AMD - Piledriver chitchat

[scaryjim]

Truth is we don't know what the TDP restriction will do to the performance on Trinity: the 17w version is a single module with low clock speeds, and at 17w it doesn't have much TDP headroom for turbo. That said, the rumours I've picked up on are that CULV Sandy Bridge had appalling graphical performance because it simply didn't have enough TDP headroom to turbo the graphics at all.

Let's conjecture:
Given desktop dual-core Llano (A4-3300) has 160 cores at 443 MHz and Mobile Llano (A4-3300M/3310MX) has 240 shaders at 444MHz, even at 327MHz the 256shaders in 17W Trinity should ensure it performs equal to or faster than either of those, and if it reaches full turbo it should be noticably faster. CPU clock speeds are similar to mobile Llano dual cores, so CPU performance should be equivalent or slightly better (as Trinity seems to have at least as good IPC as Llano) - at least for lightly threaded workloads. I suspect heavily threaded workloads might suffer somewhat due to the shared components in Trinity. Still, for most users 17W trinity should be at least as good as, if not better than, low end 35W Llano, but obviously with much better power consumption.

Hmmm, proper reviews and retail products needed ASAP, methinks

[badass]

(as Trinity seems to have at least as good IPC as Llano)

That's a good point. They have taken an architecture that was designed with high clock speeds in mind, but not high IPC that was power hungry and had a low IPC and given it IPC to match the older architecture but with the boost on clock speed and a reduction in power.
Piledriver is really looking like what Bulldozer should have been.
I just think the IGP is strangled. It might be substantially faster than Ivy Bridge but I bet with 50% more memory bandwidth, we'd see over 45% more FPS in games and an utter IVB annihilation in the Graphics front.
I'm still holding out hope that the Trinity VCE is actually as fast as the IVB one once they get the SDK out there.

[CAT-THE-FIFTH]

To me it seems the same situation as llano. The performance for most people is enough if not more than enough but the public isn't informed enough to know that. Intel win the business because in benchmarks they rule.

AFAIK,it seems both Zacate and Llano have done very well in China which is one of the most rapidly expanding computer markets in the world. I suspect at least in many markets the GPU performance for the price must be quite good against the laptops with similarly priced Intel CPUs.

The thing is though in Europe and the US the situation might be somewhat different due to the relative pricing difference between Intel and AMD laptops,which might not be the same as the rest of the world.

[Noxvayl]

Reality says they wont scratch ivy bridge but it would be good for AMD if they matched olde sandy bridge in the desktop space.

Don't know if you noticed but there isn't much difference between Ivy and Sandy Bridge: http://www.anandtech.com/bench/Product/287?vs=551

[CAT-THE-FIFTH]

I think the 8 thread Vishera CPUs will match(or be just behind) the IB Core i7 quad cores in multi-threaded applications while consuming more power. Even if Vishera only has a 10% single threaded IPC increase,it wouldn't surprise me if launched with a stock clockspeed of over 4GHZ. At 4.2GHZ,you would be looking at 25% more performance in multi-threaded applications over an FX8150.

I also suspect in lightly threaded applications it will beat the Phenom II X4 but will be behind the Core i5 CPUs. So,I still suspect the Intel CPUs will be better for gaming.

[DanceswithUnix]

Looks like BSN think desktop Trinity is out in June.

Bottom of first page, http://www.brightsideofnews.com/news...es-fusion.aspx

"But beware, this also includes desktop models, which are planned to be launched in 65W and 100W bins during Computex Taipei 2012, which is being held from June 5-9 in Taiwan."

[scaryjim]

One of the previously linked news stories (can't find it now though ) suggested that desktops with Trinity in would launch soon, but the APUs wouldn't be available through the channel until later in the year - i.e. OEMs have been given priority access to the first batches of chips. I guess that kind of makes sense for AMD (I imagine the pre-built PC market is much bigger than the home-built ) but it's a bit disappointing for those of us wanting to home build a new HTPC...

[DanceswithUnix]

Pictures of an early FM2 sample motherboard.

http://en.expreview.com/2012/05/21/a...red/23168.html

[CAT-THE-FIFTH]

I had some thoughts on the desktop A10-5800K:

http://3.bp.blogspot.com/-bB_PdG5id-...ons_leaked.jpg

This has a 800MHZ GPU.

Now lets do some Moose level numbers.

The mobile version run at a maximum of 685MHZ:

http://www.anandtech.com/show/5831/a...m-a-new-hope/6

This is a comparison of the mobile A10 and mobile HD4000:

http://images.anandtech.com/doci/583...gaming-new.png

The HD4000 in the Core i7 3720QM runs at between 650MHZ to 1250MHZ.

The HD4000 in the Core i7 3770K runs at 650MHZ to 1150MHZ.

With the higher clockspeed,desktop chipset and potential to use 1866MHZ RAM,it probably will be at least another 15% to 20% faster overall than the mobile A10,especially with the increase in lightly threaded performance with the desktop A10. The mobile HD7660G has around 20% greater performance than the mobile HD4000 GPU.

That would probably mean a 40% to 45% improvement in overall GPU performance over the desktop HD4000.

This would mean the desktop HD5570 and GT440 cards should easily be surpassed in theory. The HD6550D is already around HD5570 or GT440 level in many games:

http://www.anandtech.com/show/5771/t...3770k-review/9

This is with 1600MHZ DDR3 RAM too.

In this case,it would mean that any GPU under £50 might be not faster!

[CAT-THE-FIFTH]

First review of an IB ULV CPU:

http://www.anandtech.com/show/5843/a...ux21a-review/1

The A10-4655M has an IGP clocked 28% lower than the one in the A10-4600M,but the CPU clockspeed is only 15% lower overall.

I suspect,in most gaming situations the A10 based thin and light laptops are going to do better IMHO.

[CAT-THE-FIFTH]

It seems Kaveri might use GDDR5:

http://translate.google.com/translat...spx%3FId=33384

[DanceswithUnix]

It seems Kaveri might use GDDR5:

http://translate.google.com/translat...spx%3FId=33384

Seems very unlikely. Main memory would still have to be DDR3 for the CPU to use, so this would have to be another channel or two as well as that. return of sideport memory perhaps? How would that fit in with the aim of making all memory work/address the same through the MMU?

There would be a real potential to confuse customers here, and given more memory channels wouldn't it be better to just make them DDR3 and unified with the rest of the memory?

[scaryjim]

GDDR5 off-chip cache, perhaps? Or a stacked module with GDDR5 sitting on top of the APU silicon? (although not sure how badly that would affect thermals). The translated article - in as much as it's possible to interpret the machine-translated text, anyway - seems to suggest that there will be GPU-specific memory, and that lower-specced APUs will make do with just system memory. That suggests there'll be a secondary memory controller that can be interfaced to some GDDR5 to provide graphics-only memory, although that would take up a lot of additional silicon and doesn't sound very heterogeneous to me. A wider primary memory controller would make a lot more sense, or a move to DDR4 (which as I understand it is a point to point protocol, giving each DIMM its own channel, basically).

I'd much rather see three or four DDR3 channels with a similar bandwidth management to the current system - you could basically end up with a dedicated 128 or 192 bit path for the GPU when under heavy graphical load, with a 64bit channel "spare" to supply the CPU...

[CAT-THE-FIFTH]

Haswell is rumoured to have on-die stacked RAM for the GPU and it seems that AMD might be also working on something similar too.

SA,leaked this picture last year:

http://semiaccurate.com/2011/10/27/a...-gpu-pictured/



Some more information:

http://semiaccurate.com/forums/showp...8&postcount=26

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

[DanceswithUnix]

Beat me to it Cat

I suppose if they are doing this for Xbox, then it might make sense to do a consumer version. How heat much does a modern DDR chip push out, is it lost in the noise against a 100W CPU?

[CAT-THE-FIFTH]

Just noticed this slide:

http://www.techspot.com/news/48742-r...n-q3-2012.html