AMD - Piledriver chitchat

[CAT-THE-FIFTH]

It seems there were 45W TDP and 65W TDP AM3+ Bulldozer Opterons:

http://www.cpu-world.com/CPUs/Bulldo...50HOW4MGU.html

Supposedly,AMD is releasing the Opteron 3300 series next month:

http://www.cpu-world.com/news_2012/2...rocessors.html

Its a shame AMD has not released even one of these CPUs for review.

[Terbinator]

The one in the first link isn't Vishera based though? How odd.

[CAT-THE-FIFTH]

They are Opteron 3200 are Bulldozer based CPUs,so the Vishera ones(the Opteron 3300) will probably run at least 200MHZ higher, plus will have better IPC and improved realworld power consumption.

They are AM3+ so you would think AMD would at least send some out for review,but like with the 95W TDP FX8120 and 95W TDP FX8320,AMD will probably not send them out,let alone make them easily available. Look at the A10-5700 for instance,its almost like they don't want it reviewed and that is easily available. I never understood their reticence with sending their lower TDP models out for review?? Its almost like their marketing department does not actually bother to read reviews at times!!

The marketing department of AMD is full of very weird people.

[Terbinator]

The marketing department of AMD is full of very weird people.

Assuming they haven't all been let go yet.

[CAT-THE-FIFTH]

Assuming they haven't all been let go yet.

They seemed a bit more proactive with SA though,so perhaps they have got some better people recently??

IMHO,it is an uphill struggle against Intel,but still they are probably shooting themselves in the foot more than they need to!!

[CAT-THE-FIFTH]

This is interesting:

http://forums.anandtech.com/showthread.php?t=2283742

http://forums.anandtech.com/showpost...1&postcount=32

Skyler is Jason Garrett-Glaser who is one of the lead devs behind x264 and that is from the x264 dev channel.

There is noise of AMD licensing Freedom Fabric(groan) to other ARM licensees too.

BTW,please note that Anandtech forums is probably the Intel equivalent of AMDZone at times!!

Edit!!

Maybe not SR:

http://forums.anandtech.com/showpost...&postcount=106

Could it be Richland??

I wonder if it is PD+GCN??

[watercooled]

Could be either really; as they said I think CPU-Z refers to an internal database to list some CPU specs, but not sure if that applies to cache too. It's probably more likely to be Richland though. I wonder how it differs from Trinity? I agree it could be PD+GCN, but I wonder if they're implementing unified memory for instance? Maybe they're planning on a die shrink, a sort of tick-tock thing? It would help with the higher peak power consumption a lot of people are obsessing with, and would help to mature the process for SR. New uarch and process node in one go is always a risk.

Also, I wonder what socket motherboard he ran that CPU on?

Edit: JF-AMD is mentioned a few times in that thread, what did he supposedly say?

[CAT-THE-FIFTH]

Look at the first comment:

http://wccftech.com/amd-richland-apu...ment-716422150

Kabini is the sucessor to Brazos. If the scores are true,then in CB R10 the CPU is around E5200 level. The single CPU score is around 31% higher than a E350:

http://www.anandtech.com/show/4134/t...-for-miniitx/8

[Terbinator]

So Core 2 Duo permanence in an Atom TDP finally?

[watercooled]

Wow, other guy in comments needs help! Wouldn't happen to be the name of the guy who runs the-site-that-shall-not-be-named would it? Same incomprehensibly-stupid, angry-schoolboy writing style at least...

So, even the desktop parts are a sort of SoC, so no separate PCH/SB? Could mean good things for SFF systems!

[CAT-THE-FIFTH]

AMD Kabini is an SOC.

So Core 2 Duo permanence in an Atom TDP finally?

At least for MT performance.

[CAT-THE-FIFTH]

http://forums.anandtech.com/showpost...&postcount=139

[scaryjim]

If that's true, AMD have put desktop APU performance into a laptop TDP?

That seems unlikely unless they've added some secret special sauce ... wonder if the whole "Trinity has Resonant Clock Mesh" was a blind? Or perhaps GCN is massively more power efficient than VLIW4 in APU mode and that's freed up enough TDP to ramp the CPU cores up? Something odd's going on here...

[watercooled]

It's also meant to be on 28nm, but I wouldn't expect that much of a jump over 32nm SOI.

[scaryjim]

I wish AMD had done a lower rung of VLIW4 cards, it's really hard to try to work out what, if any, power efficiency improvements GCN brought to the table. Obviously VLIW4 is much better than VLIW5, which is why they were able to ramp up to 800MHz GPU core in Trinity. Also you can fit 512 GCN cores into a 55W TDP (7750), the best VLIW5 managed was 480 cores into 66W (6670). Then again, the only GCN cards we've got benefit from 28nm production, so the comparison of, say, 2048 cores @ 925MHz vs 1536 cores @ 880MHz (7970 v 6970, both have a 250W TDP) is meaningless: the increase in specs could just be down to the die shrink.

OTOH: we know that GCN @ 28nm can run a 512 shader GPU plus all the additional gubbins that a discrete GPU needs in a 55W TDP. Now, when nvidia dropped the 9800GT cloicks to make a green edition, the GPU only pulled 40W of the 75W TDP (see http://www.inno3d.com/products/graph...00gt_esave.htm). Assuming roughly the same proportions for the 7750 would mean the GPU only pulls 30W. Assuming a 386 shader GPU would scale linearly that'd only be 22.5W, and that still contains a few extras that the APU slice wouldn't need - its own GDDR5 memory controller, for instance. They could easily be running the GPU section with less than 20W of the TDP.

The other thing is this will be AMDs first x86 cores on 28nm. What if they've found they can clock up much higher on the bulk process than they initially expected? If they could bump the clock speed of the cores to 3GHz at 1.1v, that'd be a huge power saving compared to the 1.4+v they required on 32nm.

There's all sorts of variabled to consider, really. But if that report is accurate - that is to say 3.2GHz in a 35W envelope - it bodes very well for AMDs future 28nm APUs and CPUs - if they can get the roadmap back on target...


EDIT: actually, if you compare VLIW5 to VLIW4, both on 40nm (5870 v 6970) VLIW4 doesn't look very energy efficient. with a similar number of shaders (1600 v 1536) and at similar clock speeds (850MHz v 880MHz) the 5870 has a 20%+ lower TDP than the 6970. VLIW4 brought more performance there, but at the cost of more energy usage. Looking at Trinity, I'm pretty sure the GPU is the limiting factor, TDP-wise: e.g. the GPU clock gets cut much harder on the ULV mobile chips than the CPU clock (~28% v ~14%). If the move to 28nm and GCN can cut the GPU TDP back hard, it would give the CPU a lot more room to breathe...

[CAT-THE-FIFTH]

The major reason AMD went for VLIW4 over VLIW5 in the HD6900 series,is because apparently it took less space up:

http://www.anandtech.com/show/4061/a...deon-hd-6950/4

So perhaps,for Trinity AMD knew they needed to keep die sizes in check,so went with VLIW4 instead. Hence they could raise IGP peformance,and since the PD cores probably were more efficient too,it masked the inefficiencies of the IGP.