Pascal consumer = Maxwell on speed confirmed

[Kobz360]

3 minutes ago, GoodBytes said:

Told you guys that Pascal was Maxwell 2.0. It is not a new architecture.

As predicted, and as we see, Pascal is an improved Maxwell, with added new features and for the 1080 GDDR5X, all helping achieving greater performance.

Volta is the new architecture.

 

While not very consistent, Nvidia does the tick-tock thing. Where they release a new architecture, then releases a "2.0" version of it, then releases a new architecture, then it's 2.0, and so on and so forth. Sometimes they release a 3.0 version, if the new architecture is delayed.

 

Well in 2013 (@GTC), Pascal wasn't even on the roadmap so it's not surprising it's not a substantial architectural improvement vs. Maxwell. It wouldn't surprise me at all if it was mainly just a Maxwell shrink. That of course, doesn't change the fact that the 1080 and 1070 are still beasts of cards but yeah. 

 

 

http://www.anandtech.com/show/7900/nvidia-updates-gpu-roadmap-unveils-pascal-architecture-for-2016

 

[DocSwag]

8 minutes ago, Agost said:

 Yes, but the Nano was on Fury performance levels while pulling ~50W less and was mainly thermal throttled. Looks like GCN has an inherent clockspeed limitation

I'm not sure I see your point here.... When compared to the fury X, the nano is the fury X with slower clocks and the fury is the fury X with less shaders, which again proves my point that reducing clocks dramatically increases efficiency. Plus the fury X was slightly faster anyways.

 

Also, the nano wasn't mainly thermal throttling. It was also power throttling a lot.

[Agosto]

Just now, DocSwag said:

I'm not sure I see your point here.... When compared to the fury X, the nano is the fury X with slower clocks and the fury is the fury X with less shaders, which again proves my point that reducing clocks dramatically increases efficiency. Plus the fury X was slightly faster anyways.

 

Also, the nano wasn't mainly thermal throttling. It was also power throttling a lot.

That's obvious that reducing clocks helps with power consumption, I have nothing to say about that

The issue is that it looks like GCN has some issues with clockspeeds which were not accounted very much for Polaris. Anche the Nano was thermal throttling a lot.

[DocSwag]

12 minutes ago, Agost said:

That's obvious that reducing clocks helps with power consumption, I have nothing to say about that

The issue is that it looks like GCN has some issues with clockspeeds which were not accounted very much for Polaris. Anche the Nano was thermal throttling a lot.

That, I 100% agree with. GCN is a horrible overclocker, and so a great underclocker. Nvidias recent architectures, on the other hand, are great overclockers. Imo, if AMD wants to stay in the game, they will have to make either architectures that beat nvidia at its own game (efficiency), which is gonna be pretty hard considering AMD has been losing ever since gcn came out, or design a completely new architecture from scratch to replace gcn, which would take many, many years. Considering raja Koduri said Polaris took 3 years to design, that would take something like 4-5 years. I hope they've already started working on a gcn replacement, because if they haven't... AMD doesn't seem like they'll be doing too well until 2020 or 2021.

[BuckGup]

56 minutes ago, HKZeroFive said:

The Titan XP is not a full GP102. The full chip should have 3840 cores (evident that the Quadro P6000 has 3840 cores and is based on GP102) yet the XP only has 3584. I've been saying for days that this seems more reminiscent of Kepler rather than Maxwell.

So they are going to release something with 300 more cores for $1800? Yeah no

[patrickjp93]

36 minutes ago, Agost said:

Well, considering that TSMC has always made high power processes, while GloFo's 14nm is just a copy of Samsung's 14nm LPP lithography and they are both actually 20nm BEOL, TSMC could have better transistors for high power GPUs (look at 10x0 stock voltages, they are about 0.1V lower than that stock RX 480).

*thoughts freely derived from some @patrickjp93 words about that in some topic*

That's it in a nutshell.

 

36 minutes ago, GoodBytes said:

Told you guys that Pascal was Maxwell 2.0. It is not a new architecture.

As predicted, and as we see, Pascal is an improved Maxwell, with added new features and for the 1080 GDDR5X, all helping achieving greater performance.

Volta is the new architecture.

 

While not very consistent, Nvidia does the tick-tock thing. Where they release a new architecture, then releases a "2.0" version of it, then releases a new architecture, then it's 2.0, and so on and so forth. Sometimes they release a 3.0 version, if the new architecture is delayed.

Yeah, but this is the first time that Nvidia has done that moving across nodes. And it's more accurate to say Maxwell was Pascal 0.9/Beta. Maxwell was just stripped down to fit on 28nm when TSMC announced 20nm HP wasn't coming.

[patrickjp93]

3 minutes ago, SamStrecker said:

So they are going to release something with 300 more cores for $1800? Yeah no

Titan X Black Edition: $1500

[Valentyn]

Just now, patrickjp93 said:

Titan X Black Edition: $1500

I can really see that happening. This entire launch reminds me of Kepler, and the original Titan.

 

Will be interesting to see how it plays out. I just really hope Vega isn't too far off and those two chips bring at least comparable performance.

[GoodBytes]

Just now, patrickjp93 said:

Yeah, but this is the first time that Nvidia has done that moving across nodes. And it's more accurate to say Maxwell was Pascal 0.9/Beta. Maxwell was just stripped down to fit on 28nm when TSMC announced 20nm HP wasn't coming.

Nha. When a company like Nvidia looks into die shrink, they do production test, and see what went wrong and work with the foundry to see on ways to fix them, and it is a role that both side work on heavily (chip design modification & production process changes/improvement). Lots of R&D on both sides is put. TSMC is a foundry, so when they say "we can do 10nm", it is really "We are not a processor manufacture, but on our test sample chips that we design, we were able to shrink it down to 10nm... we have never done complex processors. Work with us to make it happen." It actually sounds like a poor deal for chip makers, but sadly that is how it works.

 

Ultimately, a decision needs to be taking. You can spend more millions with a 1-2 years delay your product release to release what you want, or switch to larger die size, let the chip foundry work out the production issues, which you know will take many months as they need to potentially design or modify new equipment to make your product originally planed happen. At the same time you dedicate a small team working to learn what needs to be modified on the design wise of the chip, so that when things are ready, then can incorporate everything they learn in the updated chip design, and everyone crosses their fingers that it works, and not that they are other things missed.

 

It is a very time consuming, and costly process, such as R&D in general.

 

 

[AlTech]

14 minutes ago, patrickjp93 said:

Titan X Black Edition: $1500

Yeah which will be full GP102 GP100 with 3840 cuda cores and HBM2 VRAM.

[AlTech]

1 hour ago, Dabombinable said:

There were no architectural changes. At all!!!!!

Anandtech said otherwise but you know you sound a lot more credible than a journalist. Not.

[AlTech]

51 minutes ago, Agost said:

 Yes, but the Nano was on Fury performance levels while pulling ~50W less and was mainly thermal throttled. Looks like GCN has an inherent clockspeed limitation

GCN doesn't do well when highly clocked.

 

Which is why it needs as many ALUs as possible.

[PlayStation 2]

3 hours ago, Dresta said:

Ngreedia

 

The was customers are played

huh

[HKZeroFive]

4 minutes ago, AluminiumTech said:

Yeah which will be full GP102 GP100 with 3840 cuda cores and HBM2 VRAM.

GP100 is basically GP102 with double precision. Great for computing, but for gaming? Forget about it. Plus, NVIDIA is prioritising HBM2 for their Teslas so I doubt they're going to be releasing an enthusiast card with HBM2 for Pascal. AMD on the other hand, might.

[xAcid9]

2 hours ago, AluminiumTech said:

nobody claimed 85w anything. I think your mistaking it for the GTX 950 vs RX 460 BF4 comparison on 1080p.

 

Polaris supports DX 12 FL12_0 whereas pascal support DX12 FL12_1.

 

But that's ok since Maxwell and Maxwell 2 Support DX12 FL11_1.

 

And why does 12_1 matter more than 12_0? DX12 games can adapt to support whatever feature level the GPU support.

Maxwell v2 support FL 12_1 as well btw. (950/960/970/980/980Ti/TX)

[Curufinwe_wins]

6 minutes ago, GoodBytes said:

snip

I think the way I (and similarly Anandtech) described Pascal as Maxwell w/ a large number of one-off improvements (dynamic scheduling, gddr5x, task/pixel preemption, multi-projection, node shrink). It makes what comes next for Nvidia quite interesting to me seeing how radical a departure Maxwell was.

 

 

 

Quote

In a by-the-numbers comparison then, Pascal does not bring any notable changes in throughput relative to Maxwell. CUDA cores, texture units, PolyMorph Engines, Raster Engines, and ROPs all have identical theoretical throughput-per-clock as compared to Maxwell. So on a clock-for-clock, unit-for-unit basis, Pascal is not any faster on paper. And while NVIDIA does not disclose the size/speed of most of their internal datapaths, so far I haven’t seen anything to suggest that these have radically changed. This continuity means that outside of its new features, GP104 behaves a lot like GM204. Though it should be noted that real world efficiency isn’t quite as cut and dry, as various factors such as the increased SM count and changes in memory technology can greatly influence this.

 

[Curufinwe_wins]

 

7 minutes ago, AluminiumTech said:

Anandtech said otherwise but you know you sound a lot more credible than a journalist. Not.

Yea you are just blatantly wrong here. There were 0 hardware path changes to the die or anything else (other than the use of newer higher quality components).

 

There were software optimizations (quite similar to the 480x patch actually) that along with the extremely mature process meant card by card performance was MUCH more uniform but the best 390(x) cards paled in comparison to the best 290x 8GB models (consistently reaching 1.15 was fairly easy, but only MSI and XFX models could even more than occasionally go above that.)

 

The biggest single improvement was the tessellation and color compression formulas that were brought over from Tonga. These improvements did make their way to the 290(x) variants a few weeks later (mid august 2015).

 

[grayperview]

1 hour ago, Agost said:

 Yes, but the Nano was on Fury performance levels while pulling ~50W less and was mainly thermal throttled. Looks like GCN has an inherent clockspeed limitation

Seems that way, it would make sense why AMD are focusing on IPC for now, while they figure out how to get the higher clock speeds..If they get the high clocks they should given their IPC advantage, overtake Nvidia.

 

 

[laminutederire]

4 hours ago, Dresta said:

 

In a couple years when games adopt more VRAM and Vulkan,the RX 480 will crush the 970 to bits,but that does not matter since Nvidia will drop Maxwell just like they did Kepler since

 

Ngreedia

 

The was customers are played

Dude, the rx 480 already owns the 970, do your research before trying to start a flame war.

[laminutederire]

4 hours ago, PerfectTemplar said:

Evidently, it wasn't meant to compete with the 1060 either

True, the 1060 is meant to compete with it

[goodtofufriday]

4 hours ago, Agost said:

What? I didn't mean failed on the economic side... it failed to deliver what I (and many others) expected from Polaris. I don't want AMD to fail, I really hoped for a good 480 in order to switch to AMD.

 

DX12 and Vulkan are still in their early stages, and the gains on these APIs on AMD cards are large because of DX11/OpenGL driver overhead/issues on the red side; Async is playing a very small role there. Pascal is still much more efficient

Rx480 was expected to be a low teir gpu that could beat the 970 and maybe reach the 980. 

 

If you expected anything more than you didnt follow what AMD was saying about their card. 

 

We are also still waiting for ither polatis chips. Amds goal was to start low end and build up. While nvidia is to start big then go small. 

[Curufinwe_wins]

13 minutes ago, goodtofufriday said:

Rx480 was expected to be a low teir gpu that could beat the 970 and maybe reach the 980. 

 

If you expected anything more than you didnt follow what AMD was saying about their card. 

 

We are also still waiting for ither polatis chips. Amds goal was to start low end and build up. While nvidia is to start big then go small. 

There are no bigger polaris chips. AMD's goal was just to not compete at all at the high end, let yields increase over time then smack with Vega at the high end.

 

That is still the goal. Unfortunately, Pascal high end has been such a huge hit that AMD is wishing it had SOMETHING to sell above the 480, hence talks about a 490 being two 480 gpus. That imho probably won't happen though and AMD will just wait it out.

[2FA]

EDIT: My reading comprehension is bad.

[AlTech]

4 minutes ago, DeadEyePsycho said:

EDIT: My reading comprehension is bad.

think of VEGA as big Polaris because it will just be bigger Polaris with GDDR5X or HBM2.

[Briggsy]

Without moving the goal-post every 5 minutes, and because Pascal and Polaris are both die shrinks of existing architecture (for the most part), the argument boils down to this: which Architecture leverages the FinFET die-shrink better - Pascal or Polaris?

 

(See spoiler below for Math)

 

If clock speed, core count and average gaming performance are removed from the equation:

RX 480 loses 8.6% performance from the die shrink and uses 61.7% of an R9 390's power draw.

GTX 1060 gains 14.22% performance from the die shrink and uses 74.4% of a GTX 980's power draw.

 

So in respect to die shrink, AMD made more headroom in efficiency, while Nvidia Made more gains in performance. We already knew that AMD were targeting efficiency and Nvidia were targeting Clock speeds. The fact that Nvidia still gains in efficiency when destroying core count of its predecessor is noteworthy.

 

Spoiler

 

AMD - Polaris vs. Hawaii

2304 vs. 2560 cores

1266Mhz vs. 1000Mhz - 26.6% increase

100% vs. 96% performance average

(100/96) / 1.266 * (2560/2304) 

= 8.6% performance loss from die shrink if clocks, core count and performance are neutralized.

163W / 264W typical gaming power draw

= 61.7% of power

 

Nvidia - Pascal vs. Maxwell

1280 vs. 2048 cores

1506 vs. 1064 base clocks - 41.5% increase

100% vs. 99% performance average

(100 / 99) / 1.415 * (2048 / 1280) 

= 14.22% performance gains from die shrink if clocks, core count and performance are neutralized.

116W / 156W typical gaming power draw

= 74.4% of power

 

 

Edited July 27, 2016 by Briggsy
derped on AMD's math