AMD's 2012 lines: Komodo, Trinity, and, Krishna/Wichita

[mAJORD]

Originally Posted by stmok

Its highly likely GlobalFoundries still doesn't have the capacity to meet AMD's needs (trying to use the same processes/procedures with 45nm and 32nm at the same time); and AMD is forced to lay out their release roadmap in a fashion that allows GF some breathing room in order to not cause another supply issue that has been experienced this year. (Remember the desktop Bulldozer delay from mid-2011 to October? Just so they can meet demand of OEMs by putting Llano on production priority.)

The more I think about it, the more I realise AMD spinning off its fabs as been the worst possible move they've made. Their original processor release roadmap for 2011 has been thrown into disarray. AMD and GF aren't acting in sync. They're dis-jointed. Its costing AMD and hurting their relationship. (Krishna/Wichita APUs were originally to be done by GF. Now they've been canceled in favour of a refresh of the existing Bobcat APU line...Made by TSMC.)

This is where Intel is winning, their coordination/cooperation with design and manufacturing branches. Its their consistency and everyone sticking to the same page that's resulting in the overall success of the company as a whole. (They're pretty aggressive given how this year they have demo'ed both their 2012 and 2013 mainstream processors at their last Developer Forum.)

...AMD needs to restructure itself in that manner. Make investments in key areas where they are lacking.

Well, 32nm has been out for a year now, and in that time we've seen next to no process improvements, so on the surface it doesn't look good.

I don't even know if GF are employing CTI still on the SOI process? If they are it must be purely focused on improving yields.

WE did see a similar situation on 65nm, so time will tell if we see better improvement over time on piledriver (Which should yield better)

[AEKaBeer]

Originally Posted by mAJORD

Well, 32nm has been out for a year now, and in that time we've seen next to no process improvements, so on the surface it doesn't look good.

Since PD is for both the desktop and notebook markets I guess they needed to focus on it's implementation over an interim refined BD, the biggest difference is going to be performance per watt over perf/clock. There will be an improvement (guessing 10-15%) with PD but we can't expect the world since they need to get steam roller out in 2013. It's a massive beat the clock game now, no more 8-9 years using the same architecture and AMD's engineers are getting their arses reamed after the bulldozer fiasco which frankly should have happened when phenom 1 came out all those years ago.

[stmok]

Some MSI-brand Socket FM2 mobos from Computex 2012.

MSI Shows off A85XA-G65 Socket FM2 Motherboard
=> http://www.techpowerup.com/167166/MS...therboard.html

ATX sized mobo using AMD's yet-to-be announced A85X chipset with two PCI-Express 2.0 x16 slots (running x8 each in SLI or Crossfire mode)...Yeah, you read right; This mobo supports Nvidia's SLI.

MSI Unveils Mini-ITX Socket FM2 Motherboard A85IA-E53
=> http://www.techpowerup.com/167168/MS...A85IA-E53.html

This uses the current AMD A75 chipset as found on Llano-oriented Socket FM1 mobos. Looks like they squeezed as much as they can into this little one.

MSI's Other Socket FM2 Motherboards Pictured
=> http://www.techpowerup.com/167169/MS...-Pictured.html

Cheaper models...
A85MA-E35 => m-ATX using current AMD A75 chipset. (Supports USB 3.0)
and
A55M-P33 (F2) => m-ATX using lower-end AMD A55 chipset. (No USB 3.0 support)


...Gigabyte brand ones as well.

GIGABYTE Also Shows Off its First Socket FM2 Motherboards
=> http://www.techpowerup.com/167116/GI...herboards.html

Three models on display; using AMD A85X and A55 chipsets for different price points. The top-end one has LucidLogix VirtuMVP.

[stmok]

AMD Trinity Desktop Chip Schedule Challenges Mobo Makers
=> http://www.tomshardware.com/news/amd...ard,15922.html

(1) Retail stock of desktop Trinity APUs isn't coming until Llano stock has been cleared. As such, enthusiasts looking to build their own Trinity-based set-up will not see it in June. The expectation is now in October.

(2) OEMs get priority of APU supplies. ie: Acer, Asus, HP, Lenovo, etc are getting Trinity chips now. The reason is because AMD wants to get OEMs ready for US's "Back-to-School" period. (They can't surrender this period to Intel.)

(3) This also means mobo makers are not happy, as they don't have sufficient retail chips to test their finalised board designs with. (You can't mass produce a mobo if you don't have a good supply of retail CPUs to make sure stuff works as intended! Think of all the mobo variants/models they have to test, and the number of models of APU they have to make sure each mobo version works with!)


The new AMD motto: Enthusiasts can wait.

...They did it with Bulldozer in 2011.
...They're doing it again with Trinity in 2012.

Yay!

[digamma]

Originally Posted by stmok

[b]
(3) This also means mobo makers are not happy, as they don't have sufficient retail chips to test their finalised board designs with. (You can't mass produce a mobo if you don't have a good supply of retail CPUs to make sure stuff works as intended! Think of all the mobo variants/models they have to test, and the number of models of APU they have to make sure each mobo version works with!)

Wouldn't most of the testing be done in software ie, on a computer, not physical? I would have thought that the design gets tested in software and then they build the thing to just make sure it works before sending it off to the fab.

The new AMD motto: Enthusiasts can wait.

...They did it with Bulldozer in 2011.
...They're doing it again with Trinity in 2012.

Yay!

As much as I hate to say it, in the money making scheme of things, enthusiasts don't contribute all that much to their quest for gold. The OEMs bring in the majority of their money, so it really would pay for them to keep their cashcows happy.

[mAJORD]

Originally Posted by digamma

Wouldn't most of the testing be done in software ie, on a computer, not physical? I would have thought that the design gets tested in software and then they build the thing to just make sure it works before sending it off to the fab.

They need a CPU to test bios functionality , Microcode, new VRM functionality

No design is ever put out without physical verification, that would be a huge risk

[digamma]

Originally Posted by mAJORD

They need a CPU to test bios functionality , Microcode, new VRM functionality

No design is ever put out without physical verification, that would be a huge risk

I realise that. But it's not really as though motherboards are making leaps and bounds in the technology. For the most part, it seems like most boards make tweaks between generations, rather than huge technology shifts. And then, most of them are utilising someone else's chipsets, and so are very limited in what their designs can be anyway. Of course they have to physically verify the design works before mass publication, but that is all it really would be, verification and test, not suck-it-and-see.

[loophole]

tom's hardware has a preview of the A10-5800K here:
AMD Trinity On The Desktop: A10, A8, And A6 Get Benchmarked!
(I haven't really read tom's hardware in years (since the 440BX was decimating the Rambus touting 820 chipset ), but it doesn't seem like a bad piece really)

They clearly know what the reader is after as they dive into some single-threaded benchmarks pitting Piledriver against 1st gen Bulldozer at the same clockspeed on page2

Short story for IPC vs Bulldozer: iTunes and 3ds Max show 15% improvement (this with the FX-8150's 8MB L3 cache left enabled).
Short story for IPC vs Stars: still not there yet but a little closer to parity (although clockspeed boost may go some way towards making up for the lost IPC), as long as you're not using FP-heavy codes.

[itsmydamnation]

Originally Posted by loophole

tom's hardware has a preview of the A10-5800K here:
AMD Trinity On The Desktop: A10, A8, And A6 Get Benchmarked!
(I haven't really read tom's hardware in years (since the 440BX was decimating the Rambus touting 820 chipset ), but it doesn't seem like a bad piece really)

They clearly know what the reader is after as they dive into some single-threaded benchmarks pitting Piledriver against 1st gen Bulldozer at the same clockspeed on page2

Short story for IPC vs Bulldozer: iTunes and 3ds Max show 15% improvement (this with the FX-8150's 8MB L3 cache left enabled).
Short story for IPC vs Stars: still not there yet but a little closer to parity (although clockspeed boost may go some way towards making up for the lost IPC), as long as you're not using FP-heavy codes.

FP heavy codes should smash vs stars for lots of reason and it generally does. FP is one thing bulldozer is very good at.


for example they obviously aren't using a FMA capable x264 with handbreak.

[mAJORD]

Impressive perf and perf/ watt over Llano.. finally.

Still no match for Intel, but at elas the BD architecture is finally showing an advantage over its predecessor in an apples to apples configuration.

Gaming perf is good. Prettymuch every game was "playable" at 1080p, perfect for the mainstream crowd..

[richan3185]

Originally Posted by mAJORD

Impressive perf and perf/ watt over Llano.. finally.

Still no match for Intel, but at elas the BD architecture is finally showing an advantage over its predecessor in an apples to apples configuration.

Gaming perf is good. Prettymuch every game was "playable" at 1080p, perfect for the mainstream crowd..

I agree...positive outcome.

AMD seem to have delivered +15% IPC over bulldozer. The FPU is very close to keeping up with the old 4 path stars core, with half the silicon resources.

A full 4m/8c piledriver chip might be an ok performer. It wont catch intel, but if the power and ipc improvements carry over...hey, looking better.

Also, with the ARM announcement today, confirming some people suspicions that there has been an ARM core in almost every APU, i think its very good news for AMD get alot of design wins soon, not least of all the all but confirmed PS4 win.

[loophole]

Originally Posted by itsmydamnation

FP heavy codes should smash vs stars for lots of reason and it generally does. FP is one thing bulldozer is very good at.


for example they obviously aren't using a FMA capable x264 with handbreak.

Agreed - for FP codes that make use of either of the FMA or AVX instruction sets Bulldozer and Trinity should be faster than the Stars cores but for those that don't (which is the majority of workloads out there at the moment) the sheer number of individual FPUs vs shared ones should put Stars ahead (especially Ph II X6 vs FX-8xxx).

But like anything else it'll just take time for developers to add support for these new instruction sets (or to move to compilers that emit these instructions).

But yes, overall Piledriver is a nice improvement to 1st gen Bulldozer. Can't wait to see how they scale clockspeed for the Vishera!

[stmok]

Originally Posted by loophole

tom's hardware has a preview of the A10-5800K here:
AMD Trinity On The Desktop: A10, A8, And A6 Get Benchmarked!
(I haven't really read tom's hardware in years (since the 440BX was decimating the Rambus touting 820 chipset ), but it doesn't seem like a bad piece really)

They clearly know what the reader is after as they dive into some single-threaded benchmarks pitting Piledriver against 1st gen Bulldozer at the same clockspeed on page2

Short story for IPC vs Bulldozer: iTunes and 3ds Max show 15% improvement (this with the FX-8150's 8MB L3 cache left enabled).
Short story for IPC vs Stars: still not there yet but a little closer to parity (although clockspeed boost may go some way towards making up for the lost IPC), as long as you're not using FP-heavy codes.

* Reads article *

Meh...It's confirmed what I've already anticipated for weeks now. They've made minor changes to offer incremental improvement. Nothing major to nibble at their rival's heels. Oh well, maybe in 2013.

[itsmydamnation]

Originally Posted by loophole

Agreed - for FP codes that make use of either of the FMA or AVX instruction sets Bulldozer and Trinity should be faster than the Stars cores but for those that don't (which is the majority of workloads out there at the moment) the sheer number of individual FPUs vs shared ones should put Stars ahead (especially Ph II X6 vs FX-8xxx).

But like anything else it'll just take time for developers to add support for these new instruction sets (or to move to compilers that emit these instructions).

But yes, overall Piledriver is a nice improvement to 1st gen Bulldozer. Can't wait to see how they scale clockspeed for the Vishera!

its not just the number of units, its the load store bus width to the FPU as well.

[Digit]

Any update on when we'll see Trinity hit these shores?

Looking at getting the A10 for a media centre, should take care of most basic games as well as HD Video recording/encoding/blu-rays right?