AMD Excavator on 28 nm, some production moving back to GF?

[bononos]

hope they bring fx back after excavator

If it comes back on the AM3+, it probably be one of the longest serving platforms, was the super7 longer?

And whats the deal with GF? Why did AMD break off in 2012?

 

[redzo]

The 1150 i7's are apus also! I see no need for them to bring the FX back. They could just as easily increase the core count + IPC and there you have it.

 

[Abwx]

@Abwx - I dont think so. why would they tape out on 28nm just for debugging when they can debug in an emulator which is easier, faster and cheaper
we dont know where GF is with their 20nm processes. I think it could be more for testing the process than the chip

There would be no point in releasing a 28nm 300mm2+ chip, assuming it will be significantly better than Kaveri power dissipation would increase notably, yet we know that it will be a 65W chip and that it is supposed to be way
better than current 65W offerings, otherwise why bother to released it
if it get the same perfs as the current line.?.

Only way this could be done on 28nm is drastic IPC improvement over Kaveri since they may as well have to reduce frequency to compensate for the larger transistors budget, from what we know of Steamroller there s not a lot that can be extracted from the available executions ressources, theses are already maxed out , prove is that SR two ALUs + AGUs have a throughput as efficient as a K10 wich has three units, same for the FP units but the decoders are wide enough since we know from Bulldozer that a single one can feed two cores quite efficently so more executions ressources can be attached to each decoders and fed well enough with the existing hardware so they have the ressource to add a ALU+AGU and get 20-25% better integer IPC out of the theorical 50%, on the FP side they can implement a modest but still worthy improvement that consist to not double the exe ressources and rather just add a 128 FMAC unit and also get 25% out of the 50% possible improvement on this front.

 

[Soulkeeper]

i'm waiting for a doubling of performance before I upgrade.

At AMD's 10% increase per year that will take 7.27yrs

 

[itsmydamnation]

If it comes back on the AM3+, it probably be one of the longest serving platforms, was the super7 longer?

And whats the deal with GF? Why did AMD break off in 2012?

they didn't. they made the small cat cores at TSMC. AMD has already confirmed that they will be moving PS4/XB1 production to GF thus taking a massive amount of the minimum waffer allocations.

How was your post received at the time? :biggrin:

full of your usually quality :thumbsup:

If people look at Karvei the CPU cores take very little of the die. you could enlarge the core quite a bit and SOC size would only modestly grow. maybe those die shots are real, 28nm EX http://i.imgur.com/HDjJSET.jpg

 

[ShintaiDK]

Anyone still in doubt that AMD is the new VIA?

And the GloFo deal runs to 2024. The worst processnode around till then.

Seriously man? Exactly just how much effort are you putting into coming up with these non-constructive posts?

We'll see you in a few days.

-ViRGE

 

[norseamd]

they didn't. they made the small cat cores at TSMC. AMD has already confirmed that they will be moving PS4/XB1 production to GF thus taking a massive amount of the minimum waffer allocations.

nice. hope they can make some chips at tsmc

 

[norseamd]

If it comes back on the AM3+, it probably be one of the longest serving platforms, was the super7 longer? And whats the deal with GF? Why did AMD break off in 2012?

maybe. might be able to put the fx on am2+ also.

not sure what you are talking about but gf is horrible from what i understand. they are not advancing the new process nodes very far

 

[ShintaiDK]

maybe. might be able to put the fx on am2+ also.

not sure what you are talking about but gf is horrible from what i understand. they are not advancing the new process nodes very far

FX is dead, it didnt sell. If you get any type of FX it will be a 2M/4T with IGP and just rebadged.

 

[Abwx]

FX is dead, it didnt sell

Give us some numbers , because 33% of AMDs production
is CPUs with APU being 67%, it just say that you are throwing
assumptions while being cluless about real numbers , anyway
if you could stop this kind of thread crapping it would be a great
plus for the forum , it is years that you constantly feed us with
your VIA irrelevant analogy...

 

[mrmt]

maybe. might be able to put the fx on am2+ also.

It's possible that you'll see FX chips again, but these might not be pure CPU processors anymore. FX chips were fused off Opteron chips, and that market is closed for AMD.

 

[USER8000]

It's an APU. FX is dead.

I know its an APU.

But it says SOC.

That means integration of the chipset into the CPU.

Kaveri is not an SOC.

Kabini and Temash are SOCs.

 

[Fjodor2001]

The ExtremeTech article in the OP says:

It’s not clear, at this point, if Kaveri and Carrizo are built on two fundamentally different types of 28nm silicon, or if the different codenames reflect subtle changes. If the 65W target for Carrizo is accurate, it’s possible that AMD is moving to a different node that emphasizes lower power and higher efficiency.

So maybe it's not so bad after all if Carrizo actually is on 28 nm instead of 20 nm?

Node shrinks seems to bring less advantages nowadays anyway, apart from the transistor density improvements. And one process at a certain node width is not directly comparable to another at the same node width. They can differ in terms of max frequency, actual transistor density, power consumption, price, and so on for the actual chips.

 

[Homeles]

Node shrinks are only less appealing at the high end of the frequency spectrum. Mobile devices benefit tremendously.

 

[norseamd]

FX chips were fused off Opteron chips, and that market is closed for AMD.

how so

 

[NTMBK]

hope they bring fx back after excavator

If they want to bring back FX, they need a CPU (not APU) which is competitive with Intel on performance/W- in the server market AMD got absolutely slaughtered. They have nothing which can compete with a 12-core Ivy Bridge EP in either perf/W or perf/rack unit, and until they have a competitive core design and fab process they won't bring out a new Opteron. Which means no FX.

 

[NTMBK]

maybe. might be able to put the fx on am2+ also.

not sure what you are talking about but gf is horrible from what i understand. they are not advancing the new process nodes very far

Even the Bulldozer parts won't work on AM2+- they don't have DDR2 memory controllers. No way that a theoretical Excavator chip would waste die area on a DDR2 memory controller.

 

[NTMBK]

I know its an APU.

But it says SOC.

That means integration of the chipset into the CPU.

Kaveri is not an SOC.

Kabini and Temash are SOCs.

Carrizo is confirmed to use the same FM2+ socket as Kaveri, meaning that it will also have a discrete chipset. The FM2+ socket doesn't have the pins for the southbridge IO.

 

[USER8000]

Carrizo is confirmed to use the same FM2+ socket as Kaveri, meaning that it will also have a discrete chipset. The FM2+ socket doesn't have the pins for the southbridge IO.

Unless for mobile they are making an SOC version??

It seems unusual for it to be called SOC on a profile,which people use for job hunting. Such a crass error would reflect badly on them.

Edit to post.

I just looked back at the roadmap from that video:

cdn4.wccftech.com/wp-content/uploads/2013/12/AMD-Carrizo-APU.jpg

The server version of Carrizo is called Toronto - I forgot that!

Both versions of Toronto(CPU and APU) seem to have SOC in the description,and it is an AMD roadmap.

Plus we know it has both a DDR3 and DDR4 memory controller.

The Kaveri based Opterons mention the Bolton SCH and the FX CPUs mention the socket too.

However,the Toronto CPUs don't mention the Bolton SCH and instead have it replaced by SOC.

 

[mrmt]

how so

AMD server offers where wiped out because Intel can offer both better IPC, larger thread count per die and more power efficiency. When Bulldozer was launched AMD couldn't compete with 8-cores SNB-EP chips, now AMD would have to beat 12-cores IVB-EP (14-core HSW-EP in Q314) with roughly the same TDP.

AMD's CPU makes no sense whatsoever, only if you need their IGP you should look for them.

 

[podspi]

Yep. FX will be back when AMD has something it can float in the server market.

Although it is debatable whether they will call it FX. Such a shame, they completely blew all the brand equity.

 

[Fjodor2001]

So what will be beyond Excavator, is anything known about that? Will it be a completely new uArch which is non-Bulldozer based?

 

[norseamd]

Even the Bulldozer parts won't work on AM2+- they don't have DDR2 memory controllers. No way that a theoretical Excavator chip would waste die area on a DDR2 memory controller.

meant fm2+. sorry

 

[norseamd]

So what will be beyond Excavator, is anything known about that? Will it be a completely new uArch which is non-Bulldozer based?

this is what i am waiting for

 

[sm625]

Well that doesn't really bode well for Carrizo bringing much to the table.

Why do you say that? Process shrinks havent really helped AMD much lately anyway. A 28nm kaveri is basically tied with 45nm lynnfield in terms of cpu performance per watt. The core design is so bad that even at 14nm it still couldnt compete with a haswell strictly in terms of performance per watt.