AMD: 2006 & Beyond

blastarr

Power Member
Encontrei esta análise no techreport.com
Quite interesting...;)

What's next for AMD?
by Scott Wasson - 10:27 am, September 2, 2005

I had a brief meeting with AMD during IDF last week and learned just a few things. The dual-core version of the Turion mobile CPU is coming, and AMD has moved up the schedule to the first half of 2006 (from a general "2006" time frame previously). The dual-core Turion will be the same basic architecture as the Athlon 64 X2, with the same sort of manufacturing process optimizations used for current Turions, on the same 90nm fab process. Dual-core Turions will slide into a new S1 socket type, which has fewer than 754 pins. Also new will be support for DDR2 memory, although AMD isn't saying how many channels of DDR2 the CPU will support just yet. They did say that number of channels will be specific to mobile parts.
There was a demo of a dual-core Turion running inside of a closed mini-tower PC case, and Task Manager did indeed show two CPUs present. Beyond that, there's little to report. AMD won't divulge this new Turion's official brand name yet, and they have no comment on the possibility that this processor could have an integrated graphics processor on die. Mum was also the word about whether the new Turion's two cores would be able to scale clock speeds and voltages independently of one another, or whether one core could turn off when not needed.

So AMD wasn't saying too much, even about its near-term plans, which was a little bit surprising given the volume of beans spilled by Intel last week. If you look at its latest official roadmap, updated in March, AMD has already delivered on nearly every product listed, and there's nothing forecast beyond that. AMD's friendly reps did tell us that the company has no plans for major microarchitectural changes for 2006, and they reminded us how firmly and how long they'd been sitting in the proverbial catbird seat in terms of CPU performance. They also said that their 65nm fab process remains "on track," which is better than off the tracks, I suppose.

All of this non-disclosure, naturally, leads one to ask what AMD might be planning for the next year, and especially for the second half of '06, when Intel's new microarchitecture is expected to arrive. Given the likely clock-for-clock performance of Intel's new design in light of the already strong IPC of the Pentium M from which it's derived, it would seem that AMD might want to do something more in order to hold on to the performance lead.

So what are the options open to AMD absent a major microarchitectural change? We can certainly speculate, for what it's worth. We know that AMD's CTO, Fred Weber, told us a few months back that the move to multi-core processors means that chip advances will be more closely tied to manufacturing advances. AMD could very well launch a quad-core version of the Opteron at 65nm in order to keep its lead in the server space. Intel's first quad-core part, code-named Whitefield, isn't expected until 2007, and perhaps now not 'til late 2007. If AMD can start cranking out quad-core Opterons before then, it will have a good chance to keep the performance lead for server CPUs—and quite possibly the performance-per-watt lead, as well, given the happy math on multi-core performance versus power consumption.

As for the desktop, well, I couldn't help but notice that Intel's new architecture doesn't look like a good candidate for odd numbers of cores per chip. Conroe, Woodcrest, and company are dual-core parts with shared L2 cache, while Whitefield looks to be two Woodcrests crammed together with two slightly larger L2 caches (see the first slide here.) Each of Whitefield's two L2 caches look to be shared by two cores. With no sharing of L2 cache, AMD's multi-core architecture should be a much better fit for a three-core design. Could it be that, when Intel's dual-core Conroe arrives on the desktop with killer clock-for-clock performance, it will have to compete with a three-core Athlon 64? I don't know, but it's fun to play this game, eh?

Another likely change for AMD's multi-core CPUs (and Opteron in particular) is the addition of a large, shared L3 cache. Weber has mentioned this possibility as a solid option for future products, and I would be surprised not to see it happen in 2006. There's also the distinct possibility that AMD could unleash another round of relatively minor architectural changes that boost performance substantially, as the revision E core did. The expected move to DDR2 memory could pep things up, as well, though I doubt by much.

So AMD has options in the face of Intel's impressive roadmap renaissance. We'd just feel more confident about them if AMD's own roadmap weren't a black hole after the end of 2005.
 
Back
Topo