AMD - Piledriver chitchat

[DanceswithUnix]

Remember that as time goes on games get more shader heavy. A good APU *should* be shader heavy on release, then it will be about right in a year or so and perhaps a bit computationally light by the time it goes EOL.

Having said that, I would love to see a 64 shader model with 6 CPU cores and some L3 cache. They could call it a Phenom 3, it would have graphics just good enough to run KDE/Aero desktops and enough grunt to not need AM3+ any more. I will stick to my Nvidia graphics card, but having a backup if the graphics card fails is nice.

L3 cache would boost the performance by maybe 5%, perhaps 10% in some games. If it used the same area as 2 more cores, which would you rather have?

[DanceswithUnix]

Yup, I knew it had an integrated memory controller - I was just under the impression there was something up with the FM2 platform regarding memory bandwidth, it must just be for the GPU though.

Is there any reason AMD wouldn't release higher end APUs with cache enabled? Seeing as we're not seeing any new AM3+ CPUs coming yet.

Cache takes a large area on the die, and right now they think that area is better used with graphics cores.

In the consumer laptop market I think they might be right. In the business laptop and desktop market, I suspect more cache would be better as people will either not care about graphics performance or will have external graphics.

[mikeo01]

They could call it a Phenom 3, it would have graphics just good enough to run KDE/Aero desktops and enough grunt to not need AM3+ any more.

Dull question incoming: Is there any reason WHY AMD can't just shrink the die and refresh it up a bit?

[DanceswithUnix]

Dull question incoming: Is there any reason WHY AMD can't just shrink the die and refresh it up a bit?

That used to be a common thing to do when a process shrink automatically made a chip faster and cheaper to make. These days there isn't much in it, so it probably isn't worth the roughly $1M it now costs to make new chip masks. You want to get more out of it than just a shrink.

I think there is also the problem that AMD have been tweaking the Athlon 64 core for years now. As fast as they tweak it, Intel tweak their now heavily modified P3 architecture and stay ahead. They really needed to do something more drastic, and although 8150 was a bit of a flop I still think they needed to do it. Their pace of change seems far better now.

[CAT-THE-FIFTH]

Another factor is that the newer AMD CPUs appear to be more modular in nature,and PD seems to work better without L3 cache too.

[scaryjim]

... I would love to see a 64 shader model with 6 CPU cores and some L3 cache ...

Judging by Kabini 128 shaders seems to be the smallest IGP AMD think is worth doing now, but yeah, the difference between a 128 shader IGP and an 800+ shader one should give plenty of room for at least a couple more cores: AFAIK the cores + L2 cache are only ~ 35% of the die area. If they're going to slowly move servers over towards APUs anyway they're eventually going to have to do 6 and 8 core APU dies just to hold position in the heavily-threaded market that they're already serving with 16-core MCM Opterons...

Any thoughts on the practicality of upping the L2 cache v. adding a separate block of L3? Sounds like it should be easier, but I'm a long way from an expert.

[DanceswithUnix]

Any thoughts on the practicality of upping the L2 cache v. adding a separate block of L3? Sounds like it should be easier, but I'm a long way from an expert.

L3 cache helps more with a single running badly threaded workloads, and also in core to core communication not having to go over the bus. As it is slower than L2, it can also have different properties like associativity. So it is kind of doing a different job.

[kalniel]

Any thoughts on the practicality of upping the L2 cache v. adding a separate block of L3? Sounds like it should be easier, but I'm a long way from an expert.

It could be harder from a die design point of view - it's got to be located a bit closer, you've got to have interconnects to each module etc. whereas L3 can be located where it's convenient, has fewer interconnects etc. I suspect you also get better accuracy (and thus yield) if you're cutting a more regular pattern in a larger block.

[DanceswithUnix]

It could be harder from a die design point of view - it's got to be located a bit closer, you've got to have interconnects to each module etc. whereas L3 can be located where it's convenient, has fewer interconnects etc. I suspect you also get better accuracy (and thus yield) if you're cutting a more regular pattern in a larger block.

I think it is simpler than that. The larger a block of cache is, the slower it is. Lots of modelling gets done against common instruction sequences to try and work out the
right balance between tiny ultra fast cache vs huge cache with better hit rate but much slower.

Put another way: if you don't find the data you need in cache, then you wait a few hundred clock cycles for it to come in from ram, which hurts. If you increase the size of the cache then you might add a clock cycle to every access that is a success, and that hurts too because that happens a lot. So if you have added L2 to the point that any more will slow the machine down, but you feel that for some workloads (games usually) you need more, then you have to add it as L3.

[mikeo01]

Not sure if this has been posted, it was posted on 21st, but I'll post it anyway.

Kaveri with 13 compute units / 832 shaders?

Planet3DNow. Apparently it's been uploaded on Sisoftware live ranker, source is at bottom of the page.

ASrock has came out with some boards.

[CAT-THE-FIFTH]

ARM based AMD tablets will be released:

http://www.sweclockers.com/nyhet/176...or-for-android

[CAT-THE-FIFTH]

BF4 Beta CPU results:

http://gamegpu.ru/images/remote/http...-bf_4_proz.jpg

An FX8350 matches a Core i7 2600K and the FX6300 destroys the Core i3 2100.

Edit!!

Some Total War Rome benchmarks:

http://gamegpu.ru/images/remote/http...e2_p2_proz.jpg
http://gamegpu.ru/images/remote/http..._p2_proz_k.jpg

The FX6300 looks faster overall when compared to a Core i3 2100,and so I suspect it and the FX6350 will compete reasonably well with the IB and Haswell Core i3 CPUs.

On the top end the Core i5 CPUs are faster though.

[scaryjim]

BF4 result makes sense, they've already talked about the game engine fully utilizing 8 CPU cores when they did the Mantle unveil. With next gen consoles having 8 x86 cores, any game thats being co-developed for both console and PC should massively benefit from core count...

[DanceswithUnix]

BF4 result makes sense, they've already talked about the game engine fully utilizing 8 CPU cores when they did the Mantle unveil. With next gen consoles having 8 x86 cores, any game thats being co-developed for both console and PC should massively benefit from core count...

Or thread count, 12 thread Intel stuff should get a boost too.

This should be interesting to watch.

[HalloweenJack]

depends - if mantle is doing lots of integer stuff - with gpu doing fpu work then octo core will shine over intel

[scaryjim]

... if mantle is doing lots of integer stuff - with gpu doing fpu work then octo core will shine over intel

Mantle gives low-level access to the GPU hardware, so is unlikely to be doing int work And while it might allow developers to offload flops to the GPU, I don't think that's the point.

It's more the case that because Mantle allows low level access to the GCN hardware, you bypass some of the CPU-managed translation layer that OpenGL/DirectX require, which reduces load on the CPU, freeing it up to handle more of the well-threaded rendering and game management that can't be done on the GPU. In that sense, it'll actually benefit Intel+AMD rigs as much as it will all AMD rigs.