AMD - Piledriver chitchat

[CAT-THE-FIFTH]

I like that hardware.fr tested a HD6670 GDDR3 too,as it seems the A10-5800K basically has an IGP with similar performance,when using 1600MHZ DDR3 it seems. So basically you are getting the performance level of a £40 to £50 graphics card.

With 2133MHZ DDR3 RAM and an IGP overclock,that is another 20% added performance:

http://hexus.net/tech/reviews/cpu/46...er-ram/?page=3

Is it just me or have SA removed + closed all comments? Can't say I blame them considering some of the rubbish which ends up posted there...

It seems they did so:

http://semiaccurate.com/forums/showthread.php?t=6651

[Biscuit]

Im not surprised, basically accusing Intel of anti-competitive business practices, without sufficient source information and evidence that's Libel if im not mistaken?

Wouldn't be the first time Intel did something like that though would it?
http://news.bbc.co.uk/1/hi/business/8047546.stm

[watercooled]

Those articles were posted yesterday and today, comments have been closed for some time now.

But Intel have done it before, would it really surprise anyone were it true?

[DanceswithUnix]

Those articles were posted yesterday and today, comments have been closed for some time now.

But Intel have done it before, would it really surprise anyone were it true?

They have done it so many times in the past that I can only boggle at their cheek of doing it again so soon. Guess they must be scared to risk more dragging through the courts.

[badass]

I also hope to see a better/more available range of mITX boards for FM2.

I've just done a search on scan for FM2 and there are loads of uATX motherboards but no section for them!

I seriously considered mITX before deciding on uATX for my current build but there wasn't enough choice. I preferred the mITX cases though.

[DanceswithUnix]

Ebuyer have gone from "pre-order" to having quite a lot of stock.
In the process, the price of the A10-5700 has dropped from 97.48 to 89.99, some of the others have shuffled by about a quid.

Still no motherboards though.

[CAT-THE-FIFTH]

[badass]

looking at the various gaming results, two things are obvious:
The A10 has some serious IG power.
The memory bandwidth is starving both the IGfx and the CPU (to a lesser extent)
I'm sure that if AMD were to make an Imaginary A10 with an on package/on interposer cache that's got lots of bandwidth, it would easily see off Haswell as well.

[CAT-THE-FIFTH]

Kaveri,will be more efficient at using memory bandwidth,since it uses GCN cores which means less contention issues between the CPU and GPU for memory bandwidth.

It seems AMD is working on interposer technology too:

http://forums.hexus.net/cpus/200143-...ml#post2355217

I just hope it makes its way into Kaveri.

Kaveri looks like it will have a monster of an IGP!!



http://www.extremetech.com/computing...en-cpu-and-gpu

It seems there is nearly a 50% jump in FLOPS going from the Trinity IGP to the Kaveri IGP.

[scaryjim]

... It seems there is nearly a 50% jump in FLOPS going from the Trinity IGP to the Kaveri IGP.

I sincerely hope Kaveri brings some serious special sauce for memory bandwidth then, since AMD have already committed to socket FM2 for another generation, meaning it's definitely stuck with a dual channel DDR3 memory interface. I'm not convinced about GCN using memory bandwidth more efficiently either - if that was true surely AMD would have stuck to 256bit for discrete GCN, rather than stepping up to a more expensive 384bit bus? Perhaps HSA can bring some improvements, but it isn't going to change the fact that the theoretical bandwidth will not increase significantly, which suggests Kaveri will still be memory-hampered. All the efficiency in the world isn't going to matter when there's still a hard limit to the amount of data you can shift...

[CAT-THE-FIFTH]

I sincerely hope Kaveri brings some serious special sauce for memory bandwidth then, since AMD have already committed to socket FM2 for another generation, meaning it's definitely stuck with a dual channel DDR3 memory interface. I'm not convinced about GCN using memory bandwidth more efficiently either - if that was true surely AMD would have stuck to 256bit for discrete GCN, rather than stepping up to a more expensive 384bit bus? Perhaps HSA can bring some improvements, but it isn't going to change the fact that the theoretical bandwidth will not increase significantly, which suggests Kaveri will still be memory-hampered. All the efficiency in the world isn't going to matter when there's still a hard limit to the amount of data you can shift...

Its more down to the fact GCN supports x86 addressing,so there will be unified memory addressing between both the CPU and GPU. ATM,this is not the case so you have more chance of contention issues between both the CPU and GPU ATM. Lower end GCN cards do use a 128 bit memory bus with GDDR5,so the memory bus has not changed for lower end discrete cards.

Edit!!

Even the CPU section of Llano seems to benefit from faster RAM:

http://www.techspot.com/review/418-a...apu/page8.html

[scaryjim]

AFAIK there's very little CPU to GPU memory copy in gaming though, so the ability to address the same data with CPU and IGP won't actually benefit gaming performance that much - it'll just make it easier for developers to leverage both CPU and GPU cores as appropriate. Although if there's evidence otherwise out there I'm happy to be corrected

As to the Llano memory speed results, they could just mean that AMDs caches and memory controller are sufficiently badly implemented that they bottleneck the CPU cores - which is a serious issue if it's true! That said, AMD specifically said that in Llano the GPU was given memory priority, so I'm not surprised the CPU responded well to increased memory bandwidth: it was probably being memory-starved with the GPU taking the vast majority of the bandwidth available at 1333MHz (let's not forget that even modern windows desktops are GPU/DX rendered, so the GPU is basically being at least lightly loaded all the time...).

[DanceswithUnix]

AFAIK there's very little CPU to GPU memory copy in gaming though, so the ability to address the same data with CPU and IGP won't actually benefit gaming performance that much - it'll just make it easier for developers to leverage both CPU and GPU cores as appropriate. Although if there's evidence otherwise out there I'm happy to be corrected

Isn't that a chicken-egg situation though?
ISTR current discrete graphics utterly suck at cpu <--> gpu data movement. If the developer wanted to do it, they basically can't and would have to find another way.

[CAT-THE-FIFTH]

LOLWTF:

http://www.youtube.com/watch?feature...&v=OmQI_NIRCgw

Trinity contain tablet?? Probably Hondo though I suspect.

[scaryjim]

ISTR current discrete graphics utterly suck at cpu <--> gpu data movement.

Yeah, CATs link demonstrated that nicely. I'm just not sure that sharing data between the CPU and GPU would make that much difference to gaming performance - do they need to work on the same large datasets that much? I thought the increased memory bandwidth helped graphics because you need to shift large textures and pixel datasets around between GPU elements - otherwise you'd expect PCIe bandwidth to have a much higher impact on gaming performance than it does...

Of course, if there is a way to hugely increase performance by sharing data between the CPU and GPU then AMDs moving in the right direction to benefit from that: but it won't be evident at Kaveri release because there won't be any games that have developed for it

Basically, I expect Kaveri to remain heavily bandwidth (and therefore memory clock) dependent on release, and potential gaming performance to suffer because of it.

[DanceswithUnix]

Phoronix playing with memory speeds:

http://www.phoronix.com/scan.php?pag...ity_ddr3&num=3