Nvidia Pascal Announcement

Well I do think Nvidia would want to accelerate the next down from GP100 to strengthen their Tesla and maybe Quadro range (think priority would be Tesla).
So sort of makes sense for the GP102 to go into production as soon as achievable for the business side, after all they are still selling Tesla K40 and K80.

This would lead on to the consumer product, so I do think it is feasible to see it in Q4 as a Titan, but that also involves a business decision and pricing for including the Samsung HBM2 (which should not be a logistics issue as they have been in mass production a little while now).
Cheers
 
Maybe Nvidia go Intel route with 6950x and have a GP102 Titan lite (at $999) with a GP100 Titan Ultra following up after some months (at $1499).
 
Hopefully it's a 600mm2 chip for maximum gaming performance, twice the GP104 and not just 400mm2 odd chip with like 30-50% better specs than gtx1080.
 
It's most likely for those VR-backpacks some companies have been showing

Actually it was just shown as a prototype, not sure it even come one day in the market. Colorful ( who is mostly unknown outside China ), seems want to show as many things they can for try enter US, EU market this year.
They have really push many many things in Computex this year. Motherboards, liquid cooling, gpu's ..
Just have to seen the announce since may, they was the first to announce their 1080 FE, they was too the first to show their custom variant.

http://en.colorful.cn/
 
Last edited:
Well I do think Nvidia would want to accelerate the next down from GP100 to strengthen their Tesla and maybe Quadro range (think priority would be Tesla). So sort of makes sense for the GP102 to go into production as soon as achievable for the business side, after all they are still selling Tesla K40 and K80.
Nvidia still hasn't released P100 as a plug-in PCIE board yet, but they will. (The white paper mentions how the existing P100 module can be put onto a carrier board). P100's performance eclipses K40 and even K80 in every metric, so it's not like NVidia is worried about their replacement. The P100 whitepaper also mentions graphic support, so it will be a high end Quadro as well. (GM210 was a Tesla-only chip with no Quadro version, so it wasn't originally clear if P100 had the same limitation. It doesn't.)

A GP104 Tesla and Quadro will be released (just using logic here, not insider knowlege). That gives more compute power at less wattage than the Maxwell M6000 and M40, so when NVidia inevitably releases both P100 and GP104 PCIeE Teslas and Quadros, they've covered the range of professional users pretty well.

That leaves GP102. Obviously it'll beat GP104 in performance and compute density, but to professional users its unique appeal over GP104 may be memory capacity. GP 104 and P100 are both limited to 16GB which may be too small for some professional workloads. NVLink memory communication mimimizes this limitation on DGX-1 systems, but it'd have more impact with PCIE board versions. GP102 likely uses a 384 bit bus and would support 24GB of memory, matching the M6000 and M40, leaving the professional Maxwell parts with no remaining advantages (similar to GTX 980 and 980Ti's eclipse).

So there's no immediate SKU hole without GP102, and there's not a desperate market pressure to accelerate its release. My unsupported best estimate is a November 2016 demonstration of a GP102 Tesla at Supercomputer 2016.
 
Nvidia still hasn't released P100 as a plug-in PCIE board yet, but they will. (The white paper mentions how the existing P100 module can be put onto a carrier board). P100's performance eclipses K40 and even K80 in every metric, so it's not like NVidia is worried about their replacement. The P100 whitepaper also mentions graphic support, so it will be a high end Quadro as well. (GM210 was a Tesla-only chip with no Quadro version, so it wasn't originally clear if P100 had the same limitation. It doesn't.)

A GP104 Tesla and Quadro will be released (just using logic here, not insider knowlege). That gives more compute power at less wattage than the Maxwell M6000 and M40, so when NVidia inevitably releases both P100 and GP104 PCIeE Teslas and Quadros, they've covered the range of professional users pretty well.

That leaves GP102. Obviously it'll beat GP104 in performance and compute density, but to professional users its unique appeal over GP104 may be memory capacity. GP 104 and P100 are both limited to 16GB which may be too small for some professional workloads. NVLink memory communication mimimizes this limitation on DGX-1 systems, but it'd have more impact with PCIE board versions. GP102 likely uses a 384 bit bus and would support 24GB of memory, matching the M6000 and M40, leaving the professional Maxwell parts with no remaining advantages (similar to GTX 980 and 980Ti's eclipse).

So there's no immediate SKU hole without GP102, and there's not a desperate market pressure to accelerate its release. My unsupported best estimate is a November 2016 demonstration of a GP102 Tesla at Supercomputer 2016.

Yes but not everyone who has a k40 or k80 is going to spend a fortune on the P100.
And those are two very old models now.
Nvidia really needs another mix-precision Tesla model that is more competitive on price, and one where its architecture spreads across at least two of the business sectors of Tesla and Quadro, that eventually also releases as Titan on Consumer.

It sounds like you are specifying the GP102 as a M6000 and M40, so they would end up replacing those and leaving K40/K80 in production in both high performance Tesla servers and workstations?
TBH I think they need to do what happened with the GK110, and have a die that crossed between Tesla/Quadro/Titan, albeit this time it is not the top Tesla die.
This would put another card into Tesla but with less DP but at a cheaper price, enable the M6000 and M40 to drop down in price or replace them, and provides the Titan.

As this is a new die, memory can be whatever they design it for; whether HBM2 or GDDR5X (but this really needs to be at least 12Gbps -120 product IMO).
Yeah I see what you mean about minimum memory for Quadro workstations, but they really need another product in the Tesla range as well as K40/K80 are too long in the tooth and not everyone is justifying the P100 upgrade costs for their research centre/business.
There is also the consideration of Intel also looking to attack Tesla business.
Problem is how to fit a die into all 3 branches of NVidia families that makes sense with regards to what they replace or position above.

Cheers

Edit:
Just to say yeah they could go the route you mention.
Problem for Nvidia is how they are going to fit next products into each of the performance families (Tesla/Quadro/Titan) without conflicting in terms of performance overlap, and they have a need to update several cards in each.
I guess it comes down to priorites, they also maybe giving themselves a headache by going with the premium large die straight off.
 
Last edited:
This well reasoned article from April surveys the existing and future Tesla lineup better than my simple speculations. In particular, it shows NVidia doesn't like to EOL Tesla parts very rapidly, with M6000 amd M40 continuing for at least another year from now. We can still have fun throwing around GP102 and PCIE P100 speculation though.
Yeah, Kepler has been around in Tesla since 2012, and with a years notification for EOL that will mean 5-6 years in production.
But then as they say Maxwell is not the right solution to replace Kepler.

Funny enough they outline what we both say for our POVs :)
Will be interesting to see what Nvidia does.
And Nvidia is not giving anything away in terms of strategy, although Marc Hamilton does mention M40 and K80 together; so maybe the focus is one Pascal GP10x for both *shrug*.
Yeah that is reaching I agree :)

The street price of the K40 and K80 are interesting and show there is a massive gap in the market; $3.2k and $4.5k these days, compared to the P100 that is estimated to be around $12.5k
Shame we did not get any hints about Quadro in the article, but different sector and top card aligns a lot with the M40 anyway.
Cheers
 
Seems reasonable to me. The lower clocks on the reported 1070m balance out the higher SP count so the performance of the two SKUs are within a couple percent of each other in FLOPs. As a guess, the mobile part's different frequency/SP balance is chosen to minimize wattage, which isn't as important for desktop.

What's much more surprising to me is that the 1070m is using same maxed out 8000 MT/sec GDDR5 frequency as the desktop part. Mobile GDDR5 tends to be clocked 20%-30% slower than desktop. (ie 5000 versus 7010 MT/sec for GM204).
 
Agreed, and at least judging from the spec sheets, 8 GT/s+ GDDR5 needs additional voltage compared to the 7-GT/s-bin for example. It would be interesting to see, if the lower power per GT/s of GDDR5X is set off by higher power requirements within the GPUs memory controllers and L2-subsystem.
 
If that thing has the same 83°C limit, it's not going to be able to hold these clocks up for long. There is just no way the cooling system in any affordable laptop can achieve the thermal goal at that power consumption. 30-60 seconds until the radiators are fully saturated, and past that point significantly reduced clocks...

Besides, why would anyone think that this is the 1070m when it has more cores than the desktop model? That make it more likely to be labeled as the 1080m. Especially considering that we probably won't see a fully uncut GP104 in any laptop part soon, as the exemplars with lowest leakage are much better monetarized to OC enthusiasts with desktop systems.
 
What is the power consumption of this part? AMD already showed with the R9 Nano what a difference careful selection and a lowered voltage can make.
 
So...the 1070m could be better than the 1070? lol what a time to be alive...

Also why the fck the laptop is so ugly? looks like a early 2000 laptop...wow.
Or the more likely case, the rumours about "full blown Pascals in laptops" were false and it's 1080m
 
Or the more likely case, the rumours about "full blown Pascals in laptops" were false and it's 1080m

Possible thatt the "full "pascal have even come from them... its a Clevo, and Clevo offtly provide laptop with desktop parts at an absolutely not affordable price. ( 6-7000$ )
 
Back
Top