Interesting AMD roadmaps

[kalniel]

Anand has two articles that are finally giving some more information on what AMD plans for the next few years.

First up, a general roadmap:
http://www.anandtech.com/cpuchipsets...oc.aspx?i=3673

Secondly, high level bulldozer and bobcat architecture diagrams:
http://www.anandtech.com/cpuchipsets...oc.aspx?i=3674

The latter is very interesting - we've been waiting for bulldozer details for ages - AMD needs this chip to be capable of beating Intel if it's going to keep/recover its CPU reputation. And the solution is interesting - rather than having 'in pipe' hyperthreading, they've doubled the integer processing units in the chip. So basically one bulldozer core will present itself as two CPUs to the OS, however FP performance will be halved. Oh noes! I hear.. especially when AMD has relied on FP advantage in the past vs Intel, but they've done it like that knowing that the APU will be providing more than enough FP performance - in a highly parrallel fashion as well.

So yes, we're going to not only have to wait for Bulldozer, but also decent APUs to sit alongside it. However it does show that fusion is going to be a nicely integrated product that should really take advantage of the best features of both CPUs and GPUs.

I'm just not sure I can wait that long!

[Domestic_Ginger]

Bulldozer seems to be using AM3. Go on treat yourself!

Not sure if motherboard support will extend that long though.....

[kalniel]

There was some talk of it using AM3rev2, which might be some adjustment that makes some boards compatible but not others. And can you envision AMD managing to cram in an APU with the same pin count?

Problem is if I'm going to buy in the next few months it'd be a lynnfield system without doubt. If thuban had any performance equality with anything Intel offered then maybe I'd consider it, but I just can't see that happening without some kind of horrible trade off in the power/heat department.

[scaryjim]

There was some talk of it using AM3rev2, which might be some adjustment that makes some boards compatible but not others.

AM3+ anyone? I agree it's hard to see how you manipulate a 938-pin package to manage all the connections a fully integrated processor would require though, unless they're doing something freaky with the existing HyperTransport bus to channel video output / PCIe lanes...

EDIT: Oooh, sudden thought - Quad-pumped HT? It's like Intel and AMD have gone back to 2002 and swapped technologies!!!

[CAT-THE-FIFTH]

If Fusion is using 32NM Phenom II cores it would be good if 32NM Phenoms do come out some time next year TBH. IMHO,AMD will find it increasingly difficult to compete with Intel if in 12 months time they are still using a 45NM process. If they they can find some way of going way past the 4GHZ mark with this process with an acceptable TDP it may help but with Thuban being a bigger processor they may have issues here.

It is a good thing that AMD has ATI still.

At this rate my next build may still be Intel based!!

[CAT-THE-FIFTH]

At least MadduckUK has got the latest C3 stepping 965 to 4GHZ:

http://forums.hexus.net/hexus-hardwa...st-review.html

[kalniel]

If Fusion is using 32NM Phenom II cores it would be good if 32NM Phenoms do come out some time next year TBH. IMHO,AMD will find it increasingly difficult to compete with Intel if in 12 months time they are still using a 45NM process.

They're nearly always a generation behind the die process, so I think next year is going to remain 45nm.

So no direct competition, but if they show enough good things about bulldozer some of us will be tempted to try Thuban next year with the new chipset if that chipset will be 100% compatible with bulldozer.

Intel have kindly left them a gap in the market too (lynnfield's limited roadmap). 'All' AMD have to do is outcompete lynnfield performance/price (while offering a future upgrade path which lynnfield is missing). Leave bloomfield/gulftown as extreme cost monopolies.

[Pob255]

You may find something like current AM3 will not be able to use onchip gpu and will have to rely on motherboard or descreat card for graphics.

The way these are going I really don't see onchip gou as a big issue, nether intel or amd are going to route of adaptability, and just truning spear cpu capasity into gpu, they both seem to be going the route of turning 1 or more cores into a gpu.
I'm failing to see how this will be any major advantage compaired to having the gpu in the chipset or seperate card.

[kalniel]

You may find something like current AM3 will not be able to use onchip gpu and will have to rely on motherboard or descreat card for graphics.

The way these are going I really don't see onchip gou as a big issue, nether intel or amd are going to route of adaptability, and just truning spear cpu capasity into gpu, they both seem to be going the route of turning 1 or more cores into a gpu.
I'm failing to see how this will be any major advantage compaired to having the gpu in the chipset or seperate card.

Think FP. They're effectively cutting down the the 'CPU' FP processing and relying on the 'GPU' to do that instead. No problem if the GPU is on die, but if it's two bridges away then you're going to take a massive latency hit.

[CAT-THE-FIFTH]

Looks like AMD has just got a $1.25 billion present from Intel:

http://forums.hexus.net/hexus-channe...-disputes.html

Much need cash for the company.

[Pob255]

Think FP. They're effectively cutting down the the 'CPU' FP processing and relying on the 'GPU' to do that instead. No problem if the GPU is on die, but if it's two bridges away then you're going to take a massive latency hit.

True
It will still require a new motherboard and chipset as this is a radical change in hardware.

And I'm still not sure I see the real benefit to this, any thing that is really heavly FP intensive then the onchip is just not going to have enough streams to make it worth it.
As an onboard graphics solution you just don't need massive bucket loads of processing.
Current top end onboard AMD graphics are more than enough for what's generally demanded of them.

The whole situation seems to be that both intel and amd have developed cpu's that are so powerful that extra processing is just going to waste so they are tring to find some other use for it.
But maybe that's just my cynical side talking.

[scaryjim]

any thing that is really heavly FP intensive[/b] then the onchip is just not going to have enough streams to make it worth it.

With the DX11 compute shader now available I *think* AMD are assuming (hoping?) that anything very heavily FP intensive will be coded for GPGPU (either by OpenCL or DX11) to take advantage of the huge processing power available in modern GPUs. The GPU on APU dies would effectively be an FP co-processor, giving much higher FP throughput than is currently available on x86 CPUs (considering the 4350 has 80 stream processors) without the need to program specifically for stream processing...

[Pob255]

Ahhh that makes sense.

CPU based physics engines would become far more viable, heck crysis might actually become playable

[kalniel]

Isn't that the same as what I said?

Doesn't need to be coded for it specifically - the scheduler will do that automatically. Just as FP operations are sent to the FP pipeline instead of the Int pipeline currently.