NVIDIA Pascal Thread

[kithylin]

HBM2 doesn't really make sense vs 384bit GDDR5X at 10-12Ghz.

HBM2 got an ECC advantage, that's great for Tesla. But in terms of gaming cards the benefit pretty much stops there and all the cons pops out.

So its natural for a GP102 Geforce chip to use GDDR5X.

I've been thinking on this lately. I think nvidia only gives us performance of about +38% to +40% over previous generation each time as an upgrade.

Like.. say for instance they do have HBM2 working with pascal. And say.. theoretically, it's +70% faster than GTX 980. Or +80% faster. From a business model stand point.. they would be shooting themselves in the foot if they released something that far ahead of a previous model when they could release something only +40% faster now.. let everyone gobble it up, and then the following year release the hbm model that's like +80% faster than GTX 980.

They're not going to give us the faster model now, they're only going to give us one so fast this time around and then something faster later.

Besides.. how damning would that be in performance to release two products back to back.. one +40% faster than 980, then another one the next year yet another +40% faster than 1080's.

 

[ShintaiDK]

What would HBM change?

I think you ask for the technology for the sake of it. Its like the native Phenom quad core for example. Technologically pure, but made no difference.

The same can be said about Fiji cards with the exception of the Nano.

If you wanted the technological best without reason, why not HMC? Its even faster.

 

[Arachnotronic]

Yes but they haven't always been fully enabled, or uncut dies, that have been used for the Titan or the x80Ti IIRC.

Not true, AMD is releasing at least one of the Vega cards with HBM2. It's just that Nvidia is having trouble with HBM2 atm & is the main reason why GP100, with HBM2, is not coming to the desktop.

As I said previously, it's just Nvidia not being able to handle HBM2 as AMD can, probably with the experience they've had since Fury. Notice how users demeaned the Fury launch for being as late as it was, & overpriced, now we'll see if the same standards apply to the big Pascal launch.

You have pretty much zero proof of your assertions. The problem with HBM2 is that it's extremely, extremely expensive.

 

[Ajay]

Well, if info is really true on the GP102 GPU, it sure makes it a really attractive long term upgrade. Or, just a nice bump in performance for AdamK

Any word on likely release dates? Maybe in time for Black Friday/Christmas sales?

 

[ShintaiDK]

I think you have to look into 2017 for it

 

[Qwertilot]

Maybe a small chance at seeing the titan before then? LIkely 1080ti == annual card refresh for 2017 though. Suppose that competition might just move the time scales forwards a bit.

HBM2 helps with the power draw and bandwidth at very high resolutions I think? Imagine that NV will use it in main stream chips once its properly mature.

 

[ShintaiDK]

GP102 may use 384bit GDDR5X, for a good reason.

 

[Ajay]

I think you have to look into 2017 for it

Wow, if that leak is vaguely correct, then it's a real scoop being this early!

 

[PeterAStoll]

Has anyone here spotted GTX 1080 driver availability yet? Also 1080-capable versions of any of the usual suspect monitoring/control/overclocking tools?

TechPowerUp asserts that version 0.8.8 of GPU-Z includes GTX 1080 support. I've not spotted others yet.

 

[R0H1T]

You have pretty much zero proof of your assertions. The problem with HBM2 is that it's extremely, extremely expensive.

What assertions are you talking about, that AMD will bring Vega with HBM2 or GP100 ships with HBM2 as default mem type? Unless you're saying none of these are true, if so then you're the one talking smack.

As for HBM2 being uber expensive, AMD releasing Vega for gamers with HBM2 disproves you're hypothesis. Nvidia can afford to release 10K+ GPU for HPC, with HBM2, yet they can't afford to bring some variant for consumers, especially when AMD can & most likely will?

Talk about arguing just for the sake of it!


Threadcrapping and trolling are not allowed. This is an Nvidia subforum, you are not allowed to discuss AMD
Markfw900

 

[kithylin]

Have any of you ever thought for just one moment that perhaps the reason we don't have HBM2 yet with nvidia.. is maybe the price of switching to HBM2 for consumer cards doesn't yield "enough" of a performance difference over GDDR5X to even be worth it, and would end up driving costs up higher than they already are?

Remember.. TESLA cards are working with very large data sets designed with exceptional parallelism. Unlike consumer games that actually (most of the time) are not very parallel in comparison. Part of why Multi-GPU doesn't scale perfectly to +100%, or even at all in some titles.

More than likely I would guess that HBM2 just isn't beneficial enough to warrant using it for consumer gaming yet.

YES AMD is doing, and Yes it does show a benefit.. to AMD's chips. The GPU core architecture for nvidia is -VASTLY- different, and what is good for them, may not be very good for nvidia. At least, not currently.

I just infracted someone for talking about AMD, knock it off !!
Markfw900

 

[zentan]

Who has yet brought or about to bring soon close to or greater than 600mm^2 die GPU with HBM2 to gaming class GPUs? None yet I guess
GP100 for consumer doesn't make sense at least at the moment. When they can have a fp64 cut down unit for much smaller die why go for the bigger die trouble.
Not to forget Nvidia with TSMC was able to get HBM2 with a huge interposer on for GP100 Tesla. HBM2 at the moment haven't hit the bandwidth it was touted to achieve which can change by or before the end of the year. 10-12Gbps GDDR5X 384 bit would be a huge improvement from the used 10Gbps 256Bit used in gtx1080 and meanwhile it would be cost significantly lower. Let's not forget HBM1 did help Fiji chips with lower power usage and higher bandwidth but that did not prevent them being beaten by a 980TI during launch. All of this is still speculation only.

 

[zlejedi]

YES AMD is doing, and Yes it does show a benefit.. to AMD's chips. The GPU core architecture for nvidia is -VASTLY- different, and what is good for them, may not be very good for nvidia. At least, not currently.

AMD used HBM because that was only way to even get GCN with 4k shaders under 300W limit


I just infracted someone for talking about AMD, knock it off !!
Markfw900

 

[3DVagabond]

AMD used HBM because that was only way to even get GCN with 4k shaders under 300W limit

LOL Yeah, that's why they did it.


I just infracted someone for talking about AMD, knock it off !!
Markfw900

 

[kithylin]

LOL Yeah, that's why they did it.

Well AMD -HAS- had a rather long history of... pretty much all of their products, cpu and gpu being quite high on the power usage side just to even remotely be some what competitive.

Not conjecture... this is a fact. AMD products are always significantly higher power usage in all regards. And run extremely hot.... so far, as of yet.. neither nvidia nor intel have released a single product that flat out -requires- water cooling for any reason.

 

[Cookie Monster]

LOL Yeah, that's why they did it.

He does have a point. The power savings using HBM compared to GDDR5 is rather significant. Its one of the key features for using HBM.

Opens up the TDP limit.

 

[ShintaiDK]

Have any of you ever thought for just one moment that perhaps the reason we don't have HBM2 yet with nvidia.. is maybe the price of switching to HBM2 for consumer cards doesn't yield "enough" of a performance difference over GDDR5X to even be worth it, and would end up driving costs up higher than they already are?

Remember.. TESLA cards are working with very large data sets designed with exceptional parallelism. Unlike consumer games that actually (most of the time) are not very parallel in comparison. Part of why Multi-GPU doesn't scale perfectly to +100%, or even at all in some titles.

More than likely I would guess that HBM2 just isn't beneficial enough to warrant using it for consumer gaming yet.

YES AMD is doing, and Yes it does show a benefit.. to AMD's chips. The GPU core architecture for nvidia is -VASTLY- different, and what is good for them, may not be very good for nvidia. At least, not currently.

The key point using HBM2 on P100 is ECC. Something that a GeForce gamer wouldn't care much about.

 

[3DVagabond]

Well AMD -HAS- had a rather long history of... pretty much all of their products, cpu and gpu being quite high on the power usage side just to even remotely be some what competitive.

Not conjecture... this is a fact. AMD products are always significantly higher power usage in all regards. And run extremely hot.... so far, as of yet.. neither nvidia nor intel have released a single product that flat out -requires- water cooling for any reason.

Fury runs just fine on air. And don't say nVidia has never made a product that required water cooling for any reason. OG Titan throttled. 780 and 780 ti reference cards throttled. Titan Z throttled. Titan X throttled. 1080 throttles. Their users don't seem to mind though. nVidia users sure minded Hawaii throttling though.

He does have a point. The power savings using HBM compared to GDDR5 is rather significant. Its one of the key features for using HBM.

Opens up the TDP limit.

That is not why they developed HBM though. It's purely advancing the state of the art. The main purpose of it going forward will be for APU's. The iGPU is bottlenecked by the system RAM and the benefit of shared RAM between the GPU and CPU. Efficiency is a side benefit.


I just infracted someone for talking about AMD, knock it off !!
Markfw900

 

[ShintaiDK]

That is not why they developed HBM though. It's purely advancing the state of the art. The main purpose of it going forward will be for APU's. The iGPU is bottlenecked by the system RAM and the benefit of shared RAM between the GPU and CPU. Efficiency is a side benefit.

That's an awful product for that segment due to the cost issue alone. Anything relating to this have been a second thought at best.

 

[R0H1T]

That's an awful product for that segment due to the cost issue alone. Anything relating to this have been a second thought at best.

Say what? If Zen is anywhere close to an IB, per core performance, than an HBM APU will blow anything Intel eDRAM laden GPU out of the water even if it's on 10nm. Guess what those eDRAM parts sell for D:

 

[ShintaiDK]

Say what? If Zen is anywhere close to an IB, per core performance, than an HBM APU will blow anything Intel eDRAM laden GPU out of the water even if it's on 10nm. Guess what those eDRAM parts sell for D:

You forget the cost structure like so many others. The interposer alone cost around what the EDRAM chip cost.

HBM still haven't moved anywhere in terms of cost. That's also why its most likely a card like the GP102 will use 384bit GDDR5X. The cost/benefit of HBM is simply way too small.

 

[R0H1T]

You forget the cost structure like so many others. The interposer alone cost around what the EDRAM chip cost.

HBM still haven't moved anywhere in terms of cost. That's also why its most likely a card like the GP102 will use 384bit GDDR5X. The cost/benefit of HBM is simply way too small.

What about PoP or TSS, for bypassing interposer, isn't that something what mobile APUs would benefit from?

 

[kithylin]

Fury runs just fine on air. And don't say nVidia has never made a product that required water cooling for any reason. OG Titan throttled. 780 and 780 ti reference cards throttled. Titan Z throttled. Titan X throttled. 1080 throttles. Their users don't seem to mind though. nVidia users sure minded Hawaii throttling though.

Throttling or not, all those items you listed actually do function 100% perfectly fine at stock on air coolers.

AMD Fury X, -HAS TO BE WATER COOLED- was in, can not run on air, no matter what, no way, not possible, can't do it. That big 5 ghz amd 8 core cpu, same.. -MUST BE WATER COOLED, CAN NOT RUN ON AIR-.

It's utterly terrible from both points.. no one should be releasing a retail product that runs that insanely hot by default.. it's just terrible.

Part of why I stick with nvidia. At least my video cards will work without water.

 

[3DVagabond]

Throttling or not, all those items you listed actually do function 100% perfectly fine at stock on air coolers.

AMD Fury X, -HAS TO BE WATER COOLED- was in, can not run on air, no matter what, no way, not possible, can't do it. That big 5 ghz amd 8 core cpu, same.. -MUST BE WATER COOLED, CAN NOT RUN ON AIR-.

It's utterly terrible from both points.. no one should be releasing a retail product that runs that insanely hot by default.. it's just terrible.

Part of why I stick with nvidia. At least my video cards will work without water.

Fury is the same chip. It runs fine on air and so would Fury X. You aren't running any 5GHz Intel CPU on a stock cooler. Try being honest.

 

[R0H1T]

Fury is the same chip. It runs fine on air and so would Fury X. You aren't running any 5GHz Intel CPU on a stock cooler. Try being honest.

You forgot the Nano, full Fury just X inches smaller :sneaky: