GP102 confirmed?

A whole fleet of Pascal chips for every price point. It looks like the Pascal family will be broader than the Maxwell family was. Nice.
 
Last edited:
I don't think we'll see HBM2 until 1080Ti/Titan V show up. No idea what the hell those chips will be called.
 
Big speculation but I think we will see like back in Kepler where the Titan/Quadro cards will have more FP64 CUDA Cores due to GP102 shared with Tesla-Quadro-Titan and possibly with some disabled on consumer Titan.
This means size of the die will have to take that into consideration.
Note this only applied to the Titan and I think the x80ti model in consumer range.

The trend was broken with Maxwell as that was focused on FP32.
Cheers
 
defaultluser said:
If they wanted beefcake FP64, they would just can GP102 and ship P100.

I think this is about saving Nvidia money, since these new process nodes cost more than the old ones. I'm sure Nvidia wasn't happy shipping the GTX 780 Ti with all those unused FP64 units for peanuts, but it wasn't as expensive as the GP100 is.

They can probably widen the bus and increase units by 50%, and still get a die a hair over 500. That's significantly smaller than P100.

Also, a thought I just had: is P100 really just a dedicated accelerator card? I mean, does it completely lack texture units and ROPs? That would certainly explain the GP102 chip more clearly!
Click to expand...

The P100 is freaking expensive when also considering yields and demands in the higher margin HPC/supercomputer sector so I doubt that will ever make it even to Quadro.

No way we will see that, also that would be a lot of FP64 CUDA cores disabled for Quadro/consumer market and another waste.

The GP102 can still have FP64 cores, it does not have to be the same amount as the P100.
Look back at Kepler, you will see that the Tesla-Quadro-Titan comparable models all had moderate DP64.
Look in that table linked in OP and you see there is a Tesla-Quadro-Titan model with GP102.

Cheers
 
defaultluser said:
But the P100 is specifically built for supercomputers, the main users of FP64. For everyone else there's enough accuracy in FP32, or they'll be willing to drop the cash for a Tesla or two. Why build a smaller chip with similar overloaded FP64? There's no market for one: you either need the FP64 throughput of P100, or you don't, and can get by with a Xeon for less.

FP64 units require much more die space, and don't perform any faster in games, so Nvidia has been phasing that out of their mainstream chips since Kepler. They ran out of die space with Maxwell, so they had to skip the dedicated FP64 monster...but now they have the die space, the FP64 beast is back again!

But now that P100 is back to the 2:1 ratio of Fermi, they don't want to ship this massive chip to everyone as a $600 part!
Click to expand...
You are forgetting the GP102 is also a Tesla-Quadro die.
And going back to Kepler which was the last generation to have reasonable DP/SP the same die used for Tesla/Quadro/Titan all had the FP64, quite a few as well.
Cheers
 
defaultluser said:
Yeah, you're forgetting that GM206 AND GM204 was also both damn Tesla dies! It didn't magically sprout better than 1/32 rate FP64. The GTX 960 rebranded M4 still gets less than 100 GFLOPS FP64!

NVIDIA Announces Tesla M40 & M4 Server Cards - Data Center Machine Learning

The point is: you don't need to give cards any special FP64 treatment to make them into Tesla parts, you just need to add ECC support to the die, BECAUSE PEOPLE CAN STILL MAKE USE OF FP32 FOR MOST THINGS!

https://devtalk.nvidia.com/default/...-bit-floating-point-ops/post/4476877/#4476877

In the PAST when Nvidia was still tryng to get customers interested in using GPU compute for accelerating Supercomputers, they gave away FAST FP64 for "free." The GTX 260 could do 1/8 rate. The second-generation GTX 460 could do 1/12. Then all Kepler parts except GK110 dropped to 1/24.

So after two generations of supporting FAST FP64 natively in mass-market carsds, Nvidia was satisfied with the uptake, and stopped supporting beefy floating point on anything but the big chip. And I imagine they want to take things to the next level and make P100 even more compute focused, which would explain why they introduced a new high-end GP102 out of fucking nowhere?

Seriously, Nvidia hasn't had a *2-series numbered chip since G92, and that's just because they didn't want to give the die-shrunk G80 card the distinction of a G90 part number :D
Click to expand...
Why do you keep going back to only Maxwell :)
Nvidia still currently sell the Kepler models for FP64 and not just the P100, showing that Maxwell was only a SP focused product and both a stop gap until Pascal.
I think we are going to disagree on this.
Anyway my context is only the same die used in models Titan-Quadro-Tesla, not the lower ones.
Cheers
 
I am not Tesla-segment savvy, maybe not all Tesla users need big fp64. No idea though ...
 
defaultluser said:
So I guess you just want to keep ignoring the fact that Nvidia already used chips outside the x00 series for Tesla units. This includes the GK104 based Tesla k10, which AGAIN had terrible double precision floating point.

You're not making a very compelling argument here for high double precision performance just because it's marketed as Tesla. I've already showed you examples from two generations in a row!

Give me a leak you can point to about details like double precision, or to me it's just another gaming part with a Tesla re-brand.
Click to expand...
I am not talking about high DP, I am talking about it takes up space on the die,disabled,removed at ASIC level :)
The Titan and same die Quadro/Tesla all share same design, including number of FP64 CUDA cores, but they may be disabled below Tesla and could be interesting what they do with Pascal Titan and 1080ti from a disabled or removed at an ASIC level.
Going back to Kepler that is still sold.
The Tesla K20/not sure what equal Quadro is probably K6000/Titan are all 1/3 FP32, now that is a pretty reasonable ratio considering Tesla P100 is 1/2 FP32.
The 780ti is 1/24 FP32.

Anyway we are digressing from my original point :)
Cheers
 
You do agree the Tesla GP102 will HAVE to include DP64?
Ergo.
Is Nvidia going to spend millions developing different dies for Tesla/Quadro/Titan with regards to Gp102?
Ask yourself why the Kepler consumer Titan had same number of FP64 CUDA cores as the primary Tesla card at the time.
Anyway I repeat the FP100 is a massive cost and die with horrendous yields, it is not going to be used for anything but core clients and the most extreme HPC/supercomputers/DL research for a long time.

Maxwell could ONLY do FP32 (its DP was pointless) :)
What are those clients going to do, install two very different cards for when they need FP64 calculation...

It has better FP32 than Kepler, obviously and why it is being sold, but Pascal can combine both just like Kepler K20-40/k6000/Titan.
Pascal is presented as a flexible multi-purpose design.

Anyway just to emphasise I am talking just about GP102 and not any other die, as it is more likely this is the top die to be used across Tesla-Quadro-consumer Pascal Titan that may have same trend as seen with Kepler shared top die.
Cheers
 
Last edited:
defaultluser said:
Since I'm talking to a wall, I'm out.
Click to expand...
Having trouble finding your posts, especially a specific one.
Because I was going to quote one of yours just to show we are coming at this from different perspective; I am specifically talking about the top tier flagship consumer and flagship Quadro that share the same die as the top or near top Tesla, and these usually have 1/3 FP32 and implications this may have with the Pascal Titan and also 1080ti (the ti model is reduced FP64 but depends if only disabled or actually removed at an ASIC level) and whether they will follow past trends with Kepler or not due to die space - notice I say if and whether.
You posted a case where you point out there is a broader range of Tesla beyond just these such as you say GK104,K10,etc;

I agree but that is why I have emphasised many many times my context and focus is specifically only the top shared die model between consumer (will be Titan)/Quadro usually a high x000/Tesla usually the top die but that is reserved for GP100 so now Tesla using GP102.
The dies below these as you show are not comparable, but outside my focus.
Thanks
 
I am excited and patiently waiting for this chip to arrive! Once this chip officially drops, then I will move on to a lesser graphics card at a cheaper price!
 
You must log in or register to reply here.
Back
Top