NVIDIA Pascal Thread

[Vesku]

If GP104 starts out on 256 bit GDDR5 (non-X) I think quite a few people in this thread are overestimating its performance. 1080 vs 980Ti might look a bit like 285 vs 7970GHz.

 

[Silverforce11]

If GP104 starts out on 256 bit GDDR5 (non-X) I think quite a few people in this thread are overestimating its performance. 1080 vs 980Ti might look a bit like 285 vs 7970GHz.

Yup, which is why I don't think it will happen. NV knows 256 GDDR5 is going to limit it's performance.

Better to wait a few months for GDDR5X.

 

[RussianSensation]

Not that I totally disagree, but to pedant - 980's are already only 165w, so would be far from obselete crap vs a 120W / identical performance die shrunk competitor. They'd have to price cut a lot of course but fine for stalling for a few months.

No, they are not. NV TDP = more often than not made up marketing. After 970/980 being out for 1.5 years, you should know better since this topic was already covered in Q4 2014

If Polaris 10's actual power usage is 120W, that's significantly below the real world power usage of a 980. That's not the point though since Polaris 10 has to compete against Pascal, not what the 970/980. Remember, NV aims to bring 2x perf/watt to the entire Pascal stack. That means potentially 75-100% more performance in a 175-200W 980 successor (let's call it 1080).

This seems like the fastest way for AMD to take its market share at single digits, if true. If Polaris 10 is as slow as you expect it to be, AMD will be offering no new SKU that would be an actual upgrade for hawaii/GM204 users and upwards.Imagine loyal AMD customers on hawaii since late 2013 waiting for something better than Fury X - this sends them straight to nvidia. AMD's only chance to sell anything midrange or higher would be to attract a small portion of the crowd that did not find Fiji worthwhile the first time round and budget buyers that still use pitcairns etc and will be looking for an upgrade in the 150-200 range. AMD will have to price their Fiji respins really low as well, since nvidia will be ruling the roost at 450+ with much higher performance. AMD getting pigeonholed in cheap brackets while competing with their 28nm SKUs smells disaster.

Supposedly 232mm2 with 2304-2560 shaders. How in the world do you expect such a chip to beat Fury X? As far as loyal Hawaii customers, most of those upgraded to 980Ti or are waiting for Vega. Think about it, if 290/290X users didn't upgrade to Fury/Fury X, you think they waited an entire year to just get Fury/Fury X performance at lower power usage (that's even if Polaris manages to match Fiji -- a tall order in itself). I would bet if someone has a 290/290X and hasn't yet upgraded to a 980Ti, they are going to want a card that's 15-20% faster than the Fury X/980Ti which is most likely going to be GP104, not Polaris 10.

The question is whether Polaris 10 will really have 2560 shading units. If yes then it may beat Fury easily with all announced improvements.

Seriously?

1300mhz @ 2560 shaders with 30% IPC (huge increase per shader)
Fury X 1050mhz @ 4096 shaders
~ theoretical tie

That doesn't even address how you'll come up with 512GB/sec ("387GB/sec real world") memory bandwidth.

Again, it makes NO sense why AMD would paper launch Radeon Pro Duo if a 2560 Polaris 10 is faster than the Fury X or even as fast. I am not sure people read my previous post as I already covered these points.

As it stands, NV's GP104 will have the market to itself until Vega and that's my view unless we get new data on Polaris. I am supposed to believe that a 120-150W 2560 Polaris chip with nearly 1/3rd the die size of Fiji, with only 256-bit bus is going to beat a 280W Fury X HBM in games and cost $299 or less (i.e., target VR spec)? Sounds like wishful thinking/hype. Notice that AMD's own roadmap has Polaris with 2.5X perf/watt of 2014 GCN? They didn't use Fiji as a point of reference for a 2.5X increase in perf/watt.

If GP104 starts out on 256 bit GDDR5 (non-X) I think quite a few people in this thread are overestimating its performance. 1080 vs 980Ti might look a bit like 285 vs 7970GHz.

1. 285 is a 384-bit bus chip, not a 256-bit. It's also overloaded with 8 ACEs for no reason given how few shaders it has. It also has TrueAudio that wastes graphics transistor space.

2. NV's next gen mid-range chip has almost always matched or outperformed the last gen's flagship card. You are saying it'll be the first time in 20 years when this isn't going to happen?

3. You can get 320GB/sec bandwidth and 384GB/sec with 256-bit and GDDR5X @ 10Gbps and 12Gbps, respectively. 980Ti has 336GB/sec. Micron already has both 10 and 12Gbps GDDR5X.
http://www.anandtech.com/show/10193/micron-begins-to-sample-gddr5x-memory

 

[Kris194]

Yup, which is why I don't think it will happen. NV knows 256 GDDR5 is going to limit it's performance.

Better to wait a few months for GDDR5X.

This, whatever Nvidia will do, they won't have a competition for few months (I believe that GP104 will be released at some point in August or even September unless it's GDDR5, not X. They have no reason to hurry.

 

[mohit9206]

I feel the upcoming Pascal graphics cards from Nvidia will only be 15-25% faster than the cards they will be replacing and also cost 15-25% higher than the cards they will be replacing.
This is because Nvidia is the undisputed king in the discrete market and they can price their products however they want.
So my prediction 1060 being 20% faster than 960 and MSRP $249, 1070 15% faster than 970 and MSRP $349, 1080 15% faster than 980 and MSRP $499 and 980ti replacement being 15% faster and MSRP $699.
These prices are the suggested MSRP from Nvidia. Obviously partner cards with custom coolers and factory overclocked models will be even higher priced.

 

[Vesku]

1. 285 is a 384-bit bus chip, not a 256-bit. It's also overloaded with 8 ACEs for no reason given how few shaders it has. It also has TrueAudio that wastes graphics transistor space.

2. NV's next gen mid-range chip has almost always matched or outperformed the last gen's flagship card. You are saying it'll be the first time in 20 years when this isn't going to happen?

3. You can get 320GB/sec bandwidth and 384GB/sec with 256-bit and GDDR5X @ 10Gbps and 12Gbps, respectively. 980Ti has 336GB/sec. Micron already has both 10 and 12Gbps GDDR5X.
http://www.anandtech.com/show/10193/micron-begins-to-sample-gddr5x-memory

1. It only implements 256-bit and the limitations of that can be seen under certain conditions.

2. Same thing could be said before Nvidia created an additional price tier with their Titan line, i.e. "they've never done that before so they won't do it ever."

3. From the article you link: "Micron claims that it is set to start mass production of the new memory this summer, which hopefully means we're going to be seeing graphics cards featuring GDDR5X before the end of the year." For a mass market position such as a traditional $400-550 *70/*80 it will be surprising to see volume retail availability before late August if they are using GDDR5X.

Note I specifically said if GP104 is GDDR5 and not GDDR5X.

If this rumored May launch is accurate than either tight supply with GDDR5X, which will burn brand goodwill with potential customers, or decent availability with GDDR5.

 

[Silverforce11]

If this rumored May launch is accurate than either tight supply with GDDR5X, which will burn brand goodwill with potential customers, or decent availability with GDDR5.

Very unlike to be May, given we haven't got any concrete leaks.

Still a few months out.

 

[RussianSensation]

1. It only implements 256-bit and the limitations of that can be seen under certain conditions.

My point was using 7970Ghz vs. R9 285/380X as a comparison to imply that 1070/1080 will flop the same way makes no sense since 285/380X have 'redundant' and 'useless' assets such as unused 384-bit memory bus and TrueAudio. 1070/1080 should be optimally designed so there is no reason at all to expect them to not beat 970/980 by at least 50% if NV is touting 100% increase in perf/watt.

2. Same thing could be said before Nvidia created an additional price tier with their Titan line, i.e. "they've never done that before so they won't do it ever."

The Titan line was an increase in price. You are implying something entirely different -- deliver barely any more performance with the main selling feature being lower power usage? NV cannot risk that since once Vega shows up, they'll need new mid-range cards again. Think about when NV was designing Pascal 3-4 years ago. They probably didn't foresee AMD bombing this hard with Fury/Fury X but the Pascal design was well underway.

If you are NV and you managed to raise mid-range price from $249 to $499 with 680 and then yet again to $550 with the 980, wouldn't you want to repeat that again? Now compare the performance increase in the same lineage:

560Ti -> 680
680 -> 980

HUGE.

Apply the same to the 980 and it should be clear unless NV failed miserably, this card should beat 980Ti by 25-40%.

3. From the article you link: "Micron claims that it is set to start mass production of the new memory this summer, which hopefully means we're going to be seeing graphics cards featuring GDDR5X before the end of the year." For a mass market position such as a traditional $400-550 *70/*80 it will be surprising to see volume retail availability before late August if they are using GDDR5X.

Note I specifically said if GP104 is GDDR5 and not GDDR5X.

If this rumored May launch is accurate than either tight supply with GDDR5X, which will burn brand goodwill with potential customers, or decent availability with GDDR5.

There are other possibilities I suppose:

1) 384-bit bus GP104 with GDDR5
2) 256-bit bus with incredible 4th gen memory compression
3) launch much later than May / June 2016 -> August/Sept 2016. Meaning NV could unveil Pascal architecture and paper launch in Q2 2016 with retail availability in Q3 2016.
4) it's possible the rumors are wrong and we'll see 950/960 replacements first. It's not the first time launch time-frame rumours were not aligned with reality.

Thing is though, when several different sources start linking similar information of 1070/1080 GDDR5X unveiling is really close, I am more likely to believe it.

Very unlike to be May, given we haven't got any concrete leaks.

Still a few months out.

I think I read retail launch around May 27-29th. That means possibly just 2 months left before GP104 reference cards are in retail on Newegg/Amazon. Would be crazy if true since that's more than 6 months ahead of Vega.

 

[AtenRa]

1070/1080 should be optimally designed so there is no reason at all to expect them to not beat 970/980 by at least 50% if NV is touting 100% increase in perf/watt.

For the GTX 1080 to be 50% faster than the GTX980, it should be close to 300mm2 and 9B transistors. That will put it 10-20% faster than GM200 and GTX 980Ti.

If they will use GDDR-5, they will need 384-bit memory. That GPU could be released in early Summer.

If they will use 256-bit memory they will need GDDR-5X. This one could only be released after Summer of 2016.

Also, 2x the perf/watt doesnt necessarily mean at the same TDP as the GTX 980.

 

[Timmah!]

For the GTX 1080 to be 50% faster than the GTX980, it should be close to 300mm2 and 9B transistors. That will put it 10-20% faster than GM200 and GTX 980Ti.

If they will use GDDR-5, they will need 384-bit memory. That GPU could be released in early Summer.

If they will use 256-bit memory they will need GDDR-5X. This one could only be released after Summer of 2016.

Also, 2x the perf/watt doesnt necessarily mean at the same TDP as the GTX 980.

With big Pascal rumored to have 17B transistors, that would be pretty underwhelming 1080. Pretty much 1/2...when 980 was 2/3 of Titan X, so was old 460/560 compared to 580... 780 was even 80 percent of Titan Black/780Ti, granted that was cut down 100/110, not a 104 class of chip...

 

[Silverforce11]

With big Pascal rumored to have 17B transistors, that would be pretty underwhelming 1080. Pretty much 1/2...when 980 was 2/3 of Titan X, so was old 460/560 compared to 580... 780 was even 80 percent of Titan Black/780Ti, granted that was cut down 100/110, not a 104 class of chip...

That's a pretty big leap, 17B transistors on TSMC 16FF, that they have said in marketing (think best case scenario) achieves 2x the density compared to 28nm.

GM200 is 8B transistors, at 600mm2. A perfect 2x is 16B. But we know it's rarely the case, not everything scales well. A more realistic target is around 14B transistors at ~550mm2 on 16nm FF.

IIRC, the FP64 target for Intel's KL is above 3TFlops, right? Say 3.5Tflops DP/FP64 is a target to reach for big Pascal. So yeah, it needs to be a huge chip.

 

[LTC8K6]

Didn't Samsung make new 8gb GDDR5 chips as well? Perhaps these will factor in somewhere? Both Micron and Samsung announced the new higher density and faster GDDR5 chips last year.

http://techreport.com/news/27676/samsung-starts-making-8gb-gddr5-memory-chips

Wiki:

As of January 15, 2015, Samsung Electronics announced in a press release that it had begun mass production of "8 Gb" (8 × 1024 bits) GDDR5 memory chips based on a 20 nm fabrication process. To meet the demand of higher resolution displays (such as 4K) becoming more mainstream, higher density chips are required in order to facilitate larger frame buffers for graphically intensive computation, namely PC gaming and other 3D rendering. Increased bandwidth of the new high-density modules equates to 8 Gbit/s per pin × 170 pins on the BGA package x 32-bits per I/O cycle, or 256 Gbit/s effective bandwidth per chip.

On January 6, 2015, Micron Technology President Mark Adams announced the successful sampling of 8 Gb GDDR5 on the company's fiscal Q1-2015 earnings call.[21][22] The company then announced, on January 25th, 2015, that it had begun commercial shipments of GDDR5 using a 20 nm process technology.[23][24][25] The formal announcement of Micron's 8 Gb GDDR5 appeared in the form of a blog post by Kristopher Kido on the company's website September 1, 2015.

 

[Creig]

I think I read retail launch around May 27-29th. That means possibly just 2 months left before GP104 reference cards are in retail on Newegg/Amazon. Would be crazy if true since that's more than 6 months ahead of Vega.

Doesn't Jen-Hsun usually parade a board (working or otherwise) around on stage prior to launch? So far we haven't even seen a Pascal chip much less a complete card. The Drive PX2 which he said had Pascal chips on it were proven to be just Maxwell MXM modules.

It's possible that Pascal will launch the end of next month. But the closer we get to the end of May without having seen Jen-Hsun up on stage with one, the less likely it is that we'll actually see retail cards by then.

 

[jpiniero]

Doesn't Jen-Hsun usually parade a board (working or otherwise) around on stage prior to launch? So far we haven't even seen a Pascal chip much less a complete card. The Drive PX2 which he said had Pascal chips on it were proven to be just Maxwell MXM modules.

I would be really surprised if we don't see Pascal running tomorrow. It may not be in games though.

Given all the Tesla hype I imagine we will see the real Drive PX2.

 

[blake0812]

I literally ordered another gtx 970 yesterday. Just my luck, huh?

 

[steve wilson]

I literally ordered another gtx 970 yesterday. Just my luck, huh?

If I was you I would cancel or return that order. Even if they only give you store credit, so you can get a 1070 or sell your current 970 and stretch to a 1080.

 

[blake0812]

If I was you I would cancel or return that order. Even if they only give you store credit, so you can get a 1070 or sell your current 970 and stretch to a 1080.

I'm planning on doing that.

 

[96Firebird]

If you happened to buy an EVGA 970, you can look into the step-up program...

 

[Glo.]

Nvidia just released mobile Quadro based on GM204 with 1140 MHz core clock, 150W and 4.7 TFLOPs of compute power.

OEMs asked them for this type of GPUs otherwise I do not see any other reason to produce them. Unless - something with Pascal or technology behind it is just plainly wrong. Lets hope that the last thing is not the case.

P.S. Fiji from S9300x2 has the same TDP and has 6.9 TFLOPs of compute power.

 

[Slaughterem]

P.S. Fiji from S9300x2 has the same TDP and has 6.9 TFLOPs of compute power.

More about this card and its usage in Science http://www.theregister.co.uk/2016/04/03/canadas_chime_telescope_taps_amd_for_gpubased_super/

While AMD didn't announce how many of its data-centre grade GPUs CHIME has ordered, this 2015 Nature article said the telescope will need 1,000 units.

Card sells for $5,999

 

[jpiniero]

OEMs asked them for this type of GPUs otherwise I do not see any other reason to produce them. Unless - something with Pascal or technology behind it is just plainly wrong. Lets hope that the last thing is not the case.

GP104 isn't going to be available at least for a few more months. As the article said it's basically a 980 for notebooks chip, so the hardware work has already been done with it.

 

[RussianSensation]

Nvidia just released mobile Quadro based on GM204 with 1140 MHz core clock, 150W and 4.7 TFLOPs of compute power.

OEMs asked them for this type of GPUs otherwise I do not see any other reason to produce them. Unless - something with Pascal or technology behind it is just plainly wrong. Lets hope that the last thing is not the case.

P.S. Fiji from S9300x2 has the same TDP and has 6.9 TFLOPs of compute power.

Another explanation is that NV has 0 competition in that space so releasing a 1.5 year old 980 at high prices maximizes profits while still delivering top of the line performance for that market segment. Why release a card 2x faster when you can milk old tech at maximum profit margin? The desktop market is different. A lot of gamers upgrade in Q3/4 as many high profile AAA games release around that time. If NV has GP104 out in volumes by Q3, they will flood the market in Q4, while Vega will be a spec on the map. This would repeat the 970/980 strategy, but with even less competition. Fiji cannot overclock well. If 1070/1080 have 20-25% OCing headroom and 1080 is 20-25% faster than 980Ti, it's going to be hard to imagine waiting 6-8 months for Vega. I keep coming back to this point: the fastest way to lose a generation is to not show up on time (for AMD specifically). Unlike many NV customers, AMD buyers/objective gamers will not wait 6-8 months for AMD to show up. If GP104 launches in June-July and AMD has weak Polaris 10, that's like 90%+ of $350-550 market share sales going to NV.

Either GP104 launching soon is not reality or AMD's Polaris is much faster or AMD is about to get hammered.

Let's look at it another way. Even if 1070 8GB just matches 980Ti, has only 15% overclocking headroom and costs $399, I personally would not buy any Fiji card for more than $299. Fury X uses 280W of power and has 4GB VRAM. I'd easily pay $100 premium for 2X the VRAM, 20%+ more OCed performance over Fury X OC, with 100-120W less power, DP1.3/HDMI 2.0a/HDR support, latest video codec, etc. I am actually scared for the GPU market since while some 7950/7970/280/290 owners skipped this gen, it's going to much harder to not upgrade to 1070/1080.

Also, historically speaking, NV cards tend to overclock better too. That's why even if AMD lowers Fury X to $349, $399 1070 ~ 980Ti would still be better. That's why I am saying Fury X @ $299 if 1070 is at least as fast as a 980Ti with 15% OC room on top. If NV prices 1070 @ $329 like 970, those things will fly off the shelves faster than Skylake K during the first 6 months. Scary times for AMD.

 

[Glo.]

If you mean that Nvidia has no competition in CUDA world - you are right, but that is changing. If you mean compute world as just like that - they have huge competition.

The absurdity of Nvidia mindshare and CUDA is that people believe that 4 TFLOPs GPU from Nvidia is magically more powerful than 4 TFLOPs GPU from any other vendor. It is the software that makes difference here. Only that. So far, the only one application that has proper implementation of OpenCL is... Final Cut Pro X. And nothing else.

 

[Head1985]

Another explanation is that NV has 0 competition in that space so releasing a 1.5 year old 980 at high prices maximizes profits while still delivering top of the line performance for that market segment. Why release a card 2x faster when you can milk old tech at maximum profit margin? The desktop market is different. A lot of gamers upgrade in Q3/4 as many high profile AAA games release around that time. If NV has GP104 out in volumes by Q3, they will flood the market in Q4, while Vega will be a spec on the map. This would repeat the 970/980 strategy, but with even less competition. Fiji cannot overclock well. If 1070/1080 have 20-25% OCing headroom and 1080 is 20-25% faster than 980Ti, it's going to be hard to imagine waiting 6-8 months for Vega. I keep coming back to this point: the fastest way to lose a generation is to not show up on time (for AMD specifically). Unlike many NV customers, AMD buyers/objective gamers will not wait 6-8 months for AMD to show up. If GP104 launches in June-July and AMD has weak Polaris 10, that's like 90%+ of $350-550 market share sales going to NV.

Either GP104 launching soon is not reality or AMD's Polaris is much faster or AMD is about to get hammered.

Let's look at it another way. Even if 1070 8GB just matches 980Ti, has only 15% overclocking headroom and costs $399, I personally would not buy any Fiji card for more than $299. Fury X uses 280W of power and has 4GB VRAM. I'd easily pay $100 premium for 2X the VRAM, 20%+ more OCed performance over Fury X OC, with 100-120W less power, DP1.3/HDMI 2.0a/HDR support, latest video codec, etc. I am actually scared for the GPU market since while some 7950/7970/280/290 owners skipped this gen, it's going to much harder to not upgrade to 1070/1080.

Also, historically speaking, NV cards tend to overclock better too. That's why even if AMD lowers Fury X to $349, $399 1070 ~ 980Ti would still be better. That's why I am saying Fury X @ $299 if 1070 is at least as fast as a 980Ti with 15% OC room on top. If NV prices 1070 @ $329 like 970, those things will fly off the shelves faster than Skylake K during the first 6 months. Scary times for AMD.

The power of GTX970 was price.GTX980 was crap from start to today.
I dont see problem if AMD will have whole market with polaris 11 and 10 from entry level to GTX980 level of performance.They will gain a ton of market share.All they need is price it well(Max 300USD for GTX980 performance with polaris 10)
NV can have 1070/1080 at 400-550USD market.But what is better?Have whole market from entry to GTX980 performance on new 14/16nm node or have 2 cards at 400-550USD market on new 16nm node and rest old obsolete crap on 28nm?

 

[poofyhairguy]

But what is better?

The $300+ market is better. Easily. Why?

1. Most of the margin/profit is made on these cards.

2. Having the fastest overall card on the market creates a halo effect for lessor cards- basically uninformed customers that buy something like a 750 Ti because they hear "nvidia is the best (as based on the 980 ti)."