AMD Exec: DDR3 Memory Not A Bottleneck, 12 GB GPUs Won’t Become The Standard Anytime Soon

[Boomstick777]

Head of global technical marketing Robert Hallock talks about the state of graphics cards.

Nvidia may still be considered the top option for graphics card enthusiasts but AMD has been showing it up recently. Aside from its hardware powering the PS4 and Xbox One, AMD also has the Radeon R9 295X2 which features a whopping 8 GB DDR5 RAM.

http://gamingbolt.com/amd-exec-ddr3...gb-gpus-wont-become-the-standard-anytime-soon

 

[eddyr]

However, PC GPUs still rely on DDR3 RAM. What kind of bottlenecks does this create for the system

Might be a simple proof reading mistake and the author meant gaming, but that in combo with the dual gpu's 'whopping' 8GB pretty much removes any faith that the author understands or cares what he is writing about. :/

 

[Boomstick777]

The questions were poorly worded, the DDR3 means for CPU and APU, not gaming GPU. Obviously currently that's GDDR5.

 

[Deceptor]

Might be a simple proof reading mistake and the author meant gaming, but that in combo with the dual gpu's 'whopping' 8GB pretty much removes any faith that the author understands or cares what he is writing about. :/

This plus the fact the titan-z has a "whopping 12GB" going by his understanding

 

[bru]

Poor article, in my opinion.
Also if the currant memory architecture on GPU's isn't a bottleneck then why are both companies heading towards stacked Dram for future GPU's.

 

[Yaayuh!]

Indeed. The guy is talking out of his posterior.

Move on, nothing to see here.

 

[Orangey]

Robert Hallock aka Thracks on this forum does in fact know what he's talking about and he was probably talking about the system as a whole. Would've been a good chance to talk up hUMA and HSA.

 

[eddyr]

Yeah the direct quotes from Thracks are fine. Just nothing new to the table (bland questions).
I'm not sure the article author would have been the best medium for talking up hUMA/HSA though.

 

[Rroff]

Poor article, in my opinion.
Also if the currant memory architecture on GPU's isn't a bottleneck then why are both companies heading towards stacked Dram for future GPU's.

VRAM bandwidth generally isn't lacking - when correctly specced :S VRAM amount is a bigger bottleneck when your using ultra high resolutions, etc.

While eventually there is going to be a need and/or advantages from having stacked RAM, etc. GDDR5 tech wise is far from tapped out for having the performance to power current and next generation GPUs.

 

[drunkenmaster]

First up the AMD guy is saying ddr4 will bring no benefits to gamers in terms of performance, the bandwidth of system memory isn't the bottleneck and he's correct to a degree. The main reason ddr4 isn't being moved to aside from the issues of timing of new cpu architectures for AMD, is that ddr4 sucks balls. There is basically within a 1-2% performance difference between fully available DDR3 and similarly priced DDR4. There is no performance advantage as yet because DDR3 scaled up way beyond where it was planned to and DDR4 took so long to get to market and hasn't scaled up very far yet that the crossing point hasn't been hit for ddr4.

AS for GDDR5 tech wise not being tapped out, it is, it's obsolete in comparison to HBM/HMC, these two technologies scale to a magnitude beyond what GDDR5 can do.

How do you get a 40% faster gpu on the same process when it's 40% bigger but uses only 10% more power... you cut out 30-40W of power from the memory usage.

Why do you think Titan had a 384bit bus and not just go 512bit, cost, size, power usage, lots of reasons. It would certainly help it's performance in compute where GPGPU is in many cases more bandwidth reliant than gaming. Why have AMD and Nvidia added memory compression hardware, this costs money, die space and R&D costs.

You can put a 1024bit memory bus on a gpu and get 600GB/s+ out of GDDR5, but you could get the same bandwidth with an even larger bus out of gddr3. It's about how feasible the idea of having a 300W gpu with 150W of power usage just in memory access leaving only 150W for the rest of the gpu. We haven't moved beyond a 512bit bus and we've had cards hit that level since at least the 2900XT. We can't get much beyond what AMD has achieved bandwidth wise with Hawaii, if at all further. It's just not sensible to use a greater portion of power on the memory than Hawaii did, 784bit bus just isn't going to happen.

HBM/HMC blow the living hell out of GDDR5, in comparison it's obsolete technology and the next generation of gpu's will struggle significantly without HBM. Something aiming to have twice the processing power of a 980gtx on 14-16nm processes and trying to run at higher resolution will absolutely find itself bandwidth limited.

HBM is going to give WAY more bandwidth AND decrease the power usage which will also mean more power available for the GPU in any given power budget. 250W gpu, 40W HBM 210W gpu instead of 80W memory and 170W gpu, have 20% more shaders in the same power budget and/or higher clocks and the HBM has the bandwidth to feed the extra shaders.

 

[Rroff]

HBM/HMC blow the living hell out of GDDR5, in comparison it's obsolete technology and the next generation of gpu's will struggle significantly without HBM.

That is as maybe but GDDR5 is still far from tapped out for current or even next generation GPUs - without having to scale it to ridiculous levels - sure certain aspects have implications on the practical side of using GDDR5 in configurations that might provide the bandwidth needed - generally seems to be more about cost cutting though lately than the actual technical constraints.

 

[Orangey]

http://dl.acm.org/citation.cfm?id=2689667

Two-level memory

>AMD paper

 

[r7slayer]

Spoiler

First up the AMD guy is saying ddr4 will bring no benefits to gamers in terms of performance, the bandwidth of system memory isn't the bottleneck and he's correct to a degree. The main reason ddr4 isn't being moved to aside from the issues of timing of new cpu architectures for AMD, is that ddr4 sucks balls. There is basically within a 1-2% performance difference between fully available DDR3 and similarly priced DDR4. There is no performance advantage as yet because DDR3 scaled up way beyond where it was planned to and DDR4 took so long to get to market and hasn't scaled up very far yet that the crossing point hasn't been hit for ddr4.

AS for GDDR5 tech wise not being tapped out, it is, it's obsolete in comparison to HBM/HMC, these two technologies scale to a magnitude beyond what GDDR5 can do.

How do you get a 40% faster gpu on the same process when it's 40% bigger but uses only 10% more power... you cut out 30-40W of power from the memory usage.

Why do you think Titan had a 384bit bus and not just go 512bit, cost, size, power usage, lots of reasons. It would certainly help it's performance in compute where GPGPU is in many cases more bandwidth reliant than gaming. Why have AMD and Nvidia added memory compression hardware, this costs money, die space and R&D costs.

You can put a 1024bit memory bus on a gpu and get 600GB/s+ out of GDDR5, but you could get the same bandwidth with an even larger bus out of gddr3. It's about how feasible the idea of having a 300W gpu with 150W of power usage just in memory access leaving only 150W for the rest of the gpu. We haven't moved beyond a 512bit bus and we've had cards hit that level since at least the 2900XT. We can't get much beyond what AMD has achieved bandwidth wise with Hawaii, if at all further. It's just not sensible to use a greater portion of power on the memory than Hawaii did, 784bit bus just isn't going to happen.

HBM/HMC blow the living hell out of GDDR5, in comparison it's obsolete technology and the next generation of gpu's will struggle significantly without HBM. Something aiming to have twice the processing power of a 980gtx on 14-16nm processes and trying to run at higher resolution will absolutely find itself bandwidth limited.

HBM is going to give WAY more bandwidth AND decrease the power usage which will also mean more power available for the GPU in any given power budget. 250W gpu, 40W HBM 210W gpu instead of 80W memory and 170W gpu, have 20% more shaders in the same power budget and/or higher clocks and the HBM has the bandwidth to feed the extra shaders.

Good read there drunken! Sometimes like to read your massive paragraphs lol. But i do sometimes think how accurate is the things you say. Not meaning that what you say isn't credible but if some of it is more your own thoughts than actual findings if you know what i mean? Will HBM truly be a big saving on power? Will it offer much greater performance or just greater bandwidth? I know more bandwidth is better for bigger resolutions but what about 1080p or possibly 1440p? Will it offer a great performance increase at these resolutions?

Also in terms of DDR4 performance in the system, doesn't DDR4 use higher timings over DDR3 to get a higher bandwidth? I always thought bigger timings was bad for gaming?

 

[drunkenmaster]

Good read there drunken! Sometimes like to read your massive paragraphs lol. But i do sometimes think how accurate is the things you say. Not meaning that what you say isn't credible but if some of it is more your own thoughts than actual findings if you know what i mean? Will HBM truly be a big saving on power? Will it offer much greater performance or just greater bandwidth? I know more bandwidth is better for bigger resolutions but what about 1080p or possibly 1440p? Will it offer a great performance increase at these resolutions?

Also in terms of DDR4 performance in the system, doesn't DDR4 use higher timings over DDR3 to get a higher bandwidth? I always thought bigger timings was bad for gaming?

As games get more demanding more bandwidth is required regardless of resolution, it's an ever increasing thing, it's not like 128GB/s is fine for 1080p and always will be, it's fine in a game that doesn't really require more to achieve a certain frame rate, but a game a year later might want 200GB/s and another game a year after that might want 150GB/s at 1080p. It also depends on frame rate, pushing 120fps vs 60fps wants far higher bandwidth.

Some people are happy with 60fps which is entirely fine, but for those won want the lowest persistence, lower motion blur and most smooth performance pushing 90fps + is where LCD tech starts to look REALLY good.

More bandwidth is required for more performance full stop, how much performance you need and at what resolution is and has always been personal choice. You might game using a 290x and only get 50-60fps in say Watchdogs and stuttering, generally poor performance, to get 90 or 120fps in that game requires a bigger gpu and it will also require more bandwidth to achieve that frame rate.

I'd take 90+ frame rate and a 120+hz screen every time over a higher res lower refresh rate, lower frame rate combination every time.

Increasing resolution increases bandwidth requirements, but so does increasing frame rate, or increasing graphical options. bandwidth requirements have never been and never will stay static. Going from 1080p to 4k certainly jumps the bandwidth requirement more than going from 60 to 120fps at 1080p, but it's still more bandwidth required. Better performance will always require more bandwidth.


HBM/HMC also reduce latency and absolutely, not just guess, reduces power, significantly. Even Intel's own papers on HMC talk about how both HMC reduces power/(GB/s) significantly but HBM is better on power usage than HMC... when Intel will say HBM saves even more power you pretty much confirm it's true.

It's relatively basic electronic behaviour, thicker wires and longer distances require more power to push a signal. PCB level traces and the size of the connections required to go from silicon to PCB scale connections are significantly bigger than when it's all done at the silicon scale which is what HBM does. You're talking 40-80nm traces with lengths maybe between 1-5cm vs traces that are massively larger at the pcb level and are pretty much up to 50CM long(there are multiple layers in a pcb and traces travel up and down several layers to get to where they are going, the straight line distance isn't indicative of actual length they'd be).

What would take around 80W for 4GB of Gddr5 to provide 512gb/s of bandwidth HBM does in 30W. Partially the power saving of the communication and partially the memory itself. Rather than using very high speed memory with a small bus, you use low speed memory with a very wide bus. So 1Ghz clock speeds with lower voltage instead of 2.5+Ghz memory with higher voltage. As with most memory, lower clock speeds means tighter timings, so you end up with, lower latency, significantly lower power memory chips all connected to a much wider bus to provide more bandwidth with lower latency, lower power. To cap it off HBM is TINY compared to normal memory because it's stacked. I forget the exact comparison but a 1GB stack is significantly smaller than a single 256MB memory chip and it's all on the single package connected to the pcb. So we also get significantly neater and smaller pcb's, we get easier to cool chips(due to vastly less complex surface mounted crap on the pcb) so better cooled VRM's, more standardisation in designs, meaning more likely that a single waterblock or air cooler will fit multiple generations of cards. If you look up pictures Nvidia showed of a mocked up Pascal and the tiny pcb size and the extremely neat look to it, that is why.

There is nothing but upsides to HBM and many of the same ones for HMC though not quite as many.

 

[eddyr]

Good read there drunken! Sometimes like to read your massive paragraphs lol. But i do sometimes think how accurate is the things you say. Not meaning that what you say isn't credible but if some of it is more your own thoughts than actual findings if you know what i mean? Will HBM truly be a big saving on power? Will it offer much greater performance or just greater bandwidth? I know more bandwidth is better for bigger resolutions but what about 1080p or possibly 1440p? Will it offer a great performance increase at these resolutions?

Also in terms of DDR4 performance in the system, doesn't DDR4 use higher timings over DDR3 to get a higher bandwidth? I always thought bigger timings was bad for gaming?

For comparable bandwidth it offers power savings, what they do with the headroom is up to them, e.g they could add more stream processors or run higher clocks and using that power headroom to increase performance/specs.

It was the same with DDR2 timings vs DDR3. DDR4 frequencies are/will go higher so it offsets the latency of having higher timings. The way memory addresses are accessed is supposed to have been altered to reduce latency also, but I don't know the intricacies of that.

 

[r7slayer]

Spoiler

As games get more demanding more bandwidth is required regardless of resolution, it's an ever increasing thing, it's not like 128GB/s is fine for 1080p and always will be, it's fine in a game that doesn't really require more to achieve a certain frame rate, but a game a year later might want 200GB/s and another game a year after that might want 150GB/s at 1080p. It also depends on frame rate, pushing 120fps vs 60fps wants far higher bandwidth.

Some people are happy with 60fps which is entirely fine, but for those won want the lowest persistence, lower motion blur and most smooth performance pushing 90fps + is where LCD tech starts to look REALLY good.

More bandwidth is required for more performance full stop, how much performance you need and at what resolution is and has always been personal choice. You might game using a 290x and only get 50-60fps in say Watchdogs and stuttering, generally poor performance, to get 90 or 120fps in that game requires a bigger gpu and it will also require more bandwidth to achieve that frame rate.

I'd take 90+ frame rate and a 120+hz screen every time over a higher res lower refresh rate, lower frame rate combination every time.

Increasing resolution increases bandwidth requirements, but so does increasing frame rate, or increasing graphical options. bandwidth requirements have never been and never will stay static. Going from 1080p to 4k certainly jumps the bandwidth requirement more than going from 60 to 120fps at 1080p, but it's still more bandwidth required. Better performance will always require more bandwidth.


HBM/HMC also reduce latency and absolutely, not just guess, reduces power, significantly. Even Intel's own papers on HMC talk about how both HMC reduces power/(GB/s) significantly but HBM is better on power usage than HMC... when Intel will say HBM saves even more power you pretty much confirm it's true.

It's relatively basic electronic behaviour, thicker wires and longer distances require more power to push a signal. PCB level traces and the size of the connections required to go from silicon to PCB scale connections are significantly bigger than when it's all done at the silicon scale which is what HBM does. You're talking 40-80nm traces with lengths maybe between 1-5cm vs traces that are massively larger at the pcb level and are pretty much up to 50CM long(there are multiple layers in a pcb and traces travel up and down several layers to get to where they are going, the straight line distance isn't indicative of actual length they'd be).

What would take around 80W for 4GB of Gddr5 to provide 512gb/s of bandwidth HBM does in 30W. Partially the power saving of the communication and partially the memory itself. Rather than using very high speed memory with a small bus, you use low speed memory with a very wide bus. So 1Ghz clock speeds with lower voltage instead of 2.5+Ghz memory with higher voltage. As with most memory, lower clock speeds means tighter timings, so you end up with, lower latency, significantly lower power memory chips all connected to a much wider bus to provide more bandwidth with lower latency, lower power. To cap it off HBM is TINY compared to normal memory because it's stacked. I forget the exact comparison but a 1GB stack is significantly smaller than a single 256MB memory chip and it's all on the single package connected to the pcb. So we also get significantly neater and smaller pcb's, we get easier to cool chips(due to vastly less complex surface mounted crap on the pcb) so better cooled VRM's, more standardisation in designs, meaning more likely that a single waterblock or air cooler will fit multiple generations of cards. If you look up pictures Nvidia showed of a mocked up Pascal and the tiny pcb size and the extremely neat look to it, that is why.

There is nothing but upsides to HBM and many of the same ones for HMC though not quite as many.

Ahh thanks Didn't really know to push more FPS, pushed up bandwidth requirements!
But yea i think id take a lower resolution with higher FPS and refresh (120Hz+) over a higher resolution and lower FPS.

Also didn't realise how much power savings HBM offered thats certainly over half the power savings, which is astonishing! Also that makes sense if you can have lower clocks you can tighten up the timings a fair bit! Can really see why AMD and intel are making the jump on this! Really quite excited for this technology now! Will HMB be possible in RAM too at some point? If it reduces latency and increases bandwidth i can see it been quite a good way to go! Obviously it may pose limits on the amount of system ram you could have i would assume? Going to check out those pics too, cheers!