64 cores per SMM, a significant departure from the GK110 Kepler architecture.

What all this means is that P100 may never be utilized in a mainstream consumer device. At best, I suspect we might see some new variant of Titan based off P100 in the future, but that could be a long way off. You see, even though Nvidia is spilling the beans on Tesla P100 today—or at least, some of the beans—and the chips are in volume production, Nvidia doesn't plan on full retail availability from OEMs until Q1'2017. That means we're far more likely to see a GP104 chip that skips all the ECC, HBM2, and FP64 stuff and potentially stuffs more FP32 cores into a smaller die than P100. Sadly, Nvidia is not commenting on any future consumer facing products at this time. Looks like we'll have to wait for Computex to hear more about the consumer lines.

So, maybe we will or maybe we won't get the full monty? Wonder what exactly this all means for those of us crunching here at SETI, especially with the now 3 different types of FP precision available? Major impact, or just an evolutionary step along the way? And love to hear what the pricing will be for the consumer level card, if/when they release one that is pretty much fully enabled?

---

*** Us code/app-challenged folk would love to hear what the volunteer devs have to say about this new memory and what it may (or may not) mean for S@H ;)

Faster, better, memory subsystem generally makes optimisation easier/simpler, though a lot will depend on the balance of cumpute horsepower to memory bandwidth. Looks like for our purposes, might amount to a near linear double of current generation (for the HBM2 variant), presumably with the GP104 without HBM somewhere in between.

Would you expect that to take advantage of HBM a specialized app would be required?
On the Radeon side comparing my R9 390X with GDDR5 to the Fury X with HBM I'm not impressed with the FuryX's run times. With 45% more cores & SP GFLOPS it only manages to run MB tasks 20% faster.

---

SETI@home classic workunits: 93,865 CPU time: 863,447 hours
Join the [url=http://tinyurl.com/8y46zvu]BP6/VP6 User Group[

So, maybe we will or maybe we won't get the full monty? Wonder what exactly this all means for those of us crunching here at SETI, especially with the now 3 different types of FP precision available? Major impact, or just an evolutionary step along the way? And love to hear what the pricing will be for the consumer level card, if/when they release one that is pretty much fully enabled?

It all depends what you mean by "not getting the full monty" - you will if you (or I) is prepared to pay for it, but it is almost certain that the consumer/domestic devices will be somewhat restricted.

The availability of three types of Floating Point precision will make very little difference to SETI as it only uses single precision, although for some other projects this could be a major selling point. One thing to remember is that generally higher precision results in lower processing rates - even when Double Precision is available "on chip".

One can but assume that the pricing will be comparable to the current levels of gtx9xx devices, with the usual "launch premium" for early adopters.

One thing that will be interesting to keep an eye out for would be the Pascal equivalent to the GTX750, that could be a very cost effective card, given the talk of "xx% more per watt".

---

Bob Smith
Member of Seti PIPPS (Pluto is a Planet Protest Society)
Somewhere in the (un)known Universe?

Yeah, it's always been if you want the nicest toys, you'd best have a thich wallet. I guess what I was referring to about the full monty comment is the statement

A fully enabled P100 has 60 SMs, giving a potential 3840 cores, but Tesla P100 disables four SMs to give 3584 total cores.

Maybe sometime down the road they'll release a fully enabled version, maybe when they get their yields up? Regardless, if they do, you know it won't be cheap.. :-)

---

Yeah, it's always been if you want the nicest toys, you'd best have a thich wallet. I guess what I was referring to about the full monty comment is the statement

A fully enabled P100 has 60 SMs, giving a potential 3840 cores, but Tesla P100 disables four SMs to give 3584 total cores.

Maybe sometime down the road they'll release a fully enabled version, maybe when they get their yields up? Regardless, if they do, you know it won't be cheap.. :-)

Perhaps on the enterprise end they reduce the number used in the event parts of it burn out. Similar to how SSD drives have more flash than is usable. So they can allocate new blocks when flash cells go dead.

---

SETI@home classic workunits: 93,865 CPU time: 863,447 hours
Join the [url=http://tinyurl.com/8y46zvu]BP6/VP6 User Group[

*** Us code/app-challenged folk would love to hear what the volunteer devs have to say about this new memory and what it may (or may not) mean for S@H ;)

Faster, better, memory subsystem generally makes optimisation easier/simpler, though a lot will depend on the balance of cumpute horsepower to memory bandwidth. Looks like for our purposes, might amount to a near linear double of current generation (for the HBM2 variant), presumably with the GP104 without HBM somewhere in between.

Would you expect that to take advantage of HBM a specialized app would be required?
On the Radeon side comparing my R9 390X with GDDR5 to the Fury X with HBM I'm not impressed with the FuryX's run times. With 45% more cores & SP GFLOPS it only manages to run MB tasks 20% faster.

In a general sense no app specialised for the memory type should be necessary (or worth it in practical senses). What is slowly happening though, is more codepaths being added, along with self tuning and dispatch, such that optional automatic scaling can take place.

I don't know much about the current OpenCL apps or AMDs various implementations, though for the NV-Cuda situation there are substantial portions that are (video) memory bound and PCIe bound. In the first case HBM should help, while the second self scaling and Petri's streaming, giving latency hiding, will help.

With Generic code on current hardware yielding ~5% compute efficiency ( Flop/TheoreticalPeakFlops ), a bit more total running multiple instances, the other ~95% is various latencies and bandwidth restriction. I'm fairly confident over time we'll approach 20% (With petri's experimental code reaching ~10% now), at least on the next, current, and maybe back to Fermi class with some diminishing returns.

---

"Living by the wisdom of computer science doesn't sound so bad after all. And unlike most advice, it's backed up by proofs." -- Algorithms to live by: The computer science of human decisions.

Hey Jason, and anyone else who knows much about this, I just started a thread about OpenCL vs. CUDA, if people who are in the know could add comments to it, that would be helpful in understanding the current situation.

---

http://www.pcworld.com/article/3053600/hardware/apples-aging-mac-pro-is-falling-way-behind-windows-rivals.html?token=%23tk.PCW_nlt_pcw_bestof_html_2016-04-08&idg_eid=a0e1f7b61040fb71f0dec4f3f7f091c8&utm_source=Sailthru&utm_medium=email&utm_campaign=Best%20of%20PCWorld%202016-04-08&utm_term=pcw_bestof_html#tk.PCW_nlt_pcw_bestof_html_2016-04-08

Just to add a bit to what is coming (and perhaps going).

---

SETI@home classic workunits 4,019
SETI@home classic CPU time 34,348 hours

OK so more rumors flying around that a mainstream GP104 chip will get announced (actual availability unknown) at Computex 2016 (from 31st to June 4th).

A couple of pictures of the chip are floating around which superficially look authentic:

http://www.fudzilla.com/news/graphics/40515-nvidia-gtx-1070-gp104-200-pascal-asic-pictured
http://videocardz.com/59266/nvidia-pascal-gp104-gpu-pictured-up-close

Even if bogus it would still make sense for the GP104 to be in the ~300mm2 range so here's a... brief history in time :)

- GeForce 9800 GTX - 324mm2 - 140W
- GeForce GTS 250 - 260mm2 - 150W
- GeForce GTX 460 - 332mm2 - 150W or 160W (2 versions)
- GeForce GTX 560 Ti - 332mm2 - 170W
- GeForce GTX 680 - 294mm2 - 195W
- GeForce GTX 770 - 294mm2 - 230W
- GeForce GTX 960 - 227mm2 - 120W
- GeForce GTX 980 - 398mm2 - 165W

So while we wait, I was thinking maybe we could see how these cards behave @Seti.
I'll probably have better luck searching for a unicorn but anybody have numbers for v8 (or v7 if not) on stock app doing one WU at a time?
Any other ideas for an apples-to-apples comparison?

For reference my little ION GPU does a "normal" ~0.43 Angle Range v8 WU in about 3 to 3 and 1/2 hours. Poor thing has a 12W TDP :)

As always, above tech specs from:
https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units

Any other ideas for an apples-to-apples comparison?

The problem is the lack of an application that makes use of the new hardware's abilities.
My GTX 750Tis are about on par with the old GTX 460 and a bit behind the GTX560Ti because of the changes in architecture that result in the current applications being nowhere nearly as effective on the new hardware as they were on the old.
As far as power consumptions is concerned they leave the old hardware for dead.
And I suspect that with an application that takes advantage of the new hardware, and is able to work around it's quirks (and particularly the current drivers) it's performance would leave the old hardware far behind as well.

---

Grant
Darwin NT

Any other ideas for an apples-to-apples comparison?

The problem is the lack of an application that makes use of the new hardware's abilities.
My GTX 750Tis are about on par with the old GTX 460 and a bit behind the GTX560Ti because of the changes in architecture that result in the current applications being nowhere nearly as effective on the new hardware as they were on the old.
As far as power consumptions is concerned they leave the old hardware for dead.
And I suspect that with an application that takes advantage of the new hardware, and is able to work around it's quirks (and particularly the current drivers) it's performance would leave the old hardware far behind as well.

It's going to be an interesting few months leading up to Pascal.

Seeing as we have a lot of new code going in, it's starting to look like some precautions/safeties, at least for default/stock operation are going to be warranted to make sure stuff doesn't break in the middle of all the tech changes going on (Drivers, GPUs, OSes, and seti processing all changing )

Having to supply these things with Fire extinguishers and miniature fusion power plants probably isn't going to work, lol.

---

"Living by the wisdom of computer science doesn't sound so bad after all. And unlike most advice, it's backed up by proofs." -- Algorithms to live by: The computer science of human decisions.

changes in architecture that result in the current applications being nowhere nearly as effective on the new hardware as they were on the old.

I could be overoptimistic or oversimplifying or both but:

For now I can't think of (or better, don't know of) a reason that couldn't change overnight. If Pascal's architecture is closer to Fermi and the 460 or 560Ti then I'm guessing the Pascal drivers could roll back into Seti app favor. That's not to say that the old apps necessarily work "out the gate", rather that they happen to be better optimized for Pascal rather than Kepler.

If that makes sense :)

(I kinda got the impression, in hindsight, that Fermi architecture was 100% compute oriented and then intentionally watered down over generations for the sake of gaming & efficiency. Pascal looks to be the same if GeForce cards end up anything like the Tesla that was announced. In other words this could be Fermi 2.0 and it again gets watered down over the generations. It looks like compute gets the advantage this round. Reminder: you guys notice that jump in frequencies on Pascal?)

I'll probably have better luck searching for a unicorn but anybody have numbers for v8 (or v7 if not) on stock app doing one WU at a time?

That was poorly written. What I meant was anybody have v8 times for any of these:

- GeForce 9800 GTX - 324mm2 - 140W
- GeForce GTS 250 - 260mm2 - 150W
- GeForce GTX 460 - 332mm2 - 150W or 160W (2 versions)
- GeForce GTX 560 Ti - 332mm2 - 170W
- GeForce GTX 680 - 294mm2 - 195W
- GeForce GTX 770 - 294mm2 - 230W
- GeForce GTX 960 - 227mm2 - 120W
- GeForce GTX 980 - 398mm2 - 165W