AMD Pulls in 32nm, Orochi, Llano and Ontario...

[flippin_waffles]

...starts 32nm production in 2010.

....Meyer, AMD's big boss, mentioned that these processors will be introduced earlier, likely the fourth quarter of 2010. This means AMD is closing in on Intel,....

http://www.hardware.info/en-US/news/...ction_in_2010/

[acidpython]

So we'll start seeing 32nm come 2011? I'm getting confused with AMD's roadmap at the moment from the bulldozer thread

[Sparky]

Nothing to be confused about. Just do what I do - ignore all roadmaps and hearsay and just play it by ear, you'll know when a CPU is getting released a few months in advance

[largon]

Whenever it comes, it better be sooner than later.
Hammer architecture is getting old after poking it around for the past 5½ years. Too bad it will be around for ~2 more years.
:\

[BulldogPO]

Phenom II is good but AMD needs now something that is significantly faster than competitors High-End products.

[AuDioFreaK39]

Intel 32nm Gulftown 6-core = Q1 2010

AMD 32nm Orochi 6 to 8-core = Q4 2010


There's still more than a half-year gap between them..

[gOJDO]

Whenever it comes, it better be sooner than later.
Hammer architecture is getting old after poking it around for the past 5½ years. Too bad it will be around for ~2 more years.
:\

Well, K8 lasted 4 years but it was improved over time: dual-core, DDR2, etc.

K10 is a new architecture, although it is reworked and improved K8. Then it got some improvements with the 45nm shrink, so you can't consider Deneb/Shanghai as a K8 architecture. In the next 2 years AMD needs to (and IMO will do) rework and improve K10.5 further. They should add more execution units, improve branching and code paralelisation, add features like memory disambiguation, improve cache latencies, increase L3 speed and add additional L3 ports, etc.

[saaya]

as soon as somebody talks about something thats more than a year away hes making guesses that can be anywhere in between from highly educated and informed ones to pure fiction and fantasy. its all about what they plan to do usually to get everybody prepared and gather feedback.

[terrace215]

They reference the Informationweek article from Feb 27, dated BEFORE the March 5 quote from AMD's Muzny to xbitlabs.

http://www.xbitlabs.com/news/cpu/dis...s_to_2011.html

Advanced Micro Devices on Thursday said that the company had reset plans for the launch if its new micro-architecture code-named Bulldozer and the first truly next-generation processors will only emerge in 2011. This slightly contradicts the optimism caused by comments of the company’s chief executive officer, who promised to “ramp up” the first chips produced at 32nm node in mid-2010.

hardware.info gets it exactly backwards. Tards.

Even in the article they cite, Meyer only promised volume production to START some time in Q4 2010, not products to be introduced then. Just terrible "journalism".

Needless to say, the title of this thread should be changed, or it should be completely deleted.

[largon]

Or, it could just be Mr. Muzny didn't get the memo from Mr. Meyer.

[informal]

Well, K8 lasted 4 years but it was improved over time: dual-core, DDR2, etc.

K10 is a new architecture, although it is reworked and improved K8. Then it got some improvements with the 45nm shrink, so you can't consider Deneb/Shanghai as a K8 architecture. In the next 2 years AMD needs to (and IMO will do) rework and improve K10.5 further. They should add more execution units, improve branching and code paralelisation, add features like memory disambiguation, improve cache latencies, increase L3 speed and add additional L3 ports, etc.

While you are mostly correct,AMD can't do some of the things you propose(while they can do some other things):
-"They should add more execution units" <-they can't do this due to time constraints and complexity of such a move
- "improve branching and code paralelisation" <- they can certainly improve the BP ,but "code parallelism" improvements are also tough to do IMO-maybe they can tweak the uncore part and achieve even better scaling from 1 to 4 cores(it's already very good,standing at 3.6-3.8x,but additional 3-5% is not out of the question).
-"add features like memory disambiguation" -similar feature is already there in K10,AMD calls it simply OOO load execution(improvements),but it's there.They can improve on it though.
-" improve cache latencies, increase L3 speed and add additional L3 ports" <-this can be done both on 45nm and 32nm node levels;1MB L2 can be a standard amount of L2 cache in 32nm shrink of Deneb while L3 can-in theory- be doubled or even tripled on 32nm node.

@teracce215

Hans De Vries summed it up in his post:

WAS reset a long time ago. Anounced at the November 2008 Analyst say.

Double confusion caused by Xbit labs. First they blunder by claiming that Bulldozer is near and coming in 2010:
http://www.xbitlabs.com/news/cpu/dis...s_in_2010.html

Instead of admitting their error they claim 3 days later that it's AMD which has rescheduled Bulldozer.....
http://www.xbitlabs.com/news/cpu/dis...s_to_2011.html

Both existing and new processor designs are being implemented on 32 nm


Regards, Hans

[flippin_waffles]

Here's some speculation from SI board.

Be clear! They are talking desktop samples of Bulldozer architected cores in late 2010. There is no update on the server version of a die design with Bulldozer architected cores. That could be sampling a lot sooner as they said early 2010 for 32nm CPUs with production in H2 2010. That server version is called Sandtiger and has 8 Bulldozer cores.

The big boost with Bulldozer was always some way to get two physical cores to speed up the execution of one physical thread which is the reverse of hyperthreading. The latest speculation is that that one core executes the most likely path at all branches and the other takes the alternate path at the first branch and then takes the likely path thereafter. If the first path turns out to be correct, the second core jumps to the first non retired branch and begins the same speculation on the less likely path again. If the less likely path is correct, the roles swap and the first core jumps to the first non retired branch of the second path and takes the less likely path. The big boost comes from the single cycle it takes to dupe the registers on the true path to the other core.

With a conditional branch in x86 taken on average of one every 5 instructions, the speed up can be substantial. Of course the best boost comes from nearly likely paths at a high percentage of branches. Given the long execution pipelines, that would boost IPC by two or more times on such "nasty" code. The other effect is that the pipelines can be longer, may be 50-100% longer. That would push the clock rate up by 20-40% and between the two, could make the new core be 50-100% quicker on the same code.

Or use the same clock rate, but far less power. Currently 2 speed bins are between 10 and 15% of clock. So 20-40% is 4 to 6 bins. 2 bins halves the needed power per core. Thus even with twice the cores needed, the overall power required for a given performance level could be 1/2 to 1/4th on clock alone. Add to that the IPC improvement and that could be that for the same performance, power could be reduced by an order of magnitude.

So you could either get 1.5 to 2 times the single thread performance at twice the power or 1x performance at 1/10th the power. In one case, you leap frog the AMD flagship way beyond Core i7 (120W TDP), you get flagship performance in a ultrathin mobile (12W TDP) or some combination in between. Anyway you look at that, that could turn the current situation on its ear. For Intel, that would be worse than the P4/RDRAM debacle.

http://siliconinvestor.advfn.com/rea...msgid=25475421

[Duh]

Nothing to be confused about. Just do what I do - ignore all roadmaps and hearsay and just play it by ear, you'll know when a CPU is getting released a few months in advance

well said.. I dont know why people round here ( I am not saying it is your case)get so frustrated and irritated when debating tech stuff ( particularly in this section). remember we are all part of XS and we are in the same boat

[Chad Boga]

Here's some speculation from SI board.



http://siliconinvestor.advfn.com/rea...msgid=25475421

It is not often that I get to see someone quote from Gerassic Park, bless you flippin', you have made my day.

[flippin_waffles]

Don't wet your pants, Chad. Like I said; 'speculation'.

[Chad Boga]

Don't wet your pants, Chad. Like I said; 'speculation'.

Yeah, Sharikou level speculation.

[Shintai]

The SI board post is simply retarded. Its written by a diehard fanboy for sure. And I see we got reverse hyperthreading in hardware back. Yet its not possible and any speculative execution requires all the work to be done in the compiler. I wouldnt be surprised if it was Sharikou himself that wrote it under another name.

[Eson]

It is not often that I get to see someone quote from Gerassic Park, bless you flippin', you have made my day.

I think its doable, why wouldn't it?

[Rammsteiner]

Intel 32nm Gulftown 6-core = Q1 2010

AMD 32nm Orochi 6 to 8-core = Q4 2010


There's still more than a half-year gap between them..

What does it matter? In the end it mostly depends on the real numbers of the CPU's, like TDP, power usage, performance etc. It's not like a compagny cant wait 6 months for an eventual better deal.

This all is too far away anyway, if some factory blows up we wont be seeing a chip for a very long time at all

[perkam]

Seeing how GPU and CPU roadmaps have been in the past year or so, only news of events reported to occur in the next 3 - 4 months can be given any credibility.

Beyond that, its all wishful thinking, unless the company starts showing final silicon test results or giving samples out before hand, like Intel.

Perkam

[terrace215]

Seeing how GPU and CPU roadmaps have been in the past year or so, only news of events reported to occur in the next 3 - 4 months can be given any credibility.

Beyond that, its all wishful thinking, unless the company starts showing final silicon test results or giving samples out before hand, like Intel.

Perkam

Not quite, it's more one-way, in that promises of "at this time" may be wrong, but they are almost always wrong in one direction: LATE, not early.

So if AMD says 2011 for Bulldozer, count on 2011 *or later*.

Your 3-4 month rule might be reasonable for a time-frame of when we have confidence that something will not be LATER than promised. But right from the get-go, we can have a high degree of confidence that products will not show up earlier than promised, as they almost never do. Particularly in AMD's case.

[ILikeCosmosS]

i would wait if this stuff is true, but sometimes waiting might be to long

(this stuff always discourages me from building my pc!)

[Nedjo]

what a misleading title, and the complete thread for that matter...

first of all, Dirk told that AMD will ramp-up 32nm in the middle of the next year! That doesn't translate in to a commercially available CPU in that time-frame... that means firing up 32nm equipment!

Second, that roadmap is the only piece of reliable info about future AMD processors, and according to that 32nm commercially available processors are coming in 2011.

Third... what's with this Bulldozer babbling? When was the last time that Bulldozer was officially mentioned or appeared in AMD roadmaps? Let's do some googling.....

here it is: year 2006, April: http://www.theinquirer.net/inquirer/...-of-the-day-is

3 years ago!!! Golden age of AMD... CTO was Phil Hester... Netburst was only thing that competition was offering! With passing of time, everything changed! Nothing is the same anymore, so why do people still think and write about Bulldozer, when AMD isn't mentioning that thing ever since 2007 tech days* is beyond me!?!

* - http://enthusiast.hardocp.com/articl...50aHVzaWFzdA==

[Sparky]

Intel 32nm Gulftown 6-core = Q1 2010

AMD 32nm Orochi 6 to 8-core = Q4 2010


There's still more than a half-year gap between them..

Eh, AMD tends to lag behind intel in process size, but that isn't necessarily a problem. Process size isn't everything. What matters more is architecture, IPC, power usage, etc..

[FischOderAal]

Eh, AMD tends to lag behind intel in process size, but that isn't necessarily a problem. Process size isn't everything. What matters more is architecture, IPC, power usage, etc..

But now the problem is Intel has both. The better (ie faster) architecture and smaller structures... Netburst just plain sucked in my books and that's why AMD could stick with 90 nm for so long without having problems.