Nvidia Pascal Announcement

[Orion]

Do you have a link? All reviews so far of custom 480 cards have not broken 1400Mhz for a stable benchmark. It's possible in the future under water but have not seen any on air.

https://lockgamer.com/2016/07/22/qual-o-ganho-de-performance-com-a-rx-480-a-1400-mhz/

Supposedly haven't translated so haven't verified might or might not be wrong.

Please define "some circumstances", you are actually saying R480 can be faster than FuryX/980Ti now? in what world? and in what benchmarks? the card can't even surapass 1060 in most benchmarks let alone above that!

In some hitman benchmarks rx 480 is at 30% over 980.
1060 is around 980 performance and supposedly the rx480 can be over 50% faster than 1060 in Doom vulkan.

In dx12 and vulkan performance is likely to shine in the future, amd's has not optimized as well as nvidia in dx11 and opengl supposedly.

 

[DavidGraham]

Yes, these are special cases related to optimizations and sponsored titles, the same can be had with other NVIDIA cards which can punch above their weight to "abnormal" levels in some games or some scenes. (for example the 1060 can surpass Fury X performance using Vulkan in Talos Principle). Comparisons using outlier cases are not valid.

Here see for example: RX480 is 20% faster than 1060 in Vulkan when not CPU limited, if CPU limited the 1060 is up to 42% faster in Vulkan as well!

 

[Deleted member 2197]

https://lockgamer.com/2016/07/22/qual-o-ganho-de-performance-com-a-rx-480-a-1400-mhz/

Supposedly haven't translated so haven't verified might or might not be wrong.

Ok. Seems he applied a bios mod and the benchmark score is valid. Only issue is the 1400 clock is not sustainable at 88c for everyday use ... I also see a 1510 clock for someone else but seems to be LN2.

 

[Grall]

Why would 88C not be sustainable? Sadly, my R9 390X goes up to as much as 94C at max load, but it runs fine at that temp. If it hadn't, it would have throttled, but it doesn't. It sits pegged at 1050MHz.

 

[CSI PC]

The same thing happened with the original Titan for like 6 months. I think Pascal is going to shake out sorta like the Kepler generation.

The first titan is partially disabled to serve as a halo product.

Then 6ish months later, that same gpu becomes fully enabled with higher clocks as the X80 Ti.

Then a few months after that, a refreshed titan drops with everything enabled, slightly higher clocks, still, and perhaps some kind of clamshell setup to get a truly egregious amount of vram.

So we've got 2016's titan now. I wouldn't be terribly surprised to see a 1080 Ti show up in a couple months to ruin Vega's day and then gp102 gets one more showing with 2017's Titan just to ensure that there's always a titan halo card in the lineup.

Although the situation is a bit different in some ways.
The GK110 that the Titan is based upon was first launched as a Tesla product as K20X, and then shared with the other segments, all had the reduced Cuda cores.
So it makes sense the full GK110 would be next iteration and that was not only the 780ti but also Tesla K40, both released at same time.
Titan Black was introduced 1 year later to the original to make Titan the flagship consumer model again by being full GK110 with highest clocks and other attributes not found on the 780ti model such as improved DP and VRAM, it was stupidly priced from what I can remember but it was also a cheaper 'prosumer' card to the Tesla range.

With the introduction of Volta late next year (definitely happening as the big die GPU is required for multiple supercomputer contracts and recently confirmed to be 16nm) I doubt we will see Pascal Titan updated in same way it was done with Kepler, but also comes down to what synergy if any there is between Pascal Titan and Tesla/Quadro, because it is not really ideal to do a GPU just for one segment these days, even back then with Kepler GK110 the full spec update was across all segments and not just Titan Black and again was released 1st as the Tesla K40 by around one quarter.
A dedicated GPU to one segment works OK with P100 models but then that is at a premium price and is a rather unique product for now with a lot of interest.
Cheers

 

[Ryan Smith]

So it is debatable whether this is the same mixed-precision FP32 Cuda core from GP100 but fixed as FP16x2 operation, or another unique kind of Cuda core.

I am going to avoid stepping out on a limb and saying it's transistor-for-transistor identical, due to the CC differences. But it serves the same function.

Now where I am a little confused, is what goes on in GP104; is there a pair of standard FP16 units ? The way it's been mentioned as a single "FP16x2" unit is confusing

There is a single CUDA core able to process a pair of vec2 packed FP16 instructions per clock cycle. It is alternatively capable of processing 1 FP32 instruction, though it is not used this way on GP104.

To differentiate it from the normal FP32 core, I call it FP16x2; NVIDIA hasn't given it a unique designation outside of the company.

 

[CarstenS]

Fury would depend on the cost of HBM. But by reasonable I mean 600-700$ for single gpu cards. AMD needs to undercut to increase marketshare. Given identical performance and price, nvidia is likelier to be chosen given their very good market perception.

But but... HBM was key to improve power efficiency and bring unprecedented memory bandwidth...
FWIW, cost seem to have dropped massively, because as of now, Fury X sells for 430 EUR and Fury for 350 EUR here in GER (unfortunately not applicable to the Nano. )

Some of the 200$ 4 GB are actually 8GB cards, just check if they've got an 8GB under the 4GB sticker and change bios.

I know (and own one of these) - but neither were or are those widely available (were they Nvidia, people had called it a paperlaunch for little stock) nor are they at 200 $ - rather 250 EUR (~270$) now.

As for AIB, some have reached over 1400Mhz, and I've read rumors some might come at over 1500Mhz.

Have reached with manual OC or factory default? Because you seem to throw a lot of ingredients into a mix that do not blend well.

FWIW, one of the more trusted sites say, their Sapphire RX 480 Nitro+ runs in Anno 2205 at 1080p with a stock boost of 1260:
https://forum.beyond3d.com/posts/1932623/
I presume it is hitting it's powerlimit, because temps do not seem to be critical from what I've seen with our sample of the Nitro+

 

[Orion]

But but... HBM was key to improve power efficiency and bring unprecedented memory bandwidth...
FWIW, cost seem to have dropped massively, because as of now, Fury X sells for 430 EUR and Fury for 350 EUR here in GER (unfortunately not applicable to the Nano. )

I know (and own one of these) - but neither were or are those widely available (were they Nvidia, people had called it a paperlaunch for little stock) nor are they at 200 $ - rather 250 EUR (~270$) now.

Don't know if it's available in europe, but one of the best cash back cards in the U.S, the citi double cash visa and also the citi costco visa cards, at least it seems(check just in case) appear to have a feature called price rewind(must register product online on their site's price rewind area), they will check hundreds of the top online sites and if a better price is found anywhere they will give you the difference back.

So at least in the U.S. with such easy to get no annual fee cards, iirc, it seems you'll get the card at or below the advertised price. Don't know about europe, though.

Have reached with manual OC or factory default? Because you seem to throw a lot of ingredients into a mix that do not blend well.

FWIW, one of the more trusted sites say, their Sapphire RX 480 Nitro+ runs in Anno 2205 at 1080p with a stock boost of 1260:
https://forum.beyond3d.com/posts/1932623/
I presume it is hitting it's powerlimit, because temps do not seem to be critical from what I've seen with our sample of the Nitro+

Probably running in quiet mode, the card has a physical button to switch between quiet mode and high performance mode.

The Sapphire TriXX 3.0 overclocking software is said to be coming soon not yet available. So its optimal OC is probably not yet easily achievable, around 7-10%performance increase with the current moderate OCs, and from comment sections it seems that's OC of just the gpu, ram OC will give additional performance which some estimate additional 7-10% more.

PS
Besides doom and hitman, will add that I heard mirror's edge at hyper settings at 1440p can use more than 6GB for optimal performance, according to some benches. Some benches show 390x 8GB beating 980ti 6GB at 1440p. Even at 1080p at hyper settings it seems Mirror's edge can't run on 970 or fury, probably due to 4GB limit.

That's a current game, that will likely run suboptimally in 1060 6GB version at highest setting 1440p. As said even the 980ti is beaten in some benches by a 390x, how will the weaker than 980ti 1060 fare?

What about future more memory intensive games. Not only some dx12, some vulkan but also some memory intensive future games should give the performance crown to rx 480.

In future memory intensive games I can see OC rx 480, even at more reasonable 1380-1390Mhz OC levels offering better than 980ti performance.

 

[seahawk]

But we also know that at higher resolution (higher workload) the RX480 struggles even more to keep the Boost up.

 

[Orion]

But we also know that at higher resolution (higher workload) the RX480 struggles even more to keep the Boost up.

Will have to check, AIBs have superior cooling and power than reference, it is said they can sustain higher clocks than reference.

 

[CSI PC]

Will have to check, AIBs have superior cooling and power than reference, it is said they can sustain higher clocks than reference.

Worth noting all reviews are mentioning (so far all air based) having to increase the fan to excessive levels to cool when manual OC even for the Sapphire Nitro 480 OC and that cooling system is very capable considering the temps it is able to maintain with default BIOS OC (early days but some reports suggesting the 480 Nitro OC needs around 80% fan speed with manual OC), which makes sense when one sees the power draw for the whole card when overclocked.
Applies also to Pascal but to a lesser degree as shown with the previous Tom's charts I linked earlier.
Cheers

 

[ieldra]

I am going to avoid stepping out on a limb and saying it's transistor-for-transistor identical, due to the CC differences. But it serves the same function.

There is a single CUDA core able to process a pair of vec2 packed FP16 instructions per clock cycle. It is alternatively capable of processing 1 FP32 instruction, though it is not used this way on GP104.

To differentiate it from the normal FP32 core, I call it FP16x2; NVIDIA hasn't given it a unique designation outside of the company.

So... In practice there are actually 129 FP32 units in an SM on GP104 ? Is there any way of using that FP16x2 for normal FP32 ?

 

[Arun]

There is a single CUDA core able to process a pair of vec2 packed FP16 instructions per clock cycle.

Has this been confirmed by NVIDIA? This would possibly imply that on GP100, it's really more like 1xFP32+1xFP16 where the FP32 unit can be used in FP16 mode (likely with clock-gating and possibly even power-gating). Note that this is a datapath implementation detail; it would not be visible in the ISA.

Also has it been 100% confirmed that GP102 doesn't support 2xFP16? It is unfortunate if so, depending on the direction that the industry goes (partly dependent on what console vendors choose for XBox Scorpio and PS5), this could have been a nice bit of future-proving if developers could use 2xFP16 for games in the future

This is already supported in DX12 just nobody's going to bother if it's not exposed (it would be nice if it could be used to reduce register pressure at least but I'm not sure NVIDIA supports free FP32<->FP16 conversion on GP104 so that might just end up wasting too many ALU cycles).

 

[lanek]

Worth noting all reviews are mentioning (so far all air based) having to increase the fan to excessive levels to cool when manual OC even for the Sapphire Nitro 480 OC and that cooling system is very capable considering the temps it is able to maintain with default BIOS OC (early days but some reports suggesting the 480 Nitro OC needs around 80% fan speed with manual OC), which makes sense when one sees the power draw for the whole card when overclocked.
Applies also to Pascal but to a lesser degree as shown with the previous Tom's charts I linked earlier.
Cheers

I have noot look in details on thoses coolers, what coldplates are used or fins material and assembly, but visualy they seems more or less as the one you would have find on gpu's as 980TI and Fury ... thoses should been able to dissipate around 300W without any problem.

 

[CSI PC]

I have noot look in details on thoses coolers, what coldplates are used or fins material and assembly, but visualy they seems more or less as the one you would have find on gpu's as 980TI and Fury ... thoses should been able to dissipate around 300W without any problem.

Yeah but it will not be identical when comparing dissipation from 28nm to 14/16nm due to various factors including density and die size, and note we are talking about when pushed beyond optimal performance envelope and scaling, which again is not identical to respective manufacturer's 28nm.
I used the Sapphire 480 Nitro OC as that has a dedicated cooling design, the only other custom 480 reviewed so far is the Asus 480 Strix and that is using the same design from their Strix 1080 so may not be ideal.

Cheers

 

[Ryan Smith]

Has this been confirmed by NVIDIA?

Yes, it has been confirmed by NVIDIA.

This would possibly imply that on GP100, it's really more like 1xFP32+1xFP16 where the FP32 unit can be used in FP16 mode (likely with clock-gating and possibly even power-gating). Note that this is a datapath implementation detail; it would not be visible in the ISA.

NVIDIA has also already confirmed that GP100's "FP32" units are all FP16x2 units. You get 1 vec2 FP16 or 1 FP32 per cycle.

 

[seahawk]

Yeah but it will not be identical when comparing dissipation from 28nm to 14/16nm due to various factors including density and die size, and note we are talking about when pushed beyond optimal performance envelope and scaling, which again is not identical to respective manufacturer's 28nm.
I used the Sapphire 480 Nitro OC as that has a dedicated cooling design, the only other custom 480 reviewed so far is the Asus 480 Strix and that is using the same design from their Strix 1080 so may not be ideal.

Cheers

It is not a problem of the cooler being unable to handle it, but the problem of getting enough heat out of the small die. The heatpipes are the limiting factor in how much heat they can draw from the die. maybe a vapour chamber would really benefit RX480.

 

[Deleted member 13524]

I know (and own one of these) - but neither were or are those widely available (were they Nvidia, people had called it a paperlaunch for little stock) nor are they at 200 $ - rather 250 EUR (~270$) now.

"Were they nVidia"? The 4GB was factually a paperlaunch for pretty much all markets but the US.
However, all RX480 cards that are being put to market are being sold really fast, so even if the demand for the "$199" 4GB model was underestimated, they're probably getting higher margins with the 8GB model nonetheless.
This was not the most consumer-friendly decision... then again AMD needs to make money ASAP.

 

[CarstenS]

Don't know if it's available in europe, but one of the best cash back cards in the U.S, the citi double cash visa and also the citi costco visa cards, at least it seems(check just in case) appear to have a feature called price rewind(must register product online on their site's price rewind area), they will check hundreds of the top online sites and if a better price is found anywhere they will give you the difference back.

So at least in the U.S. with such easy to get no annual fee cards, iirc, it seems you'll get the card at or below the advertised price. Don't know about europe, though.

Wow, that's one hell of a spin if I've ever read one.

Probably running in quiet mode, the card has a physical button to switch between quiet mode and high performance mode.

Probably got lost in translation if you cared enough to actually click and read the review, but they explicitly stated that this is the case with both settings of the BIOS switch. Originally wanted to include that info, but forgot. That's also why I arrived at the power-limit conclusion.

"Were they nVidia"? The 4GB was factually a paperlaunch for pretty much all markets but the US.
However, all RX480 cards that are being put to market are being sold really fast, so even if the demand for the "$199" 4GB model was underestimated, they're probably getting higher margins with the 8GB model nonetheless.
This was not the most consumer-friendly decision... then again AMD needs to make money ASAP.

Sold really fast or manufacturing problems getting enough out there - where they Nvidia? All a matter of perspective, it seems.

Anyway: This money-making argument would make sense, if they were genuine 4 GiByte models, but they were not.

 

[Deleted member 2197]

Nvidia Announces Quadro P6000 with 24GB gddr5x

Nvidia has announced two new graphics cards for the business market. The Quadro P6000 and P5000 are equipped with Pascal GPUs and feature Nvidia respectively, 24GB and 16GB gddr5x.

Quadro P6000 uses the P100 GPU with a full 3840 Cudacores activated, however like the Tesla P100 Accelerator the product will be available later this year. Nvidia combines the card with 24GB gddr5x.The P6000 has four DisplayPort 1.4 interfaces and dual-link DVI, which according to Nvidia 5k four screens at 60Hz van be controlled.

The Quadro P5000 has been fitted with the GP104 GPU with 2560 cuda cores. The GPU will have 16GB vram, but again it comes with gddr5x. The card has four display ports and dual-link DVI.

The video engine supports decoding hevc- and h264 video on 10bit and 12bit. The Quadro's would handle 4k images at 120Hz and 240Hz and even 8k video at 30Hz.
The Quadro P6000 and P5000 will appear in October this year.

http://www.guru3d.com/news-story/nvidia-announces-quadro-p6000-with-24gb-gddr5x.html