AMD talks Shanghai performance, features, roadmap to 2010

[Kage]

AMD's 45nm Shanghai processor has been a topic of interest for months, but precious little information was available beyond the chip's targeted time frame. But the company took the wraps off its server/workstation roadmap today, giving us a glimpse at the product ecology it intends to build with its 45nm process, as well as information on how Shanghai will evolve after its debut.

ars technica - AMD talks Shanghai performance, features, roadmap to 2010

[XJ.]

Wow, what a drag. It's looking like Shanghai may add about 10% in clock overhead and 10% more IPC for a total of a ~20% performance bump over Barcy. Shanghai needed to provide at least a 40% gain to stay in the game. AMD really blew it when designing the K10 architecture. As a former AMD user I weep, and I weep at realizing my wallet will pay the price of Intel's dominance.

[Sandog]

It seems AMD is falling behind more every chip Intel produces. With AMD throwing the 2011 date up there to be competitive there is no telling what performance Intel chips will have by then. AMD's troubles began when they drummed up fanboys and made fun of Intels P4/Prescott IMO. I guess I should roll back farther when Intel was shafting AMD but I was not biased back then and actually used a couple of AMD chips. Since they passed Intel with K8 they have went backwards. At least before K8 they competed closer performance wise. AMD needs to quit trying to beat Intel and just try to be competitive. They really should fire Ruiz who has continually blown it and get somebody in there who is smart and not bitter.

AMD is really crippling Intel. Intel is not working fast to get out the Penryn's and when they do I don't think I will be able to afford the chip I want. I might have to wait until Nehalem comes out for Penryn to drop in price but there is really no need to hurry and get Nehalem out. I guess the bright side of this for Intel is they can devote resources towards graphics and also they have more time to tweak and ready there 45nm CPU's for release so they will have less errata/bugs.

[k0NG0]

Yeah, they've screwed up big-time.

I'm still looking at a Phenom rig, since for anyone not gaming it's still a very solid platform especially with the 780G. But AMD will be playing second fiddle for a couple of years now. I do hope they manage to claw back some market share and get their financial situation in order. Everyone who likes PC hardware needs a powerful AMD.

[ctal]

A couple of things I'm wondering relative to the Magny-Cours:

I have no particular familiarity with how server software is evolving these days, but are 12 cores (no pun intended, although perhaps AMD intended one) really going to be useful to anyone just two years from now?

And is having only 12mB L3 cache (+ ? L2) really going to be competitive, especially when allocated among all those cores? The higher end 45nm Xeon's have 12mB L2 right now, serving only 4 cores.

-- Al

[tamarelle]

Originally Posted by Sandog

It seems AMD is falling behind more every chip Intel produces. With AMD throwing the 2011 date up there to be competitive there is no telling what performance Intel chips will have by then. AMD's troubles began when they drummed up fanboys and made fun of Intels P4/Prescott IMO. I guess I should roll back farther when Intel was shafting AMD but I was not biased back then and actually used a couple of AMD chips. Since they passed Intel with K8 they have went backwards. At least before K8 they competed closer performance wise. AMD needs to quit trying to beat Intel and just try to be competitive. They really should fire Ruiz who has continually blown it and get somebody in there who is smart and not bitter.

AMD is really crippling Intel. Intel is not working fast to get out the Penryn's and when they do I don't think I will be able to afford the chip I want. I might have to wait until Nehalem comes out for Penryn to drop in price but there is really no need to hurry and get Nehalem out. I guess the bright side of this for Intel is they can devote resources towards graphics and also they have more time to tweak and ready there 45nm CPU's for release so they will have less errata/bugs.

Hi Sand

I agree with what has been said here. The problem is that there is a huge market for the next generation architecture and Intel seems content to spoon it out and just rake in the $ €. It's so different from the hungry Intel which have given every enthusiast access to superlative top of the line processors. I have a Q6600 which is mind blowing and I paid less than 125€ for it. Kind of amazing that. But with AMD blowing it, Intel are clearly waiting. I won't be able to afford their top of the line next gen architecture either so maybe I will end up buying AMD shanghai. If its faster than anything but the top of the line Intel stuff and costs very little then most people will switch over rather than buy some extreme edition costing thousands.

[Sandog]

It seems the new Shanghai will make up for speed with real estate (more cache/more cores). This will put the TDP up there though IMO and these chips will be as hot as Barcelona under load probably even at 45nm.

On Intel pricing, I think we will see higher prices from Intel but I don't think it will be because they have the better product, I think it will be because Intel has to force the purchase of AMD chips to help AMD out. The Penryn's should be cheaper to produce than the 65nm parts.

[Sandog]

By the way, Intel is set to bring Nehalem out which has major changes but still based off the Core 2. I'm pretty sure you can count on %30 increase in performance with Nehalem and it will be more efficient than Penryn. I think Nehalem will be a big hit. Then comes the 8 core version at 32nm's. I don't see AMD competing well at all in the future CPU wise. I think they need to change the game with Fusion in order to win but it looks hopeless. Intel has time and money to start competing in graphics as well. I am pretty sure though that I will be buying the next ATI R770 GPU card (XT version).

[k0NG0]

Originally Posted by ctal

A couple of things I'm wondering relative to the Magny-Cours:

I have no particular familiarity with how server software is evolving these days, but are 12 cores (no pun intended, although perhaps AMD intended one) really going to be useful to anyone just two years from now?

And is having only 12mB L3 cache (+ ? L2) really going to be competitive, especially when allocated among all those cores? The higher end 45nm Xeon's have 12mB L2 right now, serving only 4 cores.

-- Al

The Penryn is still a bit lopsided due to the comparatively slow FSB. That narrow data pipe is offset by slapping on lots of L2 cache to make up for slow memory access. Nehalem will use much smaller L2 caches and a larger L3 shared among the cores, like Barcelona.

If you look at a block diagram of Nehalem, you will essentially see the same basic setup AMD has been using since the first dual-core K8 cores were released. For a quad core, it's eerily similar to a block diagram of Barcelona. (The functional units differ, of course, but it's still QPI~HT, shared L3, smaller L2, IMC etc. They're essentially siblings.)

AMD's problem isn't necessarily architectural in nature. They screwed up the 65nm transition, the Barcelona intro, they had the L3 bug. The only real architectural problem (and I'm not sure it's not a foundry issue) is they can't get speed up closer to 3GHz without heat output going all to hell (almost literally).

A great architecture means little if you can't fab chips at good speeds, in good volumes and on time. That's where they failed. Miserably, at that.

[ctal]

Excellent explanation, kONGO -- thanks!

Regards,
-- Al

[XJ.]

Originally Posted by k0NG0

AMD's problem isn't necessarily architectural in nature. They screwed up the 65nm transition, the Barcelona intro, they had the L3 bug. The only real architectural problem (and I'm not sure it's not a foundry issue) is they can't get speed up closer to 3GHz without heat output going all to hell (almost literally).

Good point, but I think there is an architectural problem in addition to an implementation problem since AMD couldn't speed up the 45nm chips. They should have been able to add stages to the pipeline or increase the fan out to either up the clock rate or IPC respectively. Neither of these things happened, which says to me that the current architecture has some serious limitations to it such that it impedes implementation improvements.

The other thing is that AMD's 45nm process doesn't move to High-K/metal gates so they don't get the rather large speed bump that Intel saw from making that change. AMD won't be improving their process till they go to 32nm (which they have stopped talking about ).

[k0NG0]

Originally Posted by XJ.

Good point, but I think there is an architectural problem in addition to an implementation problem since AMD couldn't speed up the 45nm chips. They should have been able to add stages to the pipeline or increase the fan out to either up the clock rate or IPC respectively. Neither of these things happened, which says to me that the current architecture has some serious limitations to it such that it impedes implementation improvements.

The other thing is that AMD's 45nm process doesn't move to High-K/metal gates so they don't get the rather large speed bump that Intel saw from making that change. AMD won't be improving their process till they go to 32nm (which they have stopped talking about ).

Which is why I was deliberately vague on that point. I'm not sure adding pipeline stages would do much, it would also add more heat, and I have a feeling they're hitting that limit sooner rather than later.

As for Shanghai and speeds, I've yet to see numbers and only heard rumors, and it's still 4-5 months before they hit the market, and that leaves time for tweaking. But it's clear they'll have jack squat to compete with Nehalem.

[ctal]

XJ -- Yes, it does strike me intuitively that besides all of their managerial and other non-esoteric problems, a good part of AMD's problems may simply be that Intel's high-K breakthrough was an exceptional accomplishment, that they just can't match. Last November I posted the following, which anyone with access to the IEEE Spectrum publication will find of interest:

Originally Posted by ctal

For those of you who may have access to the monthly IEEE Spectrum, either by being IEEE members or at a technical library, I highly recommend the following article in the October, 2007 issue:

Welcome to IEEE Xplore 2.0: The High-k Solution

In it, some senior Intel technical people discuss the development of the high-K gate structure being introduced in the Penryn, which represents probably the most major change in ic technology since the 1960's. After reading their descriptions of all the technical challenges they faced along the way, involving all sorts of quantum physics effects, I don't wonder that AMD (or anyone else) can't seem to keep up.

Very basically, several years ago (I suspect with either the 130nm Northwoods, or just after that with the 90nm Prescotts), the insulating layer between the gate and the source-to-drain channel of each transistor reached a thickness of 5 atoms (!!!!). Using traditional silicon-based materials, it could not be reduced any further to support smaller processes, because electrons would tend to tunnel through such a thin layer, this apparently being a major reason for the high leakage currents in Prescott.

The myriad technical challenges they overcame to work around this problem, using new materials, is an amazing story and an amazing accomplishment!

On another note, Tam -- it's nice to see you around here lately; hope all is going well for you!

Regards,
-- Al

[k0NG0]

Intel is kicking everyone's butt in terms of process technology.

45nm hasn't been a cakewalk, though. They have had problems, that's evident in the slow trickle of parts, but since AMD has been rearing centipedes just to have enough feet to shoot, it hasn't really mattered.

[ctal]

Here's a good summary of Hector's comments at AMD's recent stockholders' meeting:

X-bit labs - AMD: Financial Performance Should Not Depend on Products Performance.

I thought the most interesting sentence was:

We are re-architecting the business so that our financial success is not invariably dependent on continuous component performance leadership over our rich and dominant competitor.

He says that they are trying to re-structure the business so that they are profitable at current revenue levels. And, as might be expected, they will focus more exclusively on core businesses (presumably cpu's & graphics chips), and sell non-core and/or underperforming business units.

Regards,
-- Al