how powerful will the next APU?

[from416]

I am waiting for next genegration of kaveri APU,

how powerful will the GPU?

will like R280?

 

[ShintaiDK]

It wont be much faster than Kaveri, and its quite long time away.

 

[Qwertilot]

My impression from reviews etc is that they're more or less entirely bandwidth limited on the GPU side right now, so nothing real until that problem gets fixed one way or another, and then maybe rather substantial changes when it does.

 

[el etro]

HBM(in-die stacked RAM) can change the game for the APUs someday. Another thing that AMD depends to change is the graphical performance for the size of the graphics processor inside the APU. 512 stream processors is what they can put on the APU per now.

 

[NTMBK]

The next APU is Carrizo, which apparently will be on the same FM2+ platform. This means that it can't have any more memory bandwidth than Kaveri, unless it has some massive on-package cache like the Iris Pro.

 

[Insert_Nickname]

The next APU is Carrizo, which apparently will be on the same FM2+ platform. This means that it can't have any more memory bandwidth than Kaveri, unless it has some massive on-package cache like the Iris Pro.

I doubt that. AMDs main focus with Carrizo appears to be an integrated/on-die southbridge. Just like Kabini has. That alone will properly eat a lot of transistors and die area, even with a shrink to 20nm.

 

[el etro]

No one have any news with the 14nm GloFo process status??

 

[sandorski]

280 level performance would be sweet, but probably not for a few generations. By that time it will be treading water performance wise.

 

[skipsneeky2]

Any plans upcoming for some sort of DDR4 based Apu platform? My guess and perhaps a terrible one is a huge boost when DDR4 is here and at higher speeds and affordability is about DDR3 currently is.

 

[monstercameron]

so far there is one rumor in my subreddit about carrizo, apparently https://twitter.com/jbazzrea toured the campus and was given a sneak peak into the upcoming apu

Carrizo is going to be a nice performance increase, in more ways than one... ...expect large gains in GPU and large gains in CPU performance for multiple reasons.
Carrizo may seriously offer a solid replacement, even in an APU (seriously...news of this soon)

I can't seem to find the tweet so take with a grain of salt

 

[ShintaiDK]

so far there is one rumor in my subreddit about carrizo, apparently https://twitter.com/jbazzrea toured the campus and was given a sneak peak into the upcoming apu



I can't seem to find the tweet so take with a grain of salt

Performance/watt, not performance.

http://www.amd.com/en-us/press-releases/Pages/amd-accelerates-energy-2014jun19.aspx

 

[SPBHM]

llano is from 2011, kaveri 2014, in almost 3 years the gain was round 50%?

r9 280 seems unrealistic for a long time.

Kaveri is slower than some mid range cards from 5 years ago.

 

[frozentundra123456]

llano is from 2011, kaveri 2014, in almost 3 years the gain was round 50%?

r9 280 seems unrealistic for a long time.

Kaveri is slower than some mid range cards from 5 years ago.

If ever. That card (r9 280) has 3 times the shaders of Kaveri and a 384 bit bus. Even in a couple of generations, I would expect around 7770 level at best. That would be decent now, but by the time an APU gets that level, games will be more demanding and discrete cards hopefully will be on 20 nm and offer much better performance and lower power than today's cards.

 

[ShintaiDK]

To reach R9 280 stacked DRAM is needed and at least a real 14nm process and better yet 10nm. I am not talking about TSMC or Samsungs fake 14nm FF that is really 20nm with FF. So assuming all works out...2019?

 

[BrightCandle]

Stacked DRAM is very likely and very quickly going to change the APU equation. Getting enough bandwidth so incredibly close to the GPU core, and having it shared with the CPU could dramatically improve the amount of performance you can put into an embedded GPU so that it can directly compete with a discrete card. Very low latencies and very high bandwidth from stacked VRAM could really change the equation especially if it helps the CPU side along as well. The first company out with this tech is very likely going to have an advantage until the other releases it too, just like when memory controllers were embedded in the CPU (which is what gave AMD the upper hand in the Athlon days).

 

[ShintaiDK]

Stacked DRAM is very likely and very quickly going to change the APU equation. Getting enough bandwidth so incredibly close to the GPU core, and having it shared with the CPU could dramatically improve the amount of performance you can put into an embedded GPU so that it can directly compete with a discrete card. Very low latencies and very high bandwidth from stacked VRAM could really change the equation especially if it helps the CPU side along as well. The first company out with this tech is very likely going to have an advantage until the other releases it too, just like when memory controllers were embedded in the CPU (which is what gave AMD the upper hand in the Athlon days).

Stacked DRAM however doesnt do much as such to get to 280 performance. A double node shrink, and more likely tripple node shrink is needed for an APU to perform this way within acceptable power consumption.

There is no stacked DRAM on the roadmaps for the next APU either. For dGPUs its target is more around 2016/2017.

 

[BrightCandle]

Stacked DRAM however doesnt do much as such to get to 280 performance. A double node shrink, and more likely tripple node shrink is needed for an APU to perform this way within acceptable power consumption.

There is no stacked DRAM on the roadmaps for the next APU either. For dGPUs its target is more around 2016/2017.

I agree, its not here yet. When it is however I think APUs will get a lot closer to discrete GPUs than they can currently today. An APU today is held back enormously by memory bandwidth and latency as well as thermal headroom. When its just thermal headroom more performance on board is possible and it will be closer to a discrete card. It will still lag of course but eventually maybe HSA will make a difference. None of this is obviously soon.

 

[Ajay]

No one have any news with the 14nm GloFo process status??

IDontCare indicated that Samsung is dragging their feet a bit when it comes to helping GF moved to 14nm. They are competitors to a certain extent, so I'm not really surprised. At this rate it's going to be a at least a couple of years before GF can move any volume @ 14nm. So the "so sad saga" continues for AMD since GF isn't doing it any favors.

 

[ShintaiDK]

IDontCare indicated that Samsung is dragging their feet a bit when it comes to helping GF moved to 14nm. They are competitors to a certain extent, so I'm not really surprised. At this rate it's going to be a at least a couple of years before GF can move any volume @ 14nm. So the "so sad saga" continues for AMD since GF isn't doing it any favors.

One could add, we still need to see 20nm products. That according to various foundry roadmaps have been in production since 2012

Not to mention all the years we already used DDR4. :awe:

Or that its not really a 14nm. But a 20nm with FF that will lower power draw on the expense of performance.

 

[monstercameron]

how feasible would it be for amd to just stitch 4 jaguar cores and necessary uncore to a 280 die...and just use an off chip south bridge. Not sure but I doubt that would add 50mm^2 to that die and they already have tech for shared memory...so most of the "hardwork" has been researched.

 

[ShintaiDK]

how feasible would it be for amd to just stitch 4 jaguar cores and necessary uncore to a 280 die...and just use an off chip south bridge. Not sure but I doubt that would add 50mm^2 to that die and they already have tech for shared memory...so most of the "hardwork" has been researched.

Whats the TDP for a 280 again? :sneaky:

 

[monstercameron]

Whats the TDP for a 280 again? :sneaky:

high...and your point is?

 

[Enigmoid]

how feasible would it be for amd to just stitch 4 jaguar cores and necessary uncore to a 280 die...and just use an off chip south bridge. Not sure but I doubt that would add 50mm^2 to that die and they already have tech for shared memory...so most of the "hardwork" has been researched.

Four jaguar cores is really not going to be able to power that much GPU.

 

[monstercameron]

Four jaguar cores is really not going to be able to power that much GPU.

mantle begs to differ.

 

[Ajay]

One could add, we still need to see 20nm products. That according to various foundry roadmaps have been in production since 2012

Not to mention all the years we already used DDR4. :awe:

Or that its not really a 14nm. But a 20nm with FF that will lower power draw on the expense of performance.

Well, I don't think we know any metrics about GF's 20nm - never mind Samsung's 14nm FF.

20nm is, AFAIK, still based on IBM's IP + GF R&D - so it's decoupled from 14FF. As far as nodes go, they've pretty much been just labels for some time now, so yes, 14 FF is a new node - it just doesn't scale as well visa vi the previous decade. This is just a one time deal - 10nm and lower should have better scaling than 20nm (GF)->14 FF (Samsung) will; just now the scaling between xtor dimensions and those of the metal layer will no longer follow traditional norms.

And foundry roadmaps - they belong in the North Pole with Santa since they are clearly make believe