Why Bother with Three Cores?

The table below shows the problem with four cores:

SYSMark 2007 Overall E-Learning Video Creation Productivity 3D
Intel Celeron 420 (1 core, 512KB, 1.6GHz) 55 52 55 54 58
Intel Celeron E1200 (2 cores, 512KB, 1.6GHz) 76 68 91 70 78
% Increase from 1 to 2 cores 38% 31% 65% 30% 34%
Intel Core 2 Duo E6750 (2 cores, 4MB, 2.66GHz) 138 147 141 120 145
Intel Core 2 Quad Q6700 (4 cores, 8MB, 2.66GHz) 150 145 177 121 163
% Increase from 2 to 4 cores 8.7% 0% 26% 1% 12%

When going from one to two cores, overall system performance increases a decent amount. SYSMark encapsulates a wide variety of applications and usage models and overall performance increases by close to 40%. Obviously areas like video encoding (represented in the Video Creation tests) see the biggest gain, but all aspects of performance increase tremendously. Making the argument for two cores these days isn't a difficult one, most desktop applications can at least take some advantage of two cores.

Looking at the move from two to four cores however reveals much worse scaling. In our 1-to-2 core comparison cache size didn't increase, so the theoretical scaling could actually be even higher but in the 2-to-4 core comparison the total L2 doubles since Intel's quad-core processors are simply two dual-core die on a single package. Despite the increase in cache size however, scaling is quite poor. Overall performance goes up 8.7% percent and the E-Learning/Productivity tests see no gains at all. Once again the biggest gains come from the Video Creation tests, followed by the 3D suite.

For the vast majority of systems, four cores just aren't necessary. There are some applications that do scale very well between 2 and 4 cores, but the overall landscape is much like what we saw with dual-core CPUs circa 2005, the time for quad just isn't now. Intel's CPU shipments also reflect that both the need and demand for quad-core CPUs just isn't very high:

Currently, less than 10% of Intel's consumer desktop CPU shipments are quad-core and that number won't grow much beyond 10% by the end of 2008. But just like the early days of dual-core, we'll see a steady ramp up continuing in the years ahead.

The point here isn't that quad-core processors aren't necessary, rather they aren't quite in their prime as far as demand goes. With such a small portion of the market purchasing quad-core CPUs, the ISVs aren't exactly jumping at the opportunity to make sure all applications scale well from 2-to-4 cores. Some inherently won't scale while others may with additional effort, which requires a large install base and once more we find ourselves in the midst of an overused analogy involving chickens and eggs.

For Intel, a slow adoption of quad-core CPUs isn't much of a problem. It's just as easy to make a Core 2 Quad as it is to make a Core 2 Duo, the former simply has two dual-core die on the package instead of one. For AMD however, things are a lot more complicated.

One often cited reason for Phenom's late arrival was its "native" quad-core design. Due to its on-die memory controller/north bridge, AMD could not simply take two Athlon X2 die and individually place them on the same package; all four cores would have to be behind the memory interface, meaning that all four cores would have to be on the same die.

Number of Cores Manufacturing Process Transistor Count Die Size
AMD Phenom X4 4 65nm 450M 285 mm^2
AMD Phenom X3 3 65nm 450M 285 mm^2
AMD Athlon X2 2 90nm 243M 219 mm^2
AMD Athlon X2 2 65nm 221M 118 mm^2
Intel Core 2 Quad 4 65nm 582M 286 mm^2
Intel Core 2 Duo 2 65nm 291M 143 mm^2
Intel Core 2 Quad 4 45nm 820M 214 mm^2
Intel Core 2 Duo 2 45nm 410M 107 mm^2

Looking at the die size column you can see an issue with AMD's current processor lineup. AMD likes building the 65nm Athlon X2s, they are nice and small at 118 mm^2 per die and it can make a lot of them on a single 65nm wafer. The Athlon X2 6400+ is still built on a 90nm process and its die, by comparison, is huge; AMD doesn't like making these chips very much (update: AMD has actually ceased production of 90nm X2s altogether).

Then we have Phenom. At 285 mm^2, Phenom is huge and AMD can't make that many per wafer, plus with such a large die the yield is lower than on a smaller chip. The triple-core Phenom X3 gives AMD something to do with those quad-core die that have a single defective core, rather than throwing the entire chip away it can now be repackaged and sold as a triple-core processors.

The other problem here is that there is no dual-core Phenom, so AMD must battle Intel's 45nm dual-core processors with its very old 65nm Athlon X2s. The Phenom X3 is designed to help alleviate the burden of those poor K8s by competing with Intel's dual-core in the sub-$200 space. It's a great marketing story too: you can get three cores from AMD for the price of two from Intel.

It's more likely than not that AMD's yields aren't bad enough to have too many quad-core Phenom processors with two defective cores, which is probably why we don't see a 285 mm^2 Phenom X2.

AMD has no plans to make a separate triple-core die, simply because it would require quite a bit of engineering resources and the need for triple-core CPUs diminishes over time as quad-core adoption increases. Right now AMD is focused on bringing its 45nm Phenom processors to market and those are occupying all of AMD's availability engineering resources. Should triple-core prove to be a worthwhile addition to the lineup, AMD could always work on designing a tri-core die but for now it will fulfill its role as a stopgap solution.

Index A New CPU & Intel's Biggest Problem
POST A COMMENT

44 Comments

View All Comments

  • Locutus465 - Thursday, April 24, 2008 - link

    how do you justify this position when comparing the platforms as a whole, particularly taking into consideration budget platforms with integrated graphics. Reply
  • Roy2001 - Thursday, April 24, 2008 - link

    For HTPC, 780G paired with a 5600+ is enough. Actually, a E2200 can decode any HD/BD movies, who cares chipset?

    For multicore usage, get a Q6600/Q9450, gamers wang a E8400/E7200. Period.
    Reply
  • Roy2001 - Thursday, April 24, 2008 - link

    For HTPC, 780G paired with a 5600+ is enough. Actually, a E2200 can decode any HD/BD movies, who cares chipset?

    For multicore usage, get a Q6600/Q9450, gamers wang a E8400/E7200. Period.
    Reply
  • Nihility - Wednesday, April 23, 2008 - link

    Great review!
    The power consumption test seems to indicate the X3 requires less power per core than the X4. A 25% decrease in system power consumption after removing 1 core seems to prove that.
    I was hoping for an overclocking test. Naturally I assume it would be just as bad as the X4 version but due to the lower power requirements I'm curious if AMD managed to improve it.
    Reply
  • Schugy - Wednesday, April 23, 2008 - link

    The Phenom X3 is a great CPU just as the X4 but there's a lot of outdated software. It's obvious that the Phenom likes software that is frequently updated with its capabilities in mind. The Phenom has a lot of horsepower for Nero Recode (Imagine 45nm and 3GHz or more), Main Concept-Encoder, LDAP, UT3, AutoMKV but some software makers are extremely good in wasting it. Reply
  • bpl442 - Wednesday, April 23, 2008 - link

    Nvidia Geforce 7100 / nforce 630i based motherboards like the Gigabyte GA-73PVM-S2.


    "Unfortunately, Intel is in a not-so-great position right now when it comes to its platforms. It can't turn to ATI anymore for integrated graphics solutions, and with a full out war on NVIDIA brewing, it's left alone to provide chipsets for its processors as NVIDIA's latest IGP solutions are not yet available for Intel CPUs."
    Reply
  • nubie - Wednesday, April 23, 2008 - link

    I think they are referring to the 8200/8300 series, and the like with full HD decode, also possibly 65nm/55nm so that the power usage is less.

    nVidia isn't ready yet for Intel. (personally I don't care for onboard graphics, mainly because of their lack of TV outputs, and it seems that they are being skipped over entirely and switched to HDMI/DisplayPort if anything, I did once own a 6150 with S-Video/Composite/Component output, but that is the exception to the rule. And still no HD decode.)
    Reply
  • FodderMAN - Wednesday, April 23, 2008 - link

    Why does no one test these at faster HT speeds?

    I still stand by the statement I have made time and again that these processors are being castrated by the 200FSB HT speeds. If you run a classic athlon64 at 233, 250, or even 266 ( when attainable ) these processors start to really outshine Intel’s procs. And I can only see phenom being that much stronger. Now I know this is considered an overclock as AMD has zero procs at the higher bus speeds. This would help AMD's revenues as they would be able to add another product code for procs running various FSB / HT speeds even though they would shrink the multiplier range as these cores seem to run into stability problems at about 2.6 - 2.8ghz.

    PLEASE someone do a test at higher bus speeds so we can refer AMD to published numbers. I guess for the time being I will have to wait for a good 6 phase powered board before I buy my phenom and do the testing myself. But I always had great luck with the classic ath64's when upping the FSB / HT speeds and lowering the multipliers. The athlon64 always showed great promise at 250-266 FSB/HT and I can only imaging how well they would run at a 333 FSB / HT speed.

    Time will tell,
    The Goat
    Reply
  • retrospooty - Thursday, April 24, 2008 - link

    "a classic athlon64 at 233, 250, or even 266 ( when attainable ) these processors start to really outshine Intel’s procs."

    There isnt any situation at all, where AMD is outshining Intel CPU these days. If you OC the AMD to 266, you can OC the Intel to 500+ FSB, and AMD isnt outshing anything, not in speed, performance, power, overclockability, or bang for your buck. Intel is winning by all measurements.
    Reply
  • Assimilator87 - Wednesday, April 23, 2008 - link

    According to the guys at XS, raising the HT reference clock doesn't affect Phenom's performance in any way. It's the effective HT clock that matters, which is currently 1800Mhz using a multi of 9. Reply

Log in

Don't have an account? Sign up now