(Update)Nvidia GP104 Pascal GPU Spotted in a shipment – Roughly ~350mm^2 Die Size, May not feature HBM, but GDDR5X memory

[marldorthegreat]

It is on the image itself, it is not a text you will find using control + F. 

 

 

650 WATTS TEC is what @don_svetlio is referring to.

 

However, i personally do not see it using 650w. It might be total system consumption like others have stated, or the usual over-estimation that us done with every GPU. After all, we can run 980 Ti's on 500w PSU's, when the recommendation is 600w.

its not going to use 600 watts, it will use 250 watts. 

[don_svetlio]

its not going to use 600 watts, it will use 250 watts. 

Compute is back so more like 300 (780 TI/Titan/Titan Black used 300)

[Zah]

If it's GP104, does that mean there will be a GP100 as well like with Maxwell GM204?  I won't replace my 980ti until then.

[marldorthegreat]

Compute is back so more like 300 (780 TI/Titan/Titan Black used 300)

yeah, thats what made maxwell so efficient, so there taking a step backwards? 

[Coaxialgamer]

Would be pretty odd for Nvidia to come out swinging like that out of the gate. They normally reserve those cards until the very end, after they've already milked the enthusiasts dry with normal Ti and Titan cards. 

i just can't see a 650w card . But i can't see them using that measure for a full system in this case.

[don_svetlio]

yeah, thats what made maxwell so efficient, so there taking a step backwards? 

Either that or get creamed with DX12

[ACatWithThumbs]

If it's GP104, does that mean there will be a GP100 as well like with Maxwell GM204?  I won't replace my 980ti until then.

It's very likely and with the die size being so small the full fledged chip could be around 40% bigger.

The Titan X has a die size of 601mm the full fledged Pascal chip is probably between 500mm-600mm as well.

I highly doubt Pascal/16nm maxes out at 350mm.

 

[-BirdiE-]

Guys.. it says 650W FOR ____________

 

Likely a PSU for a test system or something. No chance that's the card's power draw under load.

[Zah]

It's very likely and with the die size being so small the full fledged chip could be around 40% bigger.

The Titan X has a die size of 601mm the full fledged Pascal chip is probably between 500mm-600mm as well.

I highly doubt Pascal/16nm maxes out at 350mm.

 

Good, guess I have about 1-2 more years out of this card then.

[FloRolf]

yes pleaz! My 970 is getting old...

[GoodBytes]

The chip doesn't consume 650W. In fact, every time you see these Wattage from the GPU, it is Nvidia/AMD is assuming a PSU for the target market average computer configuration, and assumes you'll go and buy the crappest PSU.

If you use a measurement tool, you can see that a computer under intense load it doesn't pass 400Watts. At work we measured the power usage a while back. While I don't recall the numbers (If you are intrested, I can see in getting those number again, we wrote it somewhere, just need to find it) but I remember being under 400W, and we have powerful gaming PCs, almost latest and greatest hardware at the time of purchase: GTX 980Ti, Samsung 850 Pro 512GB, CPU is a Core i7 8 core CPU with HT (Haswell-E), 32GB of RAM, ASUS X99-E WS motherboard. Kill-A-Watt device was used to measure.

So yes, many people buy overkill power supplies, but ideally, you want to be under normal load (say: gaming) at about 50% PSU load, which is where the PSU is the most efficient.

[spartaman64]

The Nano thermal throttles.

it has a really small cooler

[Lethal Seraph]

I hope nVidia releases these soon. I'm already ready to ditch my 980 Ti. Can't wait for Pascal!

It's gonna be awhile for me though. Waiting for a an HoF Pascal gpu is a bitch.

[That Norwegian Guy]

you want to be under normal load (say: gaming) at about 50% PSU load, which is where the PSU is the most efficient.

 

Spread the gospel! Too many still need this education, every time I hear "overkill PSU" I barf a little in my mouth.

[bobhays]

I wonder why these chips has so many pins? Isn't the entire point of HBM to include them on the package, meaning the only pins that actually goes into the pcb are for power delivery and signals to DisplayPort/HDMI?

 

Or are these chips non package, non HBM connected?

 

Addendum: Polaris is the architecture, the new series of AMD cards are still called Arctic Island.

I don't think HBM is on the package, it just takes less space than GDDR5. I maybe mistaken but I read somewhere about how HBM takes a lot of energy and I feel like it would be easier to cool if it wasn't together but that being said high end GPU's use hundreds of watts so whats 30-50 more.

[Citadelen]

I don't think HBM is on the package, it just takes less space than GDDR5. I maybe mistaken but I read somewhere about how HBM takes a lot of energy and I feel like it would be easier to cool if it wasn't together but that being said high end GPU's use hundreds of watts so whats 30-50 more.

HBM is around 5x more efficient than GDDR5.

[AlwaysFSX]

Wait wait wait, 650 watts? What?

Could be a Gigabyte card? They are toting their current coolers are good for 600 Watts, maybe the new ones will say they're good for 650? I highly doubt any single GPU right now is going to draw 650 Watts, neither AMD or Nvidia would do it.

[Sakkura]

It's very likely and with the die size being so small the full fledged chip could be around 40% bigger.

The Titan X has a die size of 601mm the full fledged Pascal chip is probably between 500mm-600mm as well.

I highly doubt Pascal/16nm maxes out at 350mm.

 

 

350mm² isn't really small, to be fair, but yeah there will almost certainly be a bigger version eventually. It may take a while longer though, and will probably cost a ton initially. It takes a while for yields to become good enough for a GK110 or GM200 type of chip. ~600mm² is a huge chip. I mean compare to CPUs - a Core i7-4790K clocks in at 177mm², and even that $7,000 18-core Xeon E7 is around 660mm².

[BuckGup]

Are these shipping forms numbers whatever you call it public? Or is some dude inside the company screenshoting them and stealing them. To me it seems like highly classified information

[patrickjp93]

HBM is around 5x more efficient than GDDR5.

Only at low clock speeds. I posted an in-depth article about this a while back. Above 1GHz HBM quickly becomes less efficient than GDDR5.

[patrickjp93]

350mm² isn't really small, to be fair, but yeah there will almost certainly be a bigger version eventually. It may take a while longer though, and will probably cost a ton initially. It takes a while for yields to become good enough for a GK110 or GM200 type of chip. ~600mm² is a huge chip. I mean compare to CPUs - a Core i7-4790K clocks in at 177mm², and even that $7,000 18-core Xeon E7 is around 660mm².

And now we have Knight's Landing chips that are 724mm²

[Citadelen]

Only at low clock speeds. I posted an in-depth article about this a while back. Above 1GHz HBM quickly becomes less efficient than GDDR5.

But does it need to be clocked that high, are there any benefits? HBM is really fast anyway, so bandwidth doesn't need to be increased.

[DXMember]

But does it need to be clocked that high, are there any benefits? HBM is really fast anyway, so bandwidth doesn't need to be increased.

to paraphraze linux -

we know about the drawbacks of overclocking

we know about the higher heat output

we know about the higher power consumption

and even the potentiol for hardware degradation

IT"S ABOUT THE SPEEEEEEEEEEEEEED....

 

^{source - look up hasswell-e overclocking guide}

[Geralt]

Nvidia: The way it's meant to be burned (calories)

 

 

We're gonna need an updated one.

 

Nvidia Pascal: The way it's meant to be grilled

[ShiroArikawa]

I hope so, I'm gonna have a field day if that were the case

prepare the 2000w psu