Nvidia Pascal Announcement

[Razor1]

Yes, but because it's not meant to, not because it won't be close, I was saying that in perspective of Nvidia being able to easily trump AMD, and thus doing so. If they don't release fully enabled Titan at all, AMD might as well price their offering similarly(cause it would look more appealing, heck RX 480 looks better to a lot of people than 1060 right now), which is not happening.

Full Vega might not escape all the problems Polaris is having, but coming out almost a year after it should achieve 16TF, which should put right between 3584 and 3840 Titans in performance.

IF Vega doesn't escape all the problems Polaris is having, Vega is screwed, its a bigger chip, its will be a gas guzzler and will run across the same issues Fury line ran up against.

And where does rx480 look better than the 1060, outside of Doom, Dx12 is 5% lead at most, actually closer to a tie and Dx11 is like 15% in favor of the gtx 1060. All the while the gtx 1060 having a 20% power advantage.

And Flops don't matter, the 1060 has less flops than the rx480 and yeah you see the picture.

 

[ieldra]

I don't them cutting another two SMs from GP102 to make a Ti. It would have the same vram as the Titan X (Pitan X to me) and marginally less ALU; will cannibalize Pitan X HPC sales.

Imo possibility of this Pitan X being rebranded as Ti and full GP102 with 24Gb vram (maybe even 12GT/s) becoming the Pitan X is higher

 

[Kaotik]

I don't them cutting another two SMs from GP102 to make a Ti. It would have the same vram as the Titan X (Pitan X to me) and marginally less ALU; will cannibalize Pitan X HPC sales.

Imo possibility of this Pitan X being rebranded as Ti and full GP102 with 24Gb vram (maybe even 12GT/s) becoming the Pitan X is higher

What HPC sales? By the looks of it, they cut most of the FP64 compared to GP100 to get transistor counts down by the billions

 

[CSI PC]

Can anyone explain where they're getting 44TOP/s Int8?

It see seems its just 4x the fp32 rate, but that's considering the use of FMA so it's already two ops per cycle.

22TOP/s makes sense to me as it would be 4x the int32 rate with one operation per instruction, assuming 1:1 ratio of FPU:ALU

Just read about dp4a and dp2a2, that would make sense

I quickly skimmed that thread I posted earlier, one aspect not touched upon is that a big use for int8 would be accerelated inference.
But I have concerns that those interested in this would still want a GPU with both FP32/FP16 for other associated functions-operations, they will not want multiple GPUs for the various tasks.
Cheers

Edit:
My context is to do with the decision whether it is a mixed-precision Cuda core (preferable IMO) or more traditional Cuda core as used with GP104 that does not support FP16.

 

[iMacmatician]

I don't them cutting another two SMs from GP102 to make a Ti. It would have the same vram as the Titan X (Pitan X to me) and marginally less ALU; will cannibalize Pitan X HPC sales.

Imo possibility of this Pitan X being rebranded as Ti and full GP102 with 24Gb vram (maybe even 12GT/s) becoming the Pitan X is higher

Could NVIDIA cut down the INT8 rate for a GP102-based “1080 Ti”?

 

[ieldra]

Yeah I agree it seems counterintuitive to split mixed precision (and doubled fp16) and dp4a/2a over different products. They seem complementary for the use case.

Titan X (maxwell; Mitan X) was a great cost effective card for neural networking. Pitan X will be the same, that's what I meant by hpc sales. If they had kept 2xfp16 but still gutted fp64 it would have been perfect

Could NVIDIA cut down the INT8 rate for a GP102-based “1080 Ti”?

Artificially, sure.

BTW guys, I was under the impression gp100 didn't have dedicated fp16 units, is this wrong?

 

[Deleted member 13524]

Why not? The previous Titan X is "clocked" at 1000 MHz yet it can easily be overclocked to 1450-1500 MHz. At "worst" a 45% overclock.

Hum.. what I see is a 1450MHz max boost from a reference 1215MHz, with vcore adjustments:
http://www.anandtech.com/show/9059/the-nvidia-geforce-gtx-titan-x-review/17

This is a 20% overclock, which is spectacular (mainly compared to the headroom in AMD GPUs) but not anything near the 33% overclock needed for the GP102 to reach 2GHz and even further away to the 45% that you suggest "at worst".

Maybe you need some new glasses then, because cause it's clear.
(...)
come on now, is this Beyond3d forum or a kindergarten?

This forum is also preferably not for this kind of rudeness.. probably even less so for a 66 post fresh account from late last year..

 

[DuckThor Evil]

Hum.. what I see is a 1450MHz max boost from a reference 1215MHz, with vcore adjustments:
http://www.anandtech.com/show/9059/the-nvidia-geforce-gtx-titan-x-review/17

This is a 20% overclock, which is spectacular (mainly compared to the headroom in AMD GPUs) but not anything near the 33% overclock needed for the GP102 to reach 2GHz and even further away to the 45% that you suggest "at worst".

Ok, but why are you taking the actual max boost clock from Maxwell and only the stated paper boost clock from Pascal?

In Anandtech's 1080 review, the actual max boost clock for the 1080 was 1898Mhz with stock settings, so quite a bit higher than the rated 1733Mhz (9.5% to be exact). You should assume similar actual boost clock difference for the Pascal Titan as well, meaning 1531 * 9.5% = 1676Mhz. That's a valid assumption imo, then an additional 10-15% overclock should be possible.

 

[ieldra]

Hum.. what I see is a 1450MHz max boost from a reference 1215MHz, with vcore adjustments:
http://www.anandtech.com/show/9059/the-nvidia-geforce-gtx-titan-x-review/17

This is a 20% overclock, which is spectacular (mainly compared to the headroom in AMD GPUs) but not anything near the 33% overclock needed for the GP102 to reach 2GHz and even further away to the 45% that you suggest "at worst".


This forum is also preferably not for this kind of rudeness.. probably even less so for a 66 post fresh account from late last year..

The boost clock its rated for is meaningless. This will probably clock to 1700mhz out the box, just like how the 1060 rated for 17xx boosts to 1850, same with 980ti, same with 1080, etc etc.

I think 2ghz will be very commonly achievable on this card with adequate cooling

 

[Benetanegia]

This forum is also preferably not for this kind of rudeness.. probably even less so for a 66 post fresh account from late last year..

I'll try to tone it down in the future, but really they are just ridiculous answers to ridiculous statements. I couldn't even tell at the time if I should take them seriously or not.

 

[CarstenS]

IMO the only truly good thing that came out of Fiji was the Nano, which seems like a nice deal even today because it's hitting close to 300€.

300 EUR? Link please, so I can finally get one. I have the price search tab for the nano pinned in my browser, and here in GER it's ~480 EUR for weeks now. I'd actually be happy about 3xx EUR...

 

[Ike Turner]

300 EUR? Link please, so I can finally get one. I have the price search tab for the nano pinned in my browser, and here in GER it's ~480 EUR for weeks now. I'd actually be happy about 3xx EUR...

Was lucky to get a brand new Fury X for €479 2 months ago...from Amazon..just keep your eyez peeled..I was checking every day for a month to find a good deal..

 

[DuckThor Evil]

300 EUR? Link please, so I can finally get one. I have the price search tab for the nano pinned in my browser, and here in GER it's ~480 EUR for weeks now. I'd actually be happy about 3xx EUR...

Ok not €300, but a good price anyway. Been like that for a long time already, I got one for €399 and it came with Far Cry Primal. And yeah sorry it's in Finland, but maybe it's a high time for you do that holiday trip to our somewhat lovely country (sorry for the off topic)

https://www.jimms.fi/fi/Product/Sho...d-r9-nano-4gb-hbm-dp-hdmi-pcie-3_0-tarjousera

 

[CarstenS]

Was lucky to get a brand new Fury X for €479 2 months ago...from Amazon..just keep your eyez peeled..I was checking every day for a month to find a good deal..

Thanks, the X is on sale for 430 EUR right now (and available) - but I want the Nano.

Ok not €300, but a good price anyway. Been like that for a long time already, I got one for €399 and it came with Far Cry Primal. And yeah sorry it's in Finland, but maybe it's a high time for you do that holiday trip to our somewhat lovely country (sorry for the off topic)

https://www.jimms.fi/fi/Product/Sho...d-r9-nano-4gb-hbm-dp-hdmi-pcie-3_0-tarjousera

Thanks as well. 349 EUR sounds great! Now I only need to figure out how to get through ordering with google translate and hope that delivery to GER is not another 99 EUR

 

[Deleted member 13524]

In Anandtech's 1080 review, the actual max boost clock for the 1080 was 1898Mhz with stock settings, so quite a bit higher than the rated 1733Mhz (9.5% to be exact). You should assume similar actual boost clock difference for the Pascal Titan as well, meaning 1531 * 9.5% = 1676Mhz. That's a valid assumption imo, then an additional 10-15% overclock should be possible.

That's still not going into 2GHz, 1.15*1676 = 1900MHz and that's only max boost clock where the card doesn't spend all that much time one -> Titans are traditionally sold with reference designs only, so there won't be any triple-slot coolers helping the card to maintain lower temperatures while overclocked.
But we'll see.

300 EUR? Link please, so I can finally get one. I have the price search tab for the nano pinned in my browser, and here in GER it's ~480 EUR for weeks now. I'd actually be happy about 3xx EUR...

IIRC overclockers.co.uk had the Nano at a discount down from its current £330 and the price was close to £300. And Brexit has brought the pound to its knees..

I think 300-330€ is where both the non-X Fury and the Nano will end up at sometime this year.

 

[CSI PC]

That's still not going into 2GHz, 1.15*1676 = 1900MHz and that's only max boost clock where the card doesn't spend all that much time one -> Titans are traditionally sold with reference designs only, so there won't be any triple-slot coolers helping the card to maintain lower temperatures while overclocked.
.

I do not see any mention about increasing both power and temperature target along with subtly changing the fan profile to be a bit more aggressive, which is required if maintaining boost.
This process is not complicated for any owner (not critical of any reviews because this is beyond using it out of the box), and importantly this does not affect warranty (according to some reviews I have read).

If you spend time then you can sustain 2050MHz with a fully manual configuration, HardwareCanucks required 2 hours to 'dial-in' their 2126MHz OC and memory OC of 5670MHz.
I agree most will not go that far and will just focus on power and temperature target in Boost 3 and importantly adjust the fan profile to be a bit more aggressive so it kicks in 60% at 80degrees, that is enough to stop throttling.
Unfortunately most reviews set the fan to 100% instead of finding that ideal fan speed and temp.
Also it did not help there was a fan issue with initial launch-release.

So initially looks like the reference Titan can be configured in a way without breaking warranty to provide a consistent and good boost, the only unknown is what the fan speed will need to be at the 80degree target.
But yeah, I would still prefer to wait for a 1080ti myself
Cheaper and with a quieter cooler, but that is the price for not being able to wait or needing the VRAM Pascal Titan probably has, assuming they do not do anything else to differentiate flagship Titan from the 1080ti.
Cheers

 

[ieldra]

I do not see any mention about increasing both power and temperature target along with subtly changing the fan profile to be a bit more aggressive, which is required if maintaining boost.
This process is not complicated for any owner (not critical of any reviews because this is beyond using it out of the box), and importantly this does not affect warranty (according to some reviews I have read).

If you spend time then you can sustain 2050MHz with a fully manual configuration, HardwareCanucks required 2 hours to 'dial-in' their 2126MHz OC and memory OC of 5670MHz.
I agree most will not go that far and will just focus on power and temperature target in Boost 3 and importantly adjust the fan profile to be a bit more aggressive so it kicks in 60% at 80degrees, that is enough to stop throttling.
Unfortunately most reviews set the fan to 100% instead of finding that ideal fan speed and temp.
Also it did not help there was a fan issue with initial launch-release.

So initially looks like the reference Titan can be configured in a way without breaking warranty to provide a consistent and good boost, the only unknown is what the fan speed will need to be at the 80degree target.
But yeah, I would still prefer to wait for a 1080ti myself
Cheaper and with a quieter cooler, but that is the price for not being able to wait or needing the VRAM Pascal Titan probably has, assuming they do not do anything else to differentiate flagship Titan from the 1080ti.
Cheers

Also worth mentioning that anyone buying this card for games can probably afford a waterblock, and anyone buying it with the intention of overclocking it will at least be using a third party air cooler

 

[ShaidarHaran]

Also worth mentioning that anyone buying this card for games can probably afford a waterblock, and anyone buying it with the intention of overclocking it will at least be using a third party air cooler

Ding ding ding, we have a winner! This is precisely my intention for, and interest in this card. Just need to see how it OCs before I decide between new Titan X or SLI 1070 for 4k60.

 

[RecessionCone]

Yeah I agree it seems counterintuitive to split mixed precision (and doubled fp16) and dp4a/2a over different products. They seem complementary for the use case.

Titan X (maxwell; Mitan X) was a great cost effective card for neural networking. Pitan X will be the same, that's what I meant by hpc sales. If they had kept 2xfp16 but still gutted fp64 it would have been perfect


Artificially, sure.

BTW guys, I was under the impression gp100 didn't have dedicated fp16 units, is this wrong?

Yes. This is wrong. GP100 has dedicated FP16 units. GP102, 104, 106 have dedicated Int8 units.

 

[Dangerman]

A cut-down GP102 that is likely to be under <500mm2 at 1200USD and at GBP being slightly less or equal or knowing Nvidia way over £1200? No thanks. I could see them if it was full fat GP102 with 24GB of VRAM.