Nvidia Pascal & Volta Compute Performance Figures Released, HBM2 Power-Hungry?

[patrickjp93]

You mean they saturated the supply side? That's no excuse for more than one year of production.

That's what happens when the HMC consortium fails to create an industry standard. Even if it wasn't JEDEC, they could still license out the tech to others, yet they chose not to. If NVidia was truly a powerful HMC consortium member, they should have been able to force a licensing deal with other vendors like Samsung.

 

Either way HBM has already doubled it's speed in one year. Who knows what will happen in 2017.

 

As for your fixation on HMC's lower latency, this is interesting (and goes well with what we already see)

 

As for Arctic Island, yeah it still seems to be up in the air.

They went for JEDEC and Hynix played dirty with AMD's R&D money. And the HMCC does license the tech out. Micron and Intel are only 2 of its members. Samsung is part of it too, and so is Nvidia.

 

No one expected to be blindsided by Hynix buying votes and delaying possible acceptance by 2 years (standard procedure when you fail a JEDEC vote). It had nothing to do with how powerful a member Nvidia was.

[Master Disaster]

So basically if they continue pushing HBM then we're all gonna need a 24 pin MoBo power connector in our graphics cards too.

RIP Small Form Factor builds.

[don_svetlio]

If I can get 2/3, give me Lawlz and Don. If I can only have 1, give me Don. At least Lawlz isn't an outright troll, and Trik is just paranoid and brings up 3.5GB in every damn Nvidia thread since the news broke.

I love you too honey. Did you at least do your math homework this time?

[don_svetlio]

So basically if they continue pushing HBM then we're all gonna need a 24 pin MoBo power connector in our graphics cards too.

RIP Small Form Factor builds.

that might result in better OCing though - and maybe GPUs will finally influence power bills so people aren't making that up!

[BuckGup]

So is overclocking the Fury X memory pointless? Yes it is

[AlwaysFSX]

If this is true, I can't wait for nvidia fanboys to defend the 600% increase in power consumption, for a 50% increase in performance.

 

Also if this is true, I'd have a hard time believing nvidia would openly accept hbm 2, especially since they're so big on overclocking. Would have thought they'd go for the much more efficient hbm 1, and tried to work out sourcing it en mass.

I don't think anyone will. It's pretty bad how much power the memory can draw now. They could lower the speed so it doesn't draw so much power for almost the same total speed but I'm not sure if Nvidia will do that. Time will tell what happens when cards are finally in hands so we can get concrete information. This actually makes me more concerned for AMD because more shit will fly if their new cards end up drawing even more power because of using HBM2 and it either reduces power draw for the core or increases total card draw. Both of which are a lose-lose situation.

[don_svetlio]

I don't think anyone will. It's pretty bad how much power the memory can draw now. They could lower the speed so it doesn't draw so much power for almost the same total speed but I'm not sure if Nvidia will do that. Time will tell what happens when cards are finally in hands so we can get concrete information. This actually makes me more concerned for AMD because more shit will fly if their new cards end up drawing even more power because of using HBM2 and it either reduces power draw for the core or increases total card draw. Both of which are a lose-lose situation.

We'll end up in the same situation we were when it was Kepler vs GCN 1.0/1.1 - even though both were identical people will still claim Nvidia draws much less and produces less heat since they'll have a 30-50W lower TDP

[VagabondWraith]

So is overclocking the Fury X memory pointless? Yes it is

The bandwidth is already amazing. i think HBM2 overclocking would be pointless with 1 TB/s. It will probably be locked on next gen cards like AMD has with Fury to keep power consumption down.

[VagabondWraith]

We'll end up in the same situation we were when it was Kepler vs GCN 1.0/1.1 - even though both were identical people will still claim Nvidia draws much less and produces less heat since they'll have a 30-50W lower TDP

I think both companies will be relatively the same since Maxwell is gimped on hardware.

[AlwaysFSX]

We'll end up in the same situation we were when it was Kepler vs GCN 1.0/1.1 - even though both were identical people will still claim Nvidia draws much less and produces less heat since they'll have a 30-50W lower TDP

Which we all know is BS, especially once you get to overclocking a card. Look at Maxwell if you want, those will guzzle electricity once OC'd. HBM1 has already shown that increasing memory speed doesn't effect performance all that much so I still can't see a reason to run HBM2 so fast that it requires exponential amounts of power. Maybe we'll get two generations of the memory where the first batch of HBM2 cards run at lower spec until the power draw can be solved? It's not like you'd be missing out on any serious performance even with supposedly 8 TFlops of SP compute power.

 

Oh technology. One step forward, one step back.

 

[don_svetlio]

Which we all know is BS, especially once you get to overclocking a card. Look at Maxwell if you want, those will guzzle electricity once OC'd. HBM1 has already shown that increasing memory speed doesn't effect performance all that much so I still can't see a reason to run HBM2 so fast that it requires exponential amounts of power. Maybe we'll get two generations of the memory where the first batch of HBM2 cards run at lower spec until the power draw can be solved? It's not like you'd be missing out on any serious performance even with supposedly 8 TFlops of SP compute power.

 

Oh technology. One step forward, one step back.

 

 

I think both companies will be relatively the same since Maxwell is gimped on hardware.

Nvidia will use the same trick though - run the tests for TDP at 1000MHz core and let GPU Boost take it to 1300-1400MHz so while on paper the card pulls 200W in reality it pulls 260W

[AlwaysFSX]

Nvidia will use the same trick though - run the tests for TDP at 1000MHz core and let GPU Boost take it to 1300-1400MHz so while on paper the card pulls 200W in reality it pulls 260W

Oh we already know that will happen, but it's on both sides anyways because "oh well you bought the card to run at these speeds, not past them."

 

I'm fine with the core drawing more power because you get more gain from it. The memory however is a bit more gray in that regard. If a card draws 260W that's fine by me, if it jumps to 320W because of the memory then I'll flip shit.

[don_svetlio]

Oh we already know that will happen, but it's on both sides anyways because "oh well you bought the card to run at these speeds, not past them."

 

I'm fine with the core drawing more power because you get more gain from it. The memory however is a bit more gray in that regard. If a card draws 260W that's fine by me, if it jumps to 320W because of the memory then I'll flip shit.

Thermi Mk.2?

(╯°□°）╯︵ ┻━┻

[VagabondWraith]

Nvidia will use the same trick though - run the tests for TDP at 1000MHz core and let GPU Boost take it to 1300-1400MHz so while on paper the card pulls 200W in reality it pulls 260W

I'm not so sure we'll see overclocking again like on Maxwell. I think it will be like Hawaii/Kepler like. Maxwell could OC like no tomorrow but that's because they cut out a bit of hardware to do that.

Id rather see Pascal with Async compute (and everything else cut out of Maxwell) over the ability to get crazy overclocks. If Async compute is not in Pascal, I may switch to AMD for dat pure firepower.

[don_svetlio]

I'm not so sure we'll see overclocking again like on Maxwell. I think it will be like Hawaii/Kepler like. Maxwell could OC like no tomorrow but that's because they cut out a bit of hardware to do that.

Id rather see Pascal with Async compute (and everything else cut out of Maxwell) over the ability to get crazy overclocks. If Async compute is not in Pascal, I may switch to AMD for dat pure firepower.

hehe FIREpower. Neat

[patrickjp93]

I'm not so sure we'll see overclocking again like on Maxwell. I think it will be like Hawaii/Kepler like. Maxwell could OC like no tomorrow but that's because they cut out a bit of hardware to do that.

Id rather see Pascal with Async compute (and everything else cut out of Maxwell) over the ability to get crazy overclocks. If Async compute is not in Pascal, I may switch to AMD for dat pure firepower.

Eh, if Nvidia moves to doing 2 different arches (1 for compute, one for gaming), then we may get Maxwell-like overclocking well into the future.

[Notional]

They went for JEDEC and Hynix played dirty with AMD's R&D money. And the HMCC does license the tech out. Micron and Intel are only 2 of its members. Samsung is part of it too, and so is Nvidia.

 

No one expected to be blindsided by Hynix buying votes and delaying possible acceptance by 2 years (standard procedure when you fail a JEDEC vote). It had nothing to do with how powerful a member Nvidia was.

 

You keep claiming that, providing no evidence what so ever. I doubt AMD would spend money on "Buying votes". It does not seem to fit their modus operandi, nor their economy. And since AMD doesn't seem to make any royalties on HBM, there really is no incentive for them to do such a thing. It simply makes no sense.

 

Afaik one of the biggest NAND manufacturers in the world (Samsung) doesn't even make HMC, even though they are a member? Again, makes little sense.

 

If there is money to be made in HMC, and licenses, as well as a completely supply starved market, most memory vendors, should be all over it. Yet they aren't. Why are you so fixated on HMC?

 

JEDEC or not, if the tech, the spec and the market (demand) is there, such a thing as JEDEC has no meaning.

[patrickjp93]

You keep claiming that, providing no evidence what so ever. I doubt AMD would spend money on "Buying votes". It does not seem to fit their modus operandi, nor their economy. And since AMD doesn't seem to make any royalties on HBM, there really is no incentive for them to do such a thing. It simply makes no sense.

 

Afaik one of the biggest NAND manufacturers in the world (Samsung) doesn't even make HMC, even though they are a member? Again, makes little sense.

 

If there is money to be made in HMC, and licenses, as well as a completely supply starved market, most memory vendors, should be all over it. Yet they aren't. Why are you so fixated on HMC?

 

JEDEC or not, if the tech, the spec and the market (demand) is there, such a thing as JEDEC has no meaning.

It makes perfect sense if what you want is good press and to look legitimate.

 

Samsung's in the middle of its Wide-IO push for mobile devices. I'm not surprised it hasn't jumped into the HMC fray yet. Samsung's perfectly happy to let Micron remain alive by having Oracle's business for another couple years.

 

JEDEC has a lot of meaning. Not being under JEDEC means being a less secure (for future support) product. That can shake demand clear to its foundations. HMC is vastly superior to DDR4 in every single way, and yet...

[zMeul]

If this is true, I can't wait for nvidia fanboys to defend the 600% increase in power consumption, for a 50% increase in performance.

 

Also if this is true, I'd have a hard time believing nvidia would openly accept hbm 2, especially since they're so big on overclocking. Would have thought they'd go for the much more efficient hbm 1, and tried to work out sourcing it en mass.

you do realize that AMD will be hit with this exact issue too?!  <_<

[ivan134]

you do realize that AMD will be hit with this exact issue too?!  <_<

We all know that. The reason he said it is that so many people have been parroting this myth that AMD's cards are not as efficient as Nvidia's simply because people think TDP is power consumption and Nvidia puts a lower TDP on their spec sheet than AMD does. The difference between Kepler and Hawaii was ~20W on load and Maxwell's efficiency came from gutting it's compute capabilities. Yet, we still see people, even reviewers, saying that AMD's power consumption is not competitive.

[patrickjp93]

We all know that. The reason he said it is that so many people have been parroting this myth that AMD's cards are not as efficient as Nvidia's simply because people think TDP is power consumption and Nvidia puts a lower TDP on their spec sheet than AMD does. The difference between Kepler and Hawaii was ~20W on load and Maxwell's efficiency came from gutting it's compute capabilities. Yet, we still see people, even reviewers, saying that AMD's power consumption is not competitive.

We also see the HPC world saying that, and frankly I put more stock in their actions than anyone's words on the matter here.

[AlTech]

http://wccftech.com/nvidia-pascal-volta-gpus-sc15/

 

As we all know, Pascal will be released sometime 2016. I'm personally betting late Q2 into Q3 based on Samsung's HBM 2 ramp up. Wccftech summarized Nvidia's SC'15 presentation and pulled out the pertinent details. http://images.nvidia.com/events/sc15/SC5125-energy-efficient-architectures-exascale-systems.html

 

Pascal is to top out in double precision performance of 4 Teraflops, a solid improvement, but not exactly ground-breaking, compared to AMD's S9170 FirePro rated at 2.62TFlops DP. It's not yet clear if Nvidia went with 1/2 performance or kept in step with Kepler's model of 1/3 performance. I'm betting on the latter due to the second announcement.

 

Volta will top out at 7 TFlops DP, on the same 16nmFF+ node. If Pascal was 1/3 performance, that would put SP at 12TFlops. If then Volta moved to 1/2, the corresponding SP performance will be 14TFlops, a much more believable performance improvement between generations on the same node than Pascal (at 1/2 DP performance) going from 8 TFlops SP to Volta's 14, and it's certainly more believable than Volta being 1/3 performance and having a whopping 21TFlops SP in a single-die solution.

 

14/12 = 1.1666... (16.7% improvement)

14/8 = 1.75 (75% improvement)

21/12 = 1.75 (75% improvement)

21/8 = 2.625 (162.5% improvement) Pascal at 1/2, Volta at 1/3, and obviously insane

 

 

Second, Nvidia's presenter outlined that the memory power/thermal problem is far from resolved. In fact, it gets worse when you crank up HBM 2. While below 1GHz it's a very efficient memory architecture, above that point, HBM 2 actually very quickly becomes more power hungry than GDDR5X for each additional cycle per second and every additional GB/s, At 1.2TB/s (the bandwidth of Pascal and Volta for now), the memory package is 60W of the thermal envelope on its own. whether this is with 4-hi or 8-hi stacks was not stated, but it doesn't bode well either way.

 

For reference, a 1TB/s arrangement of GDDR5x is 16 64-bit wide chips at 2GHz, and has a thermal requirement of 70W (Micron). Per clock HBM2 is actually worse if Nvidia's model is of the 4hi stacks (4*4 = 16 chips at half the clock speed).

 

All in all, the plot thickens in the GPU wars. Now to eagerly await the Arctic Islands announcements and unveiling. It looks like We'll have to find HBM 2's replacement much sooner than we'd hoped.

 

*laughs to self since HMC has lower latency and comparable bandwidth to HBM 1 and the 8 chips on KNL only need 18W of thermal dissipation power*

I will only care if it makes it to the GTX 1070 or 1080. 

 

And besides who cares about SP and DP. AMD has better SP and DP and yet they have worse gaming performance.

[patrickjp93]

I will only care if it makes it to the GTX 1070 or 1080. 

 

And besides who cares about SP and DP. AMD has better SP and DP and yet they have worse gaming performance.

1) FINALLY I CAN MAKE MY 10,000TH POST!!!

2) Don't quote the entire post

3) You care about SP because, on a relative scale, SP will tell you how much the cards have improved over the previous generation, and that improvement is generally in line with improvements in game performance if CPU bottlenecks are avoided. You care about DP only in this case because the SP numbers weren't given, but we can make a solid guess at what they are based on what Nvidia did with Kepler.

 

Yay 10,000 posts!

[AlTech]

1) FINALLY I CAN MAKE MY 10,000TH POST!!!

2) Don't quote the entire post

3) You care about SP because, on a relative scale, SP will tell you how much the cards have improved over the previous generation, and that improvement is generally in line with improvements in game performance if CPU bottlenecks are avoided. You care about DP only in this case because the SP numbers weren't given, but we can make a solid guess at what they are based on what Nvidia did with Kepler.

 

Yay 10,000 posts!

2015-11-25.png

LOl still stuck under 1500 posts. 

[Ekpyrosis]

They went for JEDEC and Hynix played dirty with AMD's R&D money. And the HMCC does license the tech out. Micron and Intel are only 2 of its members. Samsung is part of it too, and so is Nvidia.

Do you have a source for this?

Because I can't find Nvidia on this list and this article mentions explicitly that they are not a member.