• We’re currently investigating an issue related to the forum theme and styling that is impacting page layout and visual formatting. The problem has been identified, and we are actively working on a resolution. There is no impact to user data or functionality, this is strictly a front-end display issue. We’ll post an update once the fix has been deployed. Thanks for your patience while we get this sorted.

APU faster than GPU?

U

uribag

Member
If the memory bottleneck can be removed (with HBM, for instance), theoretically an APU can be faster than a similar dedicated GPU because the communication between CPU and GPU is better?

Sorry for the non technical words.
 
Well, first we dont have HBM yet, and no definite date when, except for rumors/hopes from AMD fans. Then you might be able to match a discrete gpu *of the same composition* or even exceed it if the software is written properly. Problem is, even if you remove the bandwidth restriction, you are still going to be limited to by thermals, die size and TDP to something similar to a low/midrange discrete.
 
Gikaseixas said:
28 to 20nm will help there
Click to expand...

Perhaps, but there will also be 20nm (or 16 or whatever) discrete gpus, which will hopefully
bring a large increase in performance as well, so the relative performance of an APU to dgpu may not improve. APUs are basically trying to catch up to a moving target.
 
frozentundra123456 said:
Perhaps, but there will also be 20nm (or 16 or whatever) discrete gpus, which will hopefully
bring a large increase in performance as well, so the relative performance of an APU to dgpu may not improve. APUs are basically trying to catch up to a moving target.
Click to expand...

This is true. I would then say that APUs could reach cheaper GPU solutions or even mid range previous gen cards such as 270x or GTX 760. Interesting times ahead
 
Last edited:
Latency is not really an issue gaming wise. GPGPU however is a different matter where the APU may shine.

Also you would have to look beyond all the restrictions of the APU like TDP, size etc that limits the APU as well.
 
uribag said:
If the memory bottleneck can be removed (with HBM, for instance), theoretically an APU can be faster than a similar dedicated GPU because the communication between CPU and GPU is better?

Sorry for the non technical words.
Click to expand...

yes
 
Thanks for the answers guys, but I was trying to compare more in line with ShitaiDK´s thoughts: not absolute performance of APUs in the future (nodes, TDP, etc.), but it´s relative performance in comparison with CPU+dGPU in gaming or GPGPU scenarios without the memory bottleneck.
 
uribag said:
If the memory bottleneck can be removed (with HBM, for instance), theoretically an APU can be faster than a similar dedicated GPU because the communication between CPU and GPU is better?

Sorry for the non technical words.
Click to expand...

uribag said:
Thanks for the answers guys, but I was trying to compare more in line with ShitaiDK´s thoughts: not absolute performance of APUs in the future (nodes, TDP, etc.), but it´s relative performance in comparison with CPU+dGPU in gaming or GPGPU scenarios without the memory bottleneck.
Click to expand...


It isn't so much that the GPU and CPU can communicate faster with dedicated memory, but the GPU has more bandwidth which is needed for good gaming performance.

As an example, my Radeon 7970 with a modest bump on the memory speed (1430MHz) and the 384 bit connection to the GPU provides 276GB/s of bandwidth to just the GPU, the CPU gets none of that. That is 3GB of GDDR5 dedicated to just the GPU with that amount of bandwidth. (Keep in mind my 7970 has 2048 GCN cores running at 1030MHz -a slight overclock-, a high end AMD APU has 512 GCN cores running at, I believe up ~960MHz... more powerful GPU's need more bandwidth to perform up to what they are capable of.)

My CPU has a 128 bit connection to the memory (2 x 64bit channels of DDR3 memory) running at ~1760MHz providing something like ~23GB/s of bandwidth, just to the CPU. That's the problem with APU's, both the GPU and CPU have to share that DDR3 connection which provides significantly less bandwidth compared to a discreet video card with it's own onboard memory. So making the APU's GPU portion much bigger and more powerful won't get you much in a lot of gaming situations because the memory bandwidth will often limit performance (and remember the CPU also is using some of that memory bandwidth).

With HBM or other high speed memory (the PS4 uses four DDR5 channels for a 256bit connection of fast memory that the GPU and CPU share in the APU, as an example) performance can increase and keep up with a more powerful GPU core, it can give it the bandwidth it needs.

So yea, as memory technology increases we'll see better and more powerful APU's. Right now I think they've gone about as far as they can go on dual channel DDR3, more bandwidth just isn't there right now.
 
Technically we are already there -- Where there are many APU's faster than current dedicated video cards. One my recent trip to Best Buy, they had a lot of different R5 230's, some that cost up to $69. I know that a Richland or Kaveri A8's or A10 can easily mop with an R5 230. There are definitely low end to cheaper midrange dedicated cards that already get walloped by current APU's.
 
uribag said:
If the memory bottleneck can be removed (with HBM, for instance), theoretically an APU can be faster than a similar dedicated GPU because the communication between CPU and GPU is better?

Sorry for the non technical words.
Click to expand...
If what you say is true, then shouldn't dGPU performance be bottlenecked by the latency between cpu-gpu?

dGPUs also stand to gain from additional bandwidth, so in the grand scheme of things, APUs will likely remain in the budget category. :whiste:
 
There are definitely low end to cheaper midrange dedicated cards that already get walloped by current APU's.
Click to expand...
I think a lot of that depends on what you mean by "low-end" to "midrange".

Yes, you can certainly cheap/old video cards that are slower than the fastest APUs, but those really aren't gaming cards. Those are cards to replace a part in an OEM system, or a cheap upgrade to drive extra monitors or to do basic video playback with CPUs that lack or have terrible integrated graphics.

When most people talk about mid-range cards for gaming, although they may mean many different things, I typically see this refer to some neighborhood around ~150 USD, and things like the 750(Ti)->760, and the R9-270(x)->R9-280.

The A10-7850k isn't at the same table as ANY of those.

Edit to add:
http://www.anandtech.com/print/7677/amd-kaveri-review-a8-7600-a10-7850k

It's uniformly slower than AMD's 6750 (although it is notably kind of close in CoH2) which is comparable (and probably a bit slower than) what are IMHO the entry level gaming GPUs, the 7750/R5-250 (w/ gddr5) and the 650/740 (w/ gddr5).
 
Last edited:
MiddleOfTheRoad said:
Technically we are already there -- Where there are many APU's faster than current dedicated video cards. One my recent trip to Best Buy, they had a lot of different R5 230's, some that cost up to $69. I know that a Richland or Kaveri A8's or A10 can easily mop with an R5 230. There are definitely low end to cheaper midrange dedicated cards that already get walloped by current APU's.
Click to expand...
I've no idea who would even buy an 80-stream R5 230 given 384-stream R7 250's are exactly the same $70 price on Amazon, with 640-stream R7 250X's GPU's for $10 more, and 896-stream R7 260X's not far behind. The R5 230 is total junk that would be overpriced even at $20. And as "Essence_of_War" said - even an R7 260X / 750Ti (twice the speed of an A10-7850K) is classed "low-mid" by 2014 standards, with mid-range typically being R9 270(7850) / 270X (7870) / GTX 760. By the end of next year, the "mid range" moving target may well end up a GTX 960.

The "APU's will soon beat mid-range dGPU's" always faces the same 2 problems : 1. dGPU's are a moving target. As soon as you introduce a new tech (eg, 20nm), the dGPU's incorporate it as well so there's no real relative gain, and 2. The ability to "split the heat" over two heatsinks means typically higher performance for any given thermal / noise target.
 
Stacked memory is the new tech which will disproportionally (considerably) help the APUs vs the GPUs. When that comes in it'll mostly just be the combined heat as a barrier to how many of the dGPU market APUs can swallow.
 
Qwertilot said:
Stacked memory is the new tech which will disproportionally (considerably) help the APUs vs the GPUs. When that comes in it'll mostly just be the combined heat as a barrier to how many of the dGPU market APUs can swallow.
Click to expand...

Of course, unless Intel were to implement APU type designs for AMD and Nvidia gpus, the CPU performance in an AMD part is still weak. Intel for it's part seems to have little interest chasing the high end gpu market.
 
And AMD also still well off Maxwell's efficiency I guess. Maybe they'll partially fix that sometime soon. Does matter though.

Melding a 35w i5/7 Broadwell/Skylake with Maxwell/Pascal with no memory bandwidth restrictions could produce something very effective indeed, even with staying at <100w total. If you're a chunk less power efficient on both components it gets much harder 🙁

Maybe a guess a chance that Intel will get more interested in very big iGPUs when stacked memory comes mainstream. We'll find out in time.
 
Gikaseixas said:
If HBM was to be implemented it would provide a significant performance increase. We're starting to see leaks for a possible dedicated arch with HBM from the usual suspects. Lets hope it is true for the sake of graphics advancement.

http://www.tweaktown.com/news/41040...n-gpus-internally-known-as-fiji-xt/index.html

http://www.tweaktown.com/news/40339...early-next-year-20nm-with-hbm-tech/index.html
Click to expand...

Those pics show the datarate dropping from 7Gbps to 1 Gbps for the new architecture. Is that right?
 
You must log in or register to reply here.

TRENDING THREADS

Back
Top