Intel On the Offensive: Roadmap Details and Analysis
by Kristopher Kubicki & Jarred Walton on June 14, 2005 4:33 PM EST- Posted in
- CPUs
Anand and I had a discussion about a year ago that went something along the lines of "2005 is going to be sooooo boring for Intel and AMD". But then a week later, AMD and Intel both started announcing dual core stuff internally and 2005 has been one of the better processor slugfests that we have seen yet. After passing up the last few roadmaps due to various tradeshows and NDAs, we decided to plunge into the most recent Intel roadmaps and really take an in-depth look at what is coming up for the next year or so.
We have talked extensively about Yonah - the dual core successor to Dothan - without much attention to its clock speed and other features. The latest roadmap has provided quite a bit of new information about several of the cores. We'll take a more in-depth look at Yonah, Presler and Cedar Mill in this update.
Aside from the additional cache, also note that the new 9xx series processors are the same 65nm Presler cores mentioned earlier. Presler will in fact be nothing but two Cedar Mill cores sharing the same package, just as Smithfield is only two Prescotts sharing the same core. Shrinking to 65nm frees up a considerable amount of space on the die, so moving to 2MB L2 cache seems logical. Also, all 9xx processors will have EIST enabled, whereas only the 840 and 830 Pentium D processors have EIST. That makes sense for 8xx, as EIST currently just drops the CPU speed to 2.8 GHz; we would hope that the 9xx models will improve the functionality of EIST. EM64T and XD are supported on all upcoming processors. The last final addition to the 9xx line is VT, or Vanderpool Technology.
We have discussed Vanderpool in the past (first in 2003), but it seems like there is still a large amount of confusion regarding the technology. Vanderpool - and AMD's competing technology, "Pacifica" - enable a CPU to run multiple operating systems on a single CPU at the same time. The particular demos that we have seen show four operating systems running on a single machine independent of each other. Unfortunately, this requires BIOS support, Chipset support and Processor support, not to mention OS support. Intel claims that VT will roll out on the 6xx processor line before the end of the year, so we expect 945 and 955 to fully support VT as of now, since the next chipset revision from Intel won't come until Q2'06 (Broadwater).
With 13 new performance SKUs in the next year, Intel certainly has its work cut out. The new 5x1 processors are set to launch before the end of this month, although the only real advantage that they have over the existing 5x0 and 5x0J chips is EM64T support. The interesting processors (highlighted in bold) are the upcoming 65nm Cedar Mill cores. Roadmaps hint that Cedar Mill will top out at 95W TDP per core and 3.80GHz is the highest clock on the roadmap. The Pentium 4 551 has a TDP of 84W. The roadmap also reveals that the prices of the 6xx line will cut dramatically mid-August, and VT enabled Prescott 2M (Pentium 4 672 and Pentium 4 662) will retain a premium over their non-VT counterparts.
As we mentioned earlier, two Cedar Mills running at lower clock speeds will compose a Presler dual core processor (Smithfield's replacement). HyperThreading, EIST, XD and EM64T are enabled on all of these new processors. Furthermore, all Cedar Mill chips will also receive VT technology, with the exception of the Pentium 4 631. The Pentium 4 631 and 633 are identical except the absence of VT in the 631. We're not entirely sure why the 631 is even being produced, other than to perhaps fill a niche market. All of the Cedar Mill cores clearly have VT support (as do the Prescott 2M cores), so why Intel would want to deactivate it in one model is anyone's guess.
With both 65nm Presler and 65nm Cedar Mill, the interesting thing to note is how low the FSB clock speeds remain. We had almost expected next generation Pentium processors to ramp up to 1066FSB after the most recent Pentium 4 EE utilized 1066FSB. In fact, Intel relaunched its 925X motherboard lineup (with 925XE) to support the faster bus, and all current generation motherboards make a big deal of supporting 1066FSB. As our tests revealed, 1066FSB did virtually nothing to improve performance on the 3.46EE over 800FSB. The conservative FSB speeds on Presler and Cedar Mill indicate to us that Intel probably did nothing to shorten Prescott's pipeline, although there may be other reasons why they have decided to keep the bus speed low as well.
We have talked extensively about Yonah - the dual core successor to Dothan - without much attention to its clock speed and other features. The latest roadmap has provided quite a bit of new information about several of the cores. We'll take a more in-depth look at Yonah, Presler and Cedar Mill in this update.
Desktop
Over the last month we were bombarded by dual core announcements and releases from AMD and Intel. Not only because dual core is an interesting and new concept to desktop computing, but also because after all of the hype, AMD and Intel were both able to deliver fairly good products (both AMD [RTPE: AMD Opteron Italy] and Intel [RTPE: "Pentium D"] are shipping already). Inevitably, there have been some pretty major changes to Intel's product naming, for starters:- Pentium 4 Extreme Edition processors are now simply named Pentium Extreme Edition. Although these processors are nothing more than best of breed server and desktop cores, they will now have their own product category at least. (This started with the most recent dual core Pentium XE.)
- Pentium 4 now only refers to the single core Prescott and upcoming 65nm Cedar Mill cores. Pentium D will refer to Smithfield and 65nm Presler.
- All 5x1 processors are now 64-bit enabled.
- All 6x2 processors are now VT enabled.
- All 6x3 processors are 65nm Cedar Mill (which has VT enabled).
Intel Dual Core Performance Desktop Lineup LGA775 |
||||
Processor |
Speed |
L2 Cache |
FSB |
Launch |
Pentium D 950 |
3.40GHz |
2x2MB |
800MHz |
Q1'06 |
Pentium D 940 |
3.20GHz |
2x2MB |
800MHz |
Q1'06 |
Pentium D 930 |
3.00GHz |
2x2MB |
800MHz |
Q1'06 |
Pentium D 920 |
2.80GHz |
2x2MB |
800MHz |
Q1'06 |
Pentium D 840 |
3.20GHz |
2x1MB |
800MHz |
Now |
Pentium D 830 |
3.00GHz |
2x1MB |
800MHz |
Now |
Pentium D 820 |
2.80GHz |
2x1MB |
800MHz |
Now |
Aside from the additional cache, also note that the new 9xx series processors are the same 65nm Presler cores mentioned earlier. Presler will in fact be nothing but two Cedar Mill cores sharing the same package, just as Smithfield is only two Prescotts sharing the same core. Shrinking to 65nm frees up a considerable amount of space on the die, so moving to 2MB L2 cache seems logical. Also, all 9xx processors will have EIST enabled, whereas only the 840 and 830 Pentium D processors have EIST. That makes sense for 8xx, as EIST currently just drops the CPU speed to 2.8 GHz; we would hope that the 9xx models will improve the functionality of EIST. EM64T and XD are supported on all upcoming processors. The last final addition to the 9xx line is VT, or Vanderpool Technology.
We have discussed Vanderpool in the past (first in 2003), but it seems like there is still a large amount of confusion regarding the technology. Vanderpool - and AMD's competing technology, "Pacifica" - enable a CPU to run multiple operating systems on a single CPU at the same time. The particular demos that we have seen show four operating systems running on a single machine independent of each other. Unfortunately, this requires BIOS support, Chipset support and Processor support, not to mention OS support. Intel claims that VT will roll out on the 6xx processor line before the end of the year, so we expect 945 and 955 to fully support VT as of now, since the next chipset revision from Intel won't come until Q2'06 (Broadwater).
Intel Single Core Performance Desktop Lineup LGA775 |
||||
Processor |
Speed |
L2 Cache |
FSB |
Launch |
Pentium 4 672 |
3.80GHz |
2MB |
800MHz |
Q4'05 |
Pentium 4 663 |
3.60GHz |
2MB |
800MHz |
Q1'06 |
Pentium 4 662 |
3.60GHz |
2MB |
800MHz |
Q4'05 |
Pentium 4 653 |
3.40GHz |
2MB |
800MHz |
Q1'06 |
Pentium 4 643 |
3.20GHz |
2MB |
800MHz |
Q1'06 |
Pentium 4 633 |
3.00GHz |
2MB |
800MHz |
Q1'06 |
Pentium 4 631 |
3.00GHz |
2MB |
800MHz |
Q1'06 |
Pentium 4 571 |
3.80GHz |
1MB |
800MHz |
Soon |
Pentium 4 561 |
3.60GHz |
1MB |
800MHz |
Soon |
Pentium 4 551 |
3.40GHz |
1MB |
800MHz |
Soon |
Pentium 4 541 |
3.20GHz |
1MB |
800MHz |
Soon |
Pentium 4 531 |
3.00GHz |
1MB |
800MHz |
Soon |
Pentium 4 521 |
2.80GHz |
1MB |
800MHz |
Soon |
With 13 new performance SKUs in the next year, Intel certainly has its work cut out. The new 5x1 processors are set to launch before the end of this month, although the only real advantage that they have over the existing 5x0 and 5x0J chips is EM64T support. The interesting processors (highlighted in bold) are the upcoming 65nm Cedar Mill cores. Roadmaps hint that Cedar Mill will top out at 95W TDP per core and 3.80GHz is the highest clock on the roadmap. The Pentium 4 551 has a TDP of 84W. The roadmap also reveals that the prices of the 6xx line will cut dramatically mid-August, and VT enabled Prescott 2M (Pentium 4 672 and Pentium 4 662) will retain a premium over their non-VT counterparts.
As we mentioned earlier, two Cedar Mills running at lower clock speeds will compose a Presler dual core processor (Smithfield's replacement). HyperThreading, EIST, XD and EM64T are enabled on all of these new processors. Furthermore, all Cedar Mill chips will also receive VT technology, with the exception of the Pentium 4 631. The Pentium 4 631 and 633 are identical except the absence of VT in the 631. We're not entirely sure why the 631 is even being produced, other than to perhaps fill a niche market. All of the Cedar Mill cores clearly have VT support (as do the Prescott 2M cores), so why Intel would want to deactivate it in one model is anyone's guess.
With both 65nm Presler and 65nm Cedar Mill, the interesting thing to note is how low the FSB clock speeds remain. We had almost expected next generation Pentium processors to ramp up to 1066FSB after the most recent Pentium 4 EE utilized 1066FSB. In fact, Intel relaunched its 925X motherboard lineup (with 925XE) to support the faster bus, and all current generation motherboards make a big deal of supporting 1066FSB. As our tests revealed, 1066FSB did virtually nothing to improve performance on the 3.46EE over 800FSB. The conservative FSB speeds on Presler and Cedar Mill indicate to us that Intel probably did nothing to shorten Prescott's pipeline, although there may be other reasons why they have decided to keep the bus speed low as well.
Facebook
Twitter
RSS
57 Comments
View All Comments
cssmicro - Friday, July 22, 2005 - link
50 - I don't think you realize how cost-inneffective it is to create a completely new process flow or sort flow just to turn off a feature. Creating a different chip costs many hundreds of thousands of dollars PER LAYER that's changed. The mask changes alone would cost about half of a million dollars per reticle. Forget about the engineering initiative necessary to design the changes. It would be retarded of Intel to make a move like that. 45 hit the nail right on the head. The same products will be made, and then filtered out by working/non-working parts at end of line sorting.You don't really think your graphics / non-graphics chipsets are MADE differently, do you? (PSST, they're not). Simply put, they're all tested at the end of production. Some have working graphics sub-systems, some don't. They're distributed accordingly.
cornflake - Monday, June 20, 2005 - link
I'm not sure how much Intel paid AnandTech for this article, but it must be a grundle to shovel it like this..."As a dual core solution, Yonah is the most advanced (other than perhaps Itanium 2 Monticeto) solution that we have seen out of either AMD or Intel."
Wow, so they have some cool new integrated memory controller in place to leap frog AMD. NO, well then what about the new Intel branded Hyper-transport knock off they are working on? Oh well, at least we can look forward to some big Intel adds on AnandTech in the coming months!
Perhaps when AnandTech talks about how advanced a platform is, they should wait until they can compare the performance with others as they did recently in their article...
http://www.anandtech.com/IT/showdoc.aspx?i=2447
AMD is a full generation ahead in technology and a generation behind in fluff.
Icehawk - Friday, June 17, 2005 - link
What matters is how fast Corporates embrace 64-bit outside the Enterprise app level (since most are still mainframe or at least Unix based). Generally technology and software updates are slow to rollout due to costs, contracts, stability, etc. If, somehow, 64-bit was to become part of the mainstream corporate app profile - THEN it will matter if Intel has a 64-bit mobile (laptop) processor.I would add at least one year past when Longhorn finally debuts before you see a major shift towards 64-bit. Until then it won't be a big deal IMO.
apriest - Wednesday, June 15, 2005 - link
Dual Dempsey's sounds intriguing... and dang expensive I'll wager...JarredWalton - Wednesday, June 15, 2005 - link
52:Add 1 for EM64T support on 5xx and 3xx CPUs.
Add 2 for VT support on 6xx CPUs.
Add 3 for 65nm 6xx CPUs that overlap 90nm parts (and have VT).
Otherwise, it's basically higher numbers within the same family gives better performance/features. It's a way to de-emphasize MHz/GHz, since we may be plateauing on clock speeds for a while. Intel spent so long convincing people to upgrade PCs for an extra 400 MHz that they now need to change tactics. Smart marketing, really.
cryptonomicon - Wednesday, June 15, 2005 - link
for the love of god, can someone explain INTELs not so new naming scheme? I cant make 1 bit of sense from it, it just leaves me mystified.I mean come on!! at least with AMD we knew the 3000+ meant somewhat-roughly-equivalent to a pentium of 3000mhz. gah!
PrinceGaz - Wednesday, June 15, 2005 - link
Another point with Yonah is that it is intended to be a low-power consuming Mobile design, not desktop. Most laptop users have no need of a faster CPU but do want longer battery life / smaller lighter battery etc.Given that the whole design of the Pentium M chips (Banias -> Dothan -> Yonah) is about saving power by artificially limiting the maximum intended clock speed, it makes sense for Intel to make Yonah no faster than Dothan so that they can further reduce the power consumption of the processor at any given speed compared to what it would draw if the design could reach, say 2.5GHz.
It doesn't seem at all surprising to me that Yonah will be little or no higher clocked than Dothan as there is little demand for the extra speed but plenty of demand for reduced power consumption.
JarredWalton - Wednesday, June 15, 2005 - link
45 - Possibly, but as we indicated in the article, it's just a guess. As I understand it, Vanderpool (VT) isn't adding a lot to the die size - actually, it's been present since the Prescott 2M/Irwindale cores, but deactivated. I'm more inclined to believe that Intel is just trying to separate the market: charge more for VT enabled chips, as they will go to server/workstation systems which tend to cost more.-----------
Regarding clock speed, while it's true that clockspeed isn't everything, we're essentially looking at a process shrink that isn't improving the top speed of CPUs at all. Yonah is pretty much Dothan (maybe) with dual cores. Banias topped out at 1.7 GHz, Dothan topped out at 2.26 (or will in a while), and Yonah is launching at 2.13 GHz max. It's surprising, that's all.
While MHz/GHz are not everything, the basic fact remains that similar architectures running at the same clock speed will perform similarly. Dothan 2.13 GHz will match a single core Yonah 2.13 GHz barring any drastic changes to the underlying architecture. Merom is the next major change in the underlying architecture, so we'll have a CPU that may be as much as 50% faster at the same clock speed.
Taking a more pragmatic look at the CPU environment, it's sort of interesting that the fastest (official) Northwood cores were 3.4 GHz and Prescott with 90nm and a longer pipeline only bumped that up 400 MHz. Willamette topped out at 2.4 GHz, so the 180nm to 130nm transition increased top CPU spped by 1000 MHz - 42% instead of only 12%!
Think AMD's done much better? The fastest 130nm chips from AMD were the FX-55 (2.6 GHz) and right now it doesn't look like they'll release anything above 2.8 GHz with 90nm SOI! Even if we throw out the FX-55 (which is a bit of a special case, since it's the only 130nm AMD chip with strained silicon), AMD still only went up 400 MHz with the transition - 17%.
The best we saw out of AMD/Intel for the transition to 90nm was a 33% speed bump from Banias to Dothan, but it sounds like that was also accompanies by a slightly longer pipeline.
If we got as much as a 50% speed increase going from 180nm to 130nm, and only a 33% going from 130nm to 90nm (even with adding copper, SOI, and strained silicon), what will 65nm bring? A maximum of a 20% speed increase for the same design? Maybe even less? Time will tell, but I find it an interesting trend to say the least!
piroroadkill - Wednesday, June 15, 2005 - link
Yonah was the only thing worth looking at. This is certainly no 'offensive' by Intel.Pandaren - Wednesday, June 15, 2005 - link
People keep shouting about 64-bits, but 64-bits won't make a difference to the vast majority of users. Most of the people I know who use computers surf the web, write email, print pictures, and other very basic tasks. In the corporate space, the transition to 64-bit will be glacial. The place I'm working at has used Windows 2000 for almost four years now and has no plans to change. They aren't the only ones:http://news.com.com/The+slow+road+to+Windows+XP/21...
The only ones 64-bit will make a difference to are the very small percentage of people who actually need it, and the fanbois who buy the biggest for the sake of having the biggest.