Nvidia Pascal Speculation Thread

Status
Not open for further replies.

Yeah the Drive PX2 is definitely a more serious bit of hardware-software with the jump up in its AlexNet capability in comparison to the Tegra X1 and worthy of serious recognition, NVIDIA seem really serious about breaking into this market in a successful way, and I can see other uses for this beyond cars with use in other transport systems as part of the warning/protection/assist systems and also possibly military projects (have my doubts NVIDIA can break into this).
Cheers
 
Now that I know what a DL OP is, I'll probably never acknowledge it again.

Well Deep Learning concept can be fun.
Here is an example of a simple neural network learning to play Mario Brothers; it starts from nothing and standing still to 24 hours later where it could beat the level in a very efficient way.
Fun to watch and listen to.
Audi commented using the complete NVIDIA solution (includes the DIGITS developer-platform box) ; "in less than four hours we achieved over 96 percent accuracy using Ruhr University Bochum’s traffic sign database.While others invested years of development to achieve similar levels of perception with classical computer vision algorithms "
Cheers
 
Meanwhile it seems we have been slightly misled about the PX2 as there are strong indications that the discrete GPUs on the shown PCB are actually GM204, as the MXM package looks exactly as a GTX 980M. Also the inscriptions on the die, and it's size, indicate this :(
 
Well Deep Learning concept can be fun.
Here is an example of a simple neural network learning to play Mario Brothers; it starts from nothing and standing still to 24 hours later where it could beat the level in a very efficient way.
Fun to watch and listen to.
Audi commented using the complete NVIDIA solution (includes the DIGITS developer-platform box) ; "in less than four hours we achieved over 96 percent accuracy using Ruhr University Bochum’s traffic sign database.While others invested years of development to achieve similar levels of perception with classical computer vision algorithms "
Cheers

Oh man I remember when that came out. The site it came from got a boost in members.

Still an amazing video.
 
Meanwhile it seems we have been slightly misled about the PX2 as there are strong indications that the discrete GPUs on the shown PCB are actually GM204, as the MXM package looks exactly as a GTX 980M. Also the inscriptions on the die, and it's size, indicate this :(

Sigh. It's a mock-up. :rolleyes:

PX2 won't be available till year end and it will be all Pascal then.
 
Dumb question—how is it that Drive PX 2 has 6x the ImageNet performance of Titan X when it only has ~3x the "deep learning ops" performance?
On which actual hardware performance metric (ram, flops, dl-flops) does it have 6x over Titan X?
 
Dumb question—how is it that Drive PX 2 has 6x the ImageNet performance of Titan X when it only has ~3x the "deep learning ops" performance?
On which actual hardware performance metric (ram, flops, dl-flops) does it have 6x over Titan X?
something related to their new mysterious DL TOPS thing (Deep Learning TeraOPS) on Pascal, where 1 TFLOPS = 3 DL TOPS. How they do this trick is still a secret story but if it's true, they will have a huge advantage in this generation...
 
something related to their new mysterious DL TOPS thing (Deep Learning TeraOPS) on Pascal, where 1 TFLOPS = 3 DL TOPS. How they do this trick is still a secret story but if it's true, they will have a huge advantage in this generation...

But it doesn't add up even with "deep learning ops". PX 2 has ~3x more "deep learning OPS" vs. Titan X but it has ~6x more AlexNet performance. There'a another magical "2x" something going on.
 
But it doesn't add up even with "deep learning ops". PX 2 has ~3x more "deep learning OPS" vs. Titan X but it has ~6x more AlexNet performance. There'a another magical "2x" something going on.

Edited my terrible response. I don't think I'm awake yet.

Looks like no one knows.
Ryan Smith said:
Curiously, NVIDIA also used the event to introduce a new unit of measurement – the Deep Learning Tera-Op, or DL TOPS – which at 24 is an unusual 3x higher than PX 2’s FP32 performance. Based on everything disclosed by NVIDIA about Pascal so far, we don’t have any reason to believe FP16 performance is more than 2x Pascal’s FP32 performance. So where the extra performance comes from is a mystery at the moment. NVIDIA quoted this and not FP16 FLOPS, so it may include a special case operation (ala the Fused Multiply-Add), or even including the performance of the Denver CPU cores.
 
Last edited:
But it doesn't add up even with "deep learning ops". PX 2 has ~3x more "deep learning OPS" vs. Titan X but it has ~6x more AlexNet performance. There'a another magical "2x" something going on.
Then there's the Titan X specs, which presumably destroy those of a Tegra X1 - doesn't a TX1 have about 1TF of FP16 performance? But the Titan X AlexNet score barely beats that of a Tegra X1.
 
I found some info regarding the mysterious Deep Learning OPs.
Apparently Pascal can do processing on 8-bit integers.
It's easy to understand that can be up to 4 times faster compared to 32-bit floating point operations.
See: http://www.eetimes.com/author.asp?section_id=36&doc_id=1328609&
"The NVIDIA deep learning algorithms can use specialized mixed precision instructions as low as 8-bit integer to deliver up to 24 trillion operations per second. The 8-bit integer operations are new in the Pascal GPU."
 
Then there's the Titan X specs, which presumably destroy those of a Tegra X1 - doesn't a TX1 have about 1TF of FP16 performance? But the Titan X AlexNet score barely beats that of a Tegra X1.
Where are you getting that from? Last year CES slides showed Tegra X1 having 30 images/sec for AlexNet. This year's slides show 450 images/sec for Titan X. That's 15x delta, on par with FP32 perf delta.
 
Where are you getting that from? Last year CES slides showed Tegra X1 having 30 images/sec for AlexNet. This year's slides show 450 images/sec for Titan X. That's 15x delta, on par with FP32 perf delta.

Post 541 in this thread. I guess marketing slide to marketing slide is more apples to apples... But 15x doesn't make much sense either, as X1 is supposed to have twice the FP16 throughput as FP32 ...
 
But it doesn't add up even with "deep learning ops". PX 2 has ~3x more "deep learning OPS" vs. Titan X but it has ~6x more AlexNet performance. There'a another magical "2x" something going on.
Reasearch AlexNet (papers on it out there), notice one already provided a link with measurements for the Tegra JTX1 earlier in comparison to Intel 6700k at phoronix.
It is still not ideal (better than relying on say just TFLOPs though due to its specific focus) but not something NVIDIA are making out of thin air; here is a paper to do with it: http://www.cs.toronto.edu/~fritz/absps/imagenet.pdf
Worth noting NVIDIA are talking about images processed/second for context, accuracy is more to do with their DIGITS platform.

Cheers
 
Status
Not open for further replies.
Back
Top