NVIDIA Pascal Thread

[CakeMonster]

320GB/s

I seem to remember OC'ing the mem of both my 7970 and 780 past 300GB/s. Why aren't we at higher numbers in 2016?

What are you talking about, we wont' have another graphic card for another year, or am I missing something?

Oh, we definitely will. They won't let the >$600 market go untapped for long. However, if the obscene price tag of the next chip is out of your comfort zone anyway, you might as well upgrade now. Doesn't matter how good Big(ger) Pascal is if its that's expensive.

 

[jpiniero]

BTW AT confirms that the transistor count is 7.2B. Which is quite a bit less than GM200 (8B).

64 ROPs is definately going to hurt it's performance vs GM200 at high resolutions, and could also explain why it's only 20ish% faster despite having 50% more compute power.

 

[sontin]

64 ROPs is definately going to hurt it's performance vs GM200 at high resolutions, and could also explain why it's only 20ish% faster despite having 50% more compute power.

It doesnt have 50% more compute performance over TitanX. It is 34% higher and there are no real benchmarks out.

 

[jpiniero]

It doesnt have 50% more compute performance over TitanX. It is 34% higher and there are no real benchmarks out.

I guess there is some confusion about that, since AT says 9 TF but Wikipedia says 8.2.

 

[Silverforce11]

Less ROPs but the clock differential (more front-end throughput) makes up for it, the 1080 will perform just fine at 4K.

It's the 1070 that will fall off at higher resolution, due to GDDR5.

 

[nvgpu]

9TF is the 1773MHz Boost clock numbers and Nvidia always uses that on geforce.com spec page, Wikipedia for whatever reason always uses the base clock numbers.

 

[IndyColtsFan]

Looks like a pretty nice card even with the $600 price tag. Feel bad for the old 980 Ti owner though which has been out to market for under a year and already severely obsolete. The 3X VR claim seems to be outrageous and I wonder what kind of measurement they're using but everything else is a stunning upgrade..just wow. For 180 watts...my PS is 1,000 watts... -_- AMD is going the low power route I was hoping Nvidia crank out something not so conservative...still...here's to hope it overclocks beyond 300 watts!

I'm not seeing the 980Ti as being "severely obsolete" at all, especially if the 25% figure is true. Remember, that is about the average performance gap between Skylake and Sandy Bridge and few would call Sandy Bridge obsolete at this point. Yeah, the 1080 appears to be more power efficient but that is something way down the list of priorities in a gaming system IMO.

I'll wait for benchmarks first before making any pronouncement, but I bought my 980Ti knowing Pascal was around the corner because I got a good deal and don't anticipate moving to 4K for quite some time. Unless I find a superb deal on Broadwell E, I don't anticipate a major system upgrade until Cannonlake or Skylake E and at that time, will look at a new GPU.

 

[JDG1980]

Roy Taylor... Roy Taylor... aaah THE Roy Taylor who also said :

I don't think Nvidia is going to do anything to increase the TAM, because according to everything we've seen around Pascal, it's a high-end part. I don't know what the price is gonna be, but let's say it's as low as £500/$600 and as high as £800/$1000. That price range is not going to expand the TAM for VR.

source: http://arstechnica.com/gadgets/2016/04/amd-polaris-will-be-a-mainstream-gpu/

Well... technically, he's right about the TAM part. The GTX 970 was already "VR Ready" by the metric AMD is using (SteamVR benchmark), and GTX 1070 will be slightly more expensive than that. Therefore, this wave of GP104-based Pascal cards won't "expand the TAM for VR" since they won't bring "VR Ready" performance to a lower price point.

He guessed wrong on the specific Pascal prices, but he did admit that was just speculation on his part.

My goal is not to start a flame war but only to show people that these PRs guys (green or red team) can't be trusted and we must take their blurb with huge bag of salt

Agreed. More specifically, the marketing staff usually won't tell direct lies, but they will tell selective truths and paint everything in the light that best suits their side.

 

[Cookie Monster]

So if GP104 is a 40SM part (and assuming with much lower DP rate e.g 1/8 or 1/16), I wonder if GP102 will be 60 or a 80SM part (at 384bit - GDDR5X) giving us a 5120 CC GPU since the chip will have roughly a 15B transistor budget and GK104 is only 7.2B.

 

[jpiniero]

The 1060 Ti rumors being OEM only and comparable to 970 in performance is probably true now, given they will need something that runs on the bus power only for OEMs and VR.

So if GP104 is a 40SM part (and assuming with much lower DP rate e.g 1/8 or 1/16), I wonder if GP102 will be 60 or a 80SM part (at 384bit - GDDR5X) giving us a 5120 CC GPU since the chip will have roughly a 15B transistor budget and GK104 is only 7.2B.

GP102 is likely 50 SM if it exists (3200)

 

[Cookie Monster]

The 1060 Ti rumors being OEM only and comparable to 970 in performance is probably true now, given they will need something that runs on the bus power only for OEMs and VR.



GP102 is likely 50 SM if it exists (3200)

If thats true, the die size would be quite small for a supposed high end parts. Interesting given their history of releasing ~600mm2 GPUs.

 

[sontin]

Well... technically, he's right about the TAM part. The GTX 970 was already "VR Ready" by the metric AMD is using (SteamVR benchmark), and GTX 1070 will be slightly more expensive than that. Therefore, this wave of GP104-based Pascal cards won't "expand the TAM for VR" since they won't bring "VR Ready" performance to a lower price point.

This doesnt make sense. You just need to lower the quality of content and your TAM is expanding. With Pascal nVidia is increasing the TAM in a huge way: Offering Fury Duo Pro performance for 1/4 of the price with twice the memory and 1/3 of the power consumption. No VR player cares about the $100 difference between a much slower Polaris 10 and a GTX1070 card when they need to pay more than $600 for a headset.

He guessed wrong on the specific Pascal prices, but he did admit that was just speculation on his part.

He used fake numbers to proof a nonexistent point. It was not speculation.

 

[jpiniero]

If thats true, the die size would be quite small for a supposed high end parts. Interesting given their history of releasing ~600mm2 GPUs.

They do have a 600 mm2 die - GP100.

 

[Sweepr]

Hidden Information Found In HTML Source Code of NVIDIA’s GTX 1080 Page

A considerable amount of information seems to be hidden away in the html source code of NVIDIA’s GTX 1080 page. After doing some digging through the code, I found a few interesting pieces of information that were commented out:

The benchmarks shown on the performance tab were conducted at 2560×1600 max settings(hidden in a commented out paragraph element)
The new NVIDIA SLI bridge will be available for preorder, but are limited to 1 per customer(entire commented out preorder section)
The SLI bridge will also be available at Amazon, Microcenter, and FRYS.
Unfortunately, the price is commented out and listed as “$XXX.XX”.
Maximum VGA resolution of 2048×1536(likely commented out because few people care about this except maybe CS:GO enthusiasts with CRTs)
Texture fill rate of 176 Gigatextels/sec
The “Up to 3x performance” originally had a disclaimer that it was “based on test results in graphics intensive VR gaming applications”. This disclaimer was commented out


The graph that compares the GTX 980 to the 1080 in the ‘performance’ tab doesn’t display exact numbers, meaning most readers are forced to estimate the values. I opened up the inspect element to view the width in pixels of the chart bars, and used simple math to determine the percentage difference. The GTX 1080 performs 70% better in Witcher 3 and 80% better at Rise of the Tomb Raider than the GTX 980, according to the benchmark. These benchmarks were conducted at max settings @ 2560×1600, according to a comment in the source.

Keep in mind that much of the commented out information is subject to change, and it’s possible that some of it could be inaccurate. For example, I found a commented out element that labeled the memory bandwidth as being 336.5 GB/sec, while the actual element displayed on the page states it as 320. This means that earlier on, NVIDIA may have intended higher memory clocks, but reduced them last minute for reasons such as power efficiency, stability, etc.

http://deliddedtech.com/2016/05/07/hidden-information-found-in-source-code-of-nvidias-gtx-1080-page


Rumoured Geforce GTX 1080M Specifications

From Chiphell:
GP104 with 2048 CUDA Cores
1.300 MHz
Performance close to the Geforce GTX 980 Ti

 

[James3shin]

I've been out of the GPU landscape for a while. Is it safe to presume that the 1080 (GP104) will eventually be followed up with a 1080Ti, GP100, later this year?

 

[Qwertilot]

That NV will do a big gaming chip? Utterly safe bet, yes Maybe not a direct P100 derivative this time as P100 is so specalised into compute it doesn't make such a great gaming GPU.

Timing a bit uncertain as it depends on HBM2 in big volume and the process improving. Could slip to say as late as next summer or something, especially if they want to run a titan for a bit first.

AMD's big chips are due on roughly the same sort of time scale, with the same reasons.

 

[Mopetar]

I've been out of the GPU landscape for a while. Is it safe to presume that the 1080 (GP104) will eventually be followed up with a 1080Ti, GP100, later this year?

Probably not until late 2017. Look at how heavily the 1070 is harvested. P100 is already a cut chip and the few they can get go for $10,000 a pop. Until yields improve, any good GP100 dies get sold to the HPC market that doesn't care about the price. As yields improve and HPC demand drops we'll have see a consumer card (probably even more cut) but why would NV sell a $1,000 1080 Ti when they can sell it as a Tesla for ten times as much?

 

[Head1985]

1080 25% faster than 980TI confirmed
https://www.youtube.com/watch?v=3WnmZwChW_s&feature=youtu.be&t=304
ALso 1070 will be slower than 980TI for sure because it is cutdown into oblivion.

 

[b-mac]

1080 25% faster than 980TI confirmed
https://www.youtube.com/watch?v=3WnmZwChW_s&feature=youtu.be&t=304
ALso 1070 will be slower than 980TI for sure because it is cutdown into oblivion.

Was expecting a bit more than a 25% increase, bit sad in my opinion.

 

[frowertr]

1080 25% faster than 980TI confirmed
https://www.youtube.com/watch?v=3WnmZwChW_s&feature=youtu.be&t=304
ALso 1070 will be slower than 980TI for sure because it is cutdown into oblivion.

Aren't the Titan X and 980TI basically the same GPU with the exception of less memory, a few less CCs, and a few less SMs on the 980TI? So if the claim from Nvidia is that the 1070 is faster than a Titan X, wouldn't it stand to also reason that it is also faster than the 980TI?

 

[jpiniero]

Rumoured Geforce GTX 1080M Specifications

From Chiphell:
GP104 with 2048 CUDA Cores
1.300 MHz
Performance close to the Geforce GTX 980 Ti

That's pretty hard to believe that it would be 980 Ti performance if it has 20% less cores and 25% lower base clock.

 

[Head1985]

Aren't the Titan X and 980TI basically the same GPU with the exception of less memory, a few less CCs, and a few less SMs on the 980TI? So if the claim from Nvidia is that the 1070 is faster than a Titan X, wouldn't it stand to also reason that it is also faster than the 980TI?

He talking about VR performance.Same way he says 1080 under VR is 2x faster than TITANX.

 

[kithylin]

Well, I think as long as the silicone will hit 2.1Ghz with a bit of increase in power limits, voltage, and better cooling easily, I think it's pretty good. Imagine 2.3-2.4ghz on water with voltage bumping. That would be pretty sweet!!!

And just wait until us smart people can modify and re-flash bios's, disable boost and run at fixed clock all the time when gaming at 2.1 - 2.2 ghz. That's where the real nvidia performance comes from.

1080 25% faster than 980TI confirmed
https://www.youtube.com/watch?v=3WnmZwChW_s&feature=youtu.be&t=304
ALso 1070 will be slower than 980TI for sure because it is cutdown into oblivion.

Actually the nvidia tweets say the 1070 is faster than the Titan X, and already the titan X is faster than 980 TI... so then 1070 is faster than all 900 series cards. We won't know until actual benchmarks but so far.. nvidia twitter is stating this as fact.

EDIT: Until we get hard benchmark data, no one knows exactly -HOW MUCH FASTER- the 1070 is. We just know, supposedly, "It is faster" period.

 

[PPB]

EDIT: Until we get hard benchmark data, no one knows exactly -HOW MUCH FASTER- the 1070 is. We just know, supposedly, "It is faster" period.

The statement made by the same people that says 1080 has 3x perf/watt than 980ti sure must be taken at face value :sneaky:

JHH brought alfalfa knowing the sheep would love it. Problem is some people are better informed than that. Like the guy saying 1080M will perform like 980ti with those specs. Fail at math lol.

Sent from my XT1040 using Tapatalk





Nice threadcrap.
Read the VC&G announcement, or maybe you have. You must petition the mods in an MD thread for reinstatement after your 30 day has passed.


esquared
Anandtech Forum Director

 

[MagickMan]

The hype seems to be collapsing in on itself, finally. I just want to see the real performance numbers, not the bullsh*t "performance per watt" crap and dodgy marketing spin. :thumbsdown: A 25% increase (for a 1080) over a stock 980TI is great, why not just go with that?? Geez.

I'm not tossing out my 980TIs any time soon, because they OC so well, but if someone has a single TI that's an OCing dog, I can see the 1080 being a solid upgrade (if you count the $450 you can get for the old card on ebay).