• Competitor rules

    Please remember that any mention of competitors, hinting at competitors or offering to provide details of competitors will result in an account suspension. The full rules can be found under the 'Terms and Rules' link in the bottom right corner of your screen. Just don't mention competitors in any way, shape or form and you'll be OK.

SK Hynix shows off HBM1 and HBM2

Boomstick777

Boomstick777

Boomstick777

Soldato
Joined
2 Jan 2012
Posts
12,409
Location
UK.
Did you know it’s almost 7 years since the first GDDR5 graphics card came out? (Well now you know, it was Radeon HD 4870).
Click to expand...

For the first time Hynix has shown off a single 1GB HBM1 module and even a whole HBM2 wafer. HBM2 is not expected this year, as this technology is still under development. It is said that the first architecture to benefit from HBM2 is NVIDIA’s Pascal.

The HBM1 however, is already being made, and the first manufacturer to utilize the power of new memory technology is AMD. Radeon R9 390X and R9 390 are said to be the first graphics cards to feature HBM instead of GDDR5 modules.
Click to expand...

http://videocardz.com/55259/sk-hynix-shows-off-hbm1-and-hbm2

The HBM2 will double the bandwidth and density. So theoretically 4 stacks will allow 32GB of memory to be installed on interposer. This is how Pascal will achieve 30+ GB capacities as shown yesterday during roadmap reveal by the CEO of NVIDIA.
Click to expand...

J06RqAs.jpg
 
Last edited:
i was talking about this earlier
but doesnt that mean 390 could have upto 16gb if they wanted?
or is my math bad?

and how much faster is it than gddr5? 4times?
some1 clever educate me :)
 
@ Mei I imagine expense is the limiting factor rather than feasibility.

In a couple of years the fact that we might have 30GB+ uber fast mem on GPU's is insane. Graphics are about to see a massive jump, die shrinks, new memory arch. New API's / software, VR, DX12, and Windows 10. 2016 will be massive for PC space imho.
 
yeh its whats going to make 8k gaming possible (looks at andy) lol
and some of the vr stuff looking lifelike i guess
the software part of it is a big thing tho
 
Boomstick777 said:
30GB+ !

:eek:
Click to expand...

Not for the consumer segment, and the more we hear about Pascal, the more it seems unlikely that we'll see a consumer version any time soon.

Remember, Volta has been massively delayed, and Pascal was brought in to win supercomputer / render farm contracts as Volta was postponed. AMD are about to launch their HBM 390s, so NVIDIA absolutely need something before the 2018 HMC Volta, so we will see consumer Pascal ... but nothing has so far been mentioned of it - all of the talk has been about custom / enterprise versions.

I still don't think we'll see any sign of Pascal until H2 '16 (probably Q4) in the consumer segment.

Also, I doubt that the first Pascal products will be HBM2 given the quoted bandwidth figures.
 
pmc25 said:
Not for the consumer segment, and the more we hear about Pascal, the more it seems unlikely that we'll see a consumer version any time soon.

Remember, Volta has been massively delayed, and Pascal was brought in to win supercomputer / render farm contracts as Volta was postponed. AMD are about to launch their HBM 390s, so NVIDIA absolutely need something before the 2018 HMC Volta, so we will see consumer Pascal ... but nothing has so far been mentioned of it - all of the talk has been about custom / enterprise versions.

I still don't think we'll see any sign of Pascal until H2 '16 (probably Q4) in the consumer segment.

Also, I doubt that the first Pascal products will be HBM2 given the quoted bandwidth figures.
Click to expand...

What makes you think we won't see high amounts of memory and Pascal in the consumer space? Demand is only going to get bigger. Games more demanding, 4K, VR. Of course we will see ridiculous amounts of memory on GPU's. It's always the same.. People used to say a few MB's would be enough :p In a few years from now, 4GB will seem like a joke ;)

Pascal will be used in all of Nvidia's tiers, from mobile (Tablet/Car) compute space up to desktop etc. It's designed to scale in performance and compute, mixed precision will be a boon for Nvidia.
 
Last edited:
HBM1 spec is for 2Gb = 256MB dies with 4 chip stacks for 1GB total per stack.

HBM 2 spend is for 4-8Gb = 512-1GB dies with 4-8 chip stacks for anything from 2GB to 8GB per stack. To arbitrarily say it doubles density is fairly misleading. The minimum listed spec is double the density, the highest listed density(due to chips and twice the number of chips in a stack) is 8 times higher.

I've seen it suggested that rather than 2GB dies what they are going to do is stick 2x 1GB stacks together on one interposer and then connect that to another interposer which also has the gpu on it. I didn't think/hope they'd need to do that for GPU but did wonder if they'd do that for an APU in the future. THe trouble if you can call it that being that even a single 128GB/s stack would be massive bandwidth for a cpu/apu, but if you had lets say a 4GB sensible minimum for APU's in laptops and 8-16GB on desktop, one way to achieve that would be the same thing. Stick 4-8 stacks on one interposer, then AMD can buy what is effectively a interposer based dimm of 4-8GB of memory, and stick that to it's cpu. So rather than have 1TB's of connections to the cpu with a massive memory controller you can stick to a relatively small memory controller, sensible bandwidth/power usage but still have the capacity available.

Realistically to ever get full system memory on an APU we'd want 1-2 stacks of 256GB/s 8GB per stack HBM.
 
drunkenmaster said:
HBM1 spec is for 2Gb = 256MB dies with 4 chip stacks for 1GB total per stack.

HBM 2 spend is for 4-8Gb = 512-1GB dies with 4-8 chip stacks for anything from 2GB to 8GB per stack. To arbitrarily say it doubles density is fairly misleading. The minimum listed spec is double the density, the highest listed density(due to chips and twice the number of chips in a stack) is 8 times higher.

I've seen it suggested that rather than 2GB dies what they are going to do is stick 2x 1GB stacks together on one interposer and then connect that to another interposer which also has the gpu on it. I didn't think/hope they'd need to do that for GPU but did wonder if they'd do that for an APU in the future. THe trouble if you can call it that being that even a single 128GB/s stack would be massive bandwidth for a cpu/apu, but if you had lets say a 4GB sensible minimum for APU's in laptops and 8-16GB on desktop, one way to achieve that would be the same thing. Stick 4-8 stacks on one interposer, then AMD can buy what is effectively a interposer based dimm of 4-8GB of memory, and stick that to it's cpu. So rather than have 1TB's of connections to the cpu with a massive memory controller you can stick to a relatively small memory controller, sensible bandwidth/power usage but still have the capacity available.

Realistically to ever get full system memory on an APU we'd want 1-2 stacks of 256GB/s 8GB per stack HBM.
Click to expand...

Interested to how AMD utilise HBM with their APU's. With the new architecture next year Zen or whatever it's called + HBM, that's some awesome potential especially in small form factor. Laptop, Steam Box, maybe even future consoles etc.

I hope they do well in that space.
 
Boomstick777 said:
What makes you think we won't see high amounts of memory and Pascal in the consumer space? Demand is only going to get bigger. Games more demanding, 4K, VR. Of course we will see ridiculous amounts of memory on GPU's. It's always the same.. People used to say a few MB's would be enough :p In a few years from now, 4GB will seem like a joke ;)

Pascal will be used in all of Nvidia's tiers, from mobile (Tablet/Car) compute space up to desktop etc. It's designed to scale in performance and compute, mixed precision will be a boon for Nvidia.
Click to expand...

You going for that 5 monitor 8K setup that'll cost you £250k and you won't actually be able to connect to the graphics cards and feed with native 8K per screen, because an interface doesn't yet exist with the bandwidth? Also, again, you'd never be able to feed enough bandwidth to the GPU.

Yeah, and it's not going to happen soon ... Pascal was brought in as a stop gap because Volta got gigantically delayed.
 
You must log in or register to reply here.
Back
Top Bottom