Originally Posted by TheLAWNOOB
I guess the 28nm process with the steamroller core will help the APUs to lower power usage and have a nice boost in performance
However, I think I heard somewhere that GDDR5 has bad latency compared to DDR3. Maybe HSA will fix that since the CPU and GPU uses the same RAM
Not just 28nm, AMD is also using high density libraries as of Excavator (Although the die-shot that some people were thinking was SR had it too if I recall correctly) which is equivalent to a full process node change.
Originally Posted by anothergeek
Yes, GDDR5 produces more bandwidth, which is well suited to 3D applications. I use lower latency DDR3, as desktop and operating processes benefit from that.
If AMD manufactures kaveri chipsets for mainstream performance notebooks, they may very well solder 4 memory chips to the APU interface, which would negate latency to begin with, thus bringing on HSA again. It wouldn't need to be any memory type in particular, directly attaching high speed RAM to the memory interface will greatly improve performance, versus using a typical memory controller. Given time, the sample can be 4x4 allotments, negating a standard DDR4 memory controller altogether.
HSA is just aiming to put X86 6 feet under
iirc the latency for GDDR5 isn't much higher than DDR3 when you take everything into account, plus can you actually name a program other than WinRAR that shows benefits from faster memory/tighter timings? RAM has been more than fast enough for CPUs for a while now and has been increasing at a rate more than fast enough to match the CPUs needs.
Originally Posted by A Bad Day
A well optimized software could run on an A10 as well as an i7, in the most ideal situation.
Essentially AMD is attempting to undercut Intel's CPU advantage through GPU computing.
"So your CPU cores are 20% faster? A shame that they're not going to match my GCN cores."
There's a reason why software rendering mode doesn't exist anymore for almost all new games.
Indeed, a build of Handbrake running OpenCL and AVX on an A10-6800k is slightly faster than the normal build on an i7 3960X, both chips at stock. (I'd wager that the 3960X would be a hell of a lot closer when you have AVX enabled on it too, if not ahead and this is probably one of the best case scenarios for the A10, but it shows how good the benefits of OpenCL with a fast IGP can be)
Originally Posted by IvantheDugtrio
When it detects a dedicated GPU (that isn't a 6670) it will shut down the integrated graphics and run off the dedicated GPU. .
You can also run the IGP as well regardless of GPU (It might only be if you run AMD dedicated though, not entirely sure) if you want the OpenCL performance benefits or extra screens, I was doing it with a 790GX and even the HD 4000 graphics in my i5 support it.
Originally Posted by OverclockerFox
I guess this is slightly good news. AMD's weakness in platform is in not having PCI-E 3, as well as other things I can't remember right now.
Are they saving Steamroller on desktop until they have a smaller process node available?
How is it really a weakness? AMD has 16x/16x whereas 115* has 8x/8x, meaning with PCIe on Intel the bandwidth is matched...When AMD gets PCIe 3.0 they'll be matching the bandwidth of Socket 2011 systems, not the mainstream ones.
Not to mention, not many people buy an AMD CPU and then two GPUs.
Originally Posted by james8
^the improvement should be quite noticeable in certain applications. one example being Flight Simulator X
Not when AMD is able to run 16x/16x and Intel is running 8x/8x, otherwise you're not really going to come close to maxing out the PCIe bus with a single GPU regardless of game. (iirc only OpenCL and the like would show a difference and even then, it's not often)
Do we know how many modules/cores these APUs will have? I'm hoping we get a 3 module SR APU, I'd go for that over a 4 module CPU as Handbrake can make use of the iGPU capabilities and provides insane price/performance when it does