Nvidia's Pascal GP104 GPU may opt for GDDR5 over HBM *Updated GDDR5X JDEC Specs

[2is]

As a long time reader of this forum and other tech forum, I know that goal post shifting trend is true.

When the 480 used 2x the power the 5870 did, power usage was not important. When NV's dustbuster was ridiculously noisy, noise wasn't important.

When NV's GPU was slower in single card, 680 vs 7970Ghz, multi-GPU was important because the 680 SLI was smoother. When AMD CF was smoother, smoother was no longer important. When AMD's Fury X CF stomps on 980TI SLI, multi-GPU is no longer important.

The same for OC potential, when the 7950 pulled 50% OC, OC wasn't important. Now Maxwell OC greats, OC is THE metric.

If Pascal has functional Async Compute, then I bet AC will be important... otherwise, "Async Compute? Meh!"..

And if Pascal is a compute focused design that sacrifices perf/w and loses to Polaris in power efficiency, power usage will no longer be important again. We're going full circle.

Whatever happened to simple and logical metrics such as bang for buck? Are you people actually happy to pay ridiculous prices for your GPUs?

As a long time reader of this forum, I know that people like to rewrite history to suite their own agenda, as you just did here. You should be more honest with yourself and the readers here. The dustbsuter (5800) series from NVidia was a disaster even by NVidia fans standards. It was loud, it was hot, it was power hungry and it underperformed. It was not good in any metric.

 

[Ajay]

So it is 4096 cores, clocked at around 950-1000 MHz. With TDP for each GPU at 100W.

Basically GTX980 with 100W of TDP.

So, NV must have something different in store for gamers - otherwise this won't sell well. I'm pretty sure that gamers care about price, performance and efficiency in that order. Probably the main reason they care about efficiency (in the US) is because of all of Intel's marketing on perf/watt.

 

[3DVagabond]

Goal post shifting isn't a NDF only past time. The ADF loves to do it just as much. Computer hardware company fans are all pitiful regardless of the color.

I'll be buying the "Flagship" GPU of whichever company releases first. If the opposing company delivers faster after that then I'll be buying that as a replacement. I need my shiny precious things.

All fans do it. You're right. It's when the Press do it that it gets frustrating. Since the Press is a business it makes you ask why?

 

[jpiniero]

So, NV must have something different in store for gamers - otherwise this won't sell well. I'm pretty sure that gamers care about price, performance and efficiency in that order. Probably the main reason they care about efficiency (in the US) is because of all of Intel's marketing on perf/watt.

This was pretty much inevitable because of the lack of a $/transistor benefit. The lower power draw will really help the mobile models though, so look for nVidia to push that hard.

And if you want better performance, well there's always GP100 and GP102.

 

[Glo.]

16 nm has around 55% density over 28 nm.

So we are looking at a smear over 200mm2 GPUs. GTX1060 or something.

 

[Ajay]

This was pretty much inevitable because of the lack of a $/transistor benefit. The lower power draw will really help the mobile models though, so look for nVidia to push that hard.

And if you want better performance, well there's always GP100 and GP102.

I wonder if that's NV's plan. Selling more (even cut down) compute GPUs helps them with volume to fund high profit Pro cards. Good point on mobiles, so a real good thing for notebook gamers (definitely a big thing with college kids).

 

[Silverforce11]

As a long time reader of this forum, I know that people like to rewrite history to suite their own agenda, as you just did here. You should be more honest with yourself and the readers here. The dustbsuter (5800) series from NVidia was a disaster even by NVidia fans standards. It was loud, it was hot, it was power hungry and it underperformed. It was not good in any metric.

Not even the dust buster, the 480. Hot, loud, very power hungry. Defended here heavily by the same individuals who throw scorn at AMD for those metrics in recent times.

I attacked the 7970 pricing on release, calling it overpriced for a mid-range chip.

I attacked heavily the R290/X reference design, calling it what it is, an awful cooler.

I'm consistent. I value bang for buck, regardless of brand. I value OC potential. I've praised Kepler/Maxwell their ease of OC-ability, especially NV's auto-boost mechanism. I've praised the custom 980Ti for being a great GPU, stomping on Fury X in single configs.

So no, there's no revisionism from me. Everything I stated about goal post shifting is fact.

I guarantee you, if Pascal is worse in efficiency, perf/w will no longer be touted as an important metric, it will shift to something else that Pascal is better at. Just you wait and see. History will repeat itself.

 

[2is]

Google NVidia dustbuster, you'll get hits for the FX 5800, not the 480. At least not on the 1st page.

 

[3DVagabond]

I guarantee you, if Pascal is worse in efficiency, perf/w will no longer be touted as an important metric, it will shift to something else that Pascal is better at. Just you wait and see. History will repeat itself.

Like DP performance making Titan a worthwhile purchase. Even though virtually nobody that bought it did anything that required DP compute.

 

[2is]

Performance per watt will absolutely be talked about. It's all the AMD fans will bring up. Performacne per watt has always been second to performance for dGPU's, you guys are saying this like you're new here. Pray tell when performance per watt was a more important metric then performance for high end dGPU's? I've been around for many years, and in all tha time, the answer has always been, never.

It's brought up now because not only does NVidia have the performance crown, but they also have the performance/watt crown.

It almost seems like you ADF guys are nervous AMD will still be behind in performance so you're bringing up nonsensical scenarios so you can feel a little better and say "I told you so" even though it's a non-factor and never has been.

Performance/watt is merely icing on the cake, a tie-breaker at best.

 

[Silverforce11]

@2is
When the 7970Ghz beat the 680, you frequently hear stuff like how the 680 has better perf/w and that makes it better... see, suddenly power usage was so important.

How about CF R290X vs 780TI SLI, cheaper, faster, smoother... but it uses more power so the 780TI is "better". Didn't work out as well for Kepler..

And ofc power usage is important. But be consistent at least.

This is why I raised this point, if Pascal end up with worse perf/w, you will see the same crowd that once put perf/w on the altar suddenly move to worship a new god.

 

[Arachnotronic]

Whichever company is out first with 14/16nm GPUs aimed at the high end will get my money. Period.

 

[2is]

I'm always consistent. When I bought a 680 it wasn't because it was more power efficient. It was because it was the better performer then the 7970 (ghz wasn't out and drivers with performance improvements did not exist yet) Not only was it the faster card, but it was also cheaper.

When we are embraced with Polaris vs Pascal discussion, performance will be the main focus there too. Of course, there will be the fans that will point of the strengths of their "favorite" brand, but that applies to both sides, though you seem to only be calling out one of them.

 

[Goatsecks]

Not even the dust buster, the 480. Hot, loud, very power hungry. Defended here heavily by the same individuals who throw scorn at AMD for those metrics in recent times...

I guarantee you, if Pascal is worse in efficiency, perf/w will no longer be touted as an important metric, it will shift to something else that Pascal is better at. Just you wait and see. History will repeat itself.

Ah, the hypocrisies of the schizophrenic hive mind. It is a confusing creature. :awe:

 

[Goatsecks]

...When we are embraced with Polaris vs Pascal discussion, performance will be the main focus there too. Of course, there will be the fans that will point of the strengths of their "favorite" brand, but that applies to both sides...

Well said.

Additionally people seem reluctant to accept a reality of economics, the taboo of taboos: brand preference. A lot of arguments on this forum could be saved if people did not feel the need to defend their, or attack others, brand preference but, instead, accept it.

 

[3DVagabond]

Well said.

Additionally people seem reluctant to accept a reality of economics, the taboo of taboos: brand preference. A lot of arguments on this forum could be saved if people did not feel the need to defend their, or attack others, brand preference but, instead, accept it.

Well, when you have people who recommend not to buy AMD GPU's and use Bulldozer as the reason, or Crossfire because of the issues when the 7970 was first released, it gets hard to simply allow that advice to go out unchallenged.

 

[2is]

Well, when you have people who recommend not to buy AMD GPU's and use Bulldozer as the reason, or Crossfire because of the issues when the 7970 was first released, it gets hard to simply allow that advice to go out unchallenged.

Thought you were complaining about performance/watt just a few posts ago, or agreeing with someone who was anyway.

You know what I see here that's quite funny? The Polaris vs Pascal debate hasn't even started yet, but the ADF is already arguing about it and shifting goal posts as you just did here. At least pretend to have some confidence in your brand of choice and wait until the products are released before you start complaining about what people will say or do.

 

[Ajay]

Hey, I have and idea here, let's pretend this is a thread about Nvidia - kind of like the title says it is :awe:

Maybe a kind mod will move it to Nvidia...

 

[ShintaiDK]

Samsung says in their press release that HBM2 is for HPC to begin with.
https://benchlife.info/samsung-electronics-mass-production-hbm2-with-20nm-process-01192016/

Translated it means GP200 and equal. So another pointer that GP104 is GDDR5(X).

 

[Ajay]

Samsung says in their press release that HBM2 is for HPC to begin with.
https://benchlife.info/samsung-electronics-mass-production-hbm2-with-20nm-process-01192016/

Translated it means GP200 and equal. So another pointer that GP104 is GDDR5(X).

Thanks. That makes allot of sense, IMHO.

 

[cm123]

GDDR5X is your standard GDDR5 memory however, opposed to delivering 32 byte/access to the memory cells, this is doubled up towards 64 byte/access. And that in theory could double up graphics card memory bandwith. Early indications according to the presentation show numbers with the memory capable of doing up-to 10 to 12 Gbps, and in the future 16 Gbps. So your high-end graphics cardsthese days hover at say 400 GB/s. With GDDR5X that could increase to 800~1000 GB/sec and thus these are very significant improvements, actually they are competitive enough with HBM.

Jedec:

Derived from the widely adopted GDDR5 SGRAM JEDEC standard, GDDR5X specifies key elements related to the design and operability of memory chips for applications requiring very high memory bandwidth. With the intent to address the needs of high-performance applications demanding ever higher data rates, GDDR5X is targeting data rates of 10 to 14 Gb/s, a 2X increase over GDDR5. In order to allow a smooth transition from GDDR5, GDDR5X utilizes the same, proven pseudo open drain (POD) signaling as GDDR5.

“GDDR5X represents a significant leap forward for high end GPU design,” said Mian Quddus, JEDEC Board of Directors Chairman. “Its performance improvements over the prior standard will help enable the next generation of graphics and other high-performance applications.”



http://www.guru3d.com/news-story/jedec-announces-publication-of-gddr5x-graphics-memory-standard.html

 

[CakeMonster]

So HBM still has the area and power usage advantage?

 

[2is]

HBM's power advantage is what I'm most interested in. With modern cards having a TDP wall, the less power the memory needs equals that much more power available to the GPU.

 

[Golgatha]

Tesla/Quadro/Firepro needs up to 32. They already got 24/32 today.

Its more like 4/8/16 for gaming GPUs.

This would be my guess as well. Probably not 16 for even the high end though unless it's a Titan or dual GPU card.

 

[tviceman]

So HBM still has the area and power usage advantage?

HBM's power advantage is what I'm most interested in. With modern cards having a TDP wall, the less power the memory needs equals that much more power available to the GPU.

HBM's power usage advantage is what matters most. The area usage isn't important on smaller dies, especially when taking into account the associated costs with going the HBM route.



On a side note, a 256-bit bus equipped GPU with 14ghz GDDR5X has 448 gb/s bandwidth - twice that of GTX980, and 33% more than GTX Titan X / 980 TI. A 384-bit bus equipped GPU would have 672 gb/s, a 31% increase over Fury X. Next gen cards should have no noticeable bandwidth bottlenecks with respects to graphics performance.