Original Link: https://www.anandtech.com/show/926
AMD's 0.13-micron Thoroughbred - Cool but Slipping at 2200+
by Anand Lal Shimpi on June 10, 2002 2:55 AM EST- Posted in
- CPUs
It has been a long three months since AMD said farewell to the Palomino core; what once was a highly anticipated extension of the Athlon processor line quietly faded away as AMD made room for much more exciting microprocessors.
With all of the press that AMD's next-generation architecture has been receiving you'd think that Hammer was a mere few weeks away from its launch. Unfortunately reality sets in and it becomes clear that there is a large gap between now and the end of the year when judgment day begins for AMD's bet on Hammer.
We've provided you with countless roadmaps and other hints to give you an idea of what to expect from the two microprocessor giants over the coming year. Today is a day when one of those boxes on AMD's roadmap is introduced and excitement ensues. Today is the day that AMD's Palomino core gets the boot and is replaced with a thinner, cheaper and cooler running successor - Thoroughbred.
Before you get overly excited about a new Athlon XP core, realize that Thoroughbred offers no tangible performance improvements over the Palomino core. The core is smaller, uses slightly fewer transistors and can run at higher clock speeds because of this. The questions you should be asking yourself are:
1) How much cheaper will the processors be for me?
2) How much further can they be overclocked?
3) How much cooler do they run, and
4) What's compatibility like with current motherboards?
We'll answer every last one of those questions as well as do the usual performance analysis spiel. To lessen hopes a bit more before diving into the review, the Thoroughbred core is currently shipping at no higher than 1.8GHz, only a 66MHz bump over the previous Palomino core. Let's start dissecting…
How does AMD do that voodoo they do so well?
For several months after Intel introduced the Pentium 4 you'd be silly to even consider purchasing the processor. Performance aside, the Pentium 4 was a very expensive processor to own and in the end it didn't come down to paying for the Intel name but it was due to the fact that it was an expensive processor to manufacture. What factors influence CPU manufacturing cost?
1) The number of useable/good CPUs on a single silicon wafer (yield)
2) The size of the die being manufactured
3) The manufacturing process being used
4) The size of the wafer that the dice are being produced on
5) The cost of packaging the CPU (everything that goes along with connection the silicon to the pins on the packaged CPU)
Intel's Pentium 4 core is quite large and going forward there will be even more features tacked onto the core. By this time next year Prescott will be readying for release with a full 1MB L2 cache on the processor die itself. Large caches are easily the biggest consumers of transistors and die space on today's CPUs. Intel's solution to dealing with their large processors is to quickly transition to smaller manufacturing processes and move to larger silicon wafers (thus yielding more chips per wafer). This is exemplified by Intel's early transition to a 0.13-micron manufacturing process, Intel's commitment to ship Prescott on a 0.09-micron process and the move to 12" wafers.
AMD takes a different approach that suits their needs and abilities better. The Athlon XP and even the forthcoming Hammer processors both have a very small die size, a benefit of being devoid of any trace caches among other things. AMD is a bit late to the 0.13-micron game as Intel has been producing on the process since January of this year, but they're finally making their way along. AMD won't have the ability to move production to 12" wafers until 2005 forcing die sizes to remain relatively small in the near future for most mainstream parts.
|
Palomino
|
Thoroughbred
|
 |
 |
Palomino vs. Thoroughbred, Intel's Northwood is bigger than both.
AMD's ability to offer much lower prices than Intel for their highest end parts comes with their small die sizes; by not adding any features to the Athlon XP processor when transitioning to the 0.13-micron Thoroughbred core, AMD buys themselves much more room to adjust prices. This means that Thoroughbred Athlon XP processors won't necessarily be noticeably cheaper to purchase immediately, but if necessary AMD can drop prices a bit more aggressively now.
The downside to this, as we mentioned at the start of this article, is that these new Athlon XPs don't perform any differently from their predecessors.
How to spot a Thoroughbred
It is in AMD's best interests to move their production to 0.13-micron as soon as possible since it increases profit margins on their wonderful little CPUs. Exactly how small are these things? The first time we saw Thoroughbred a couple of months ago we were simply shocked at how small it was. We had seen the specifications on paper but there's something about reading 80mm^2 and actually seeing a core in person that makes all the difference in the world.
From left to right - Thoroughbred (0.13-micron), Palomino (0.18-micron), and Thunderbird (0.18-micron)
The Thoroughbred core will be introduced at clock speeds that overlap with the older Palomino parts; more specifically, model numbers from 1700+ up to 2200+ will be made available in 0.13-micron Thoroughbred versions. The picture above shows the Thoroughbred with a brown organic package instead of the green substrate that was introduced with the 2100+, eventually all of the CPUs will be built on the green substrate it just so happens that all review samples were built on brown substrates. The only differences are strictly aesthetic.
AMD had to relocate the power regulating capacitors from the Palomino (right) to the top of the Thoroughbred chip simply because it's too difficult to run the wires from the bottom of the chip to the very small 80 mm^2 core.
How will you be able to make sure that the chip you're ordering is a Thoroughbred and not a Palomino core? The designation is actually quite simple; if you look at the Ordering Part Number (OPN) you'll see that Palomino based Athlon XPs followed this general form:
AX 1900 MT3C
The first two characters were always AX for Athlon XP processors and the first three were AMP for Athlon MP processors. The following four characters represented the clock speed or model number of the processor (for the first Athlon MPs it was clock speed but for all other CPUs it is the model number of the CPU, AMD does not list the clock speed anywhere on the chip itself). The final string of characters provide some information about the core itself but aren't useful in determining whether you have a Thoroughbred or not.
When ordering a Thoroughbred online make sure that the first four characters in the OPN read AXDA instead of simply AX for Athlon XP processors; the 'DA' suffix indicates the new Thoroughbred core. For example, the OPN for our 0.13-micron Athlon XP 2200+ was AXDA2200DV3C.
What makes Thoroughbred's clock tick
The architectural specifications of Thoroughbred are identical to its 0.18-micron predecessor which we've written about extensively here and here. Here are some of the details of the Thoroughbred core that do set it apart from Palomino:
|
Thoroughbred
vs. Palomino vs. Intel's Northwood
|
|||
| Code Name |
Palomino
|
Thoroughbred
|
Northwood
|
| Manufacturing Process |
0.18-micron
|
0.13-micron
|
0.13-micron
|
| Die Size |
128
mm^2
|
80
mm^2
|
146
mm^2
(eventually 131 mm^2) |
| Transistor Count |
37.5
Million
|
37.2
Million
|
55
Million
|
| Voltage |
1.750V
|
1.50V
- 1.65V
|
1.50V
|
| Clock Speeds |
1.2
- 1.73GHz
|
1.47
- 1.8GHz+
|
1.6
- 2.53GHz+
|
You'll notice that there is a small decrease in transistor count when going to the Thoroughbred core. AMD was able to accomplish this by optimizing the core during the die shrink similar to what happened with the Palomino. You'll also notice that the Thoroughbred core requires much higher voltages to hit much lower clock speeds than Intel's Northwood, keep that in mind especially when we talk about the overclocking potential of the new core later on in the review.
Thermal Comparison
|
Processor
Model |
Operating
Frequency (MHz) |
Nominal
Voltage |
Typical
Thermal Power |
Maximum
Thermal Power |
Typical
Current Working State |
Max
Current Working State |
| AXDA 1700+ |
1467
|
1.50V
|
44.9W
|
49.4W
|
29.9A
|
32.9A
|
| AXDA 1800+ |
1533
|
46.3W
|
51.0W
|
30.9A
|
34.0A
|
|
| AXDA 1900+ |
1600
|
47.7W
|
52.5W
|
31.8A
|
35.0A
|
|
| AXDA 2000+ |
1667
|
1.60V
|
54.7W
|
60.3W
|
34.2A
|
37.7A
|
| AXDA 2100+ |
1733
|
56.4W
|
62.1W
|
35.2A
|
38.8A
|
|
| AXDA 2200+ |
1800
|
1.65V
|
61.7W
|
67.9W
|
37.4A
|
41.2A
|
| 1500+ |
1333
|
1.75V
|
53.8W
|
60W
|
30.8A
|
34.3A
|
| 1600+ |
1400
|
56.3W
|
62.8W
|
32.2A
|
35.9A
|
|
| 1700+ |
1467
|
57.4W
|
64W
|
32.8A
|
36.6A
|
|
| 1800+ |
1533
|
59.2W
|
66W
|
33.8A
|
37.7A
|
|
| 1900+ |
1600
|
60.7W
|
68W
|
34.7A
|
38.9A
|
|
| 2000+ |
1667
|
62.5W
|
70W
|
35.7A
|
40A
|
|
| 2100+ |
1733
|
64.3W
|
72W
|
36.7A
|
41.1A
|
|
|
Pentium
4
|
||||||
| 2.0A |
2000
|
1.50V
|
52.4W
|
44.3A
|
||
| 2.2 |
2200
|
55.1W
|
47.1A
|
|||
| 2.26 |
2266
|
56W
|
48A
|
|||
| 2.40B |
2400
|
57.8W
|
49.8A
|
|||
| 2.53 |
2533
|
59.3W
|
51.5A
|
|||
Note: AXDA denotes Thoroughbred core
The main thing to take away from this chart is that the Thoroughbred puts out around 12% less heat than an equivalently clocked Palomino core. Despite the reduction in heat dissipation the core still puts out more heat than Intel's Northwood core which is understandable because of the Athlon XP's higher IPC. You'll also notice that the Northwood draws more peak current (Intel didn't provide Max thermal power or typical current data) than the Athlon XP; it does so at a lower voltage and thus gets away with less heat.
Clock Speeds and Expectations
With the only tangible performance improvement offered today being the 66MHz speed bump of the Athlon XP 2200+, you shouldn't expect to see too much on the benchmark side. Remember that going from March's XP 2100+ launch to today's 2200+ release actually only spans a 3.9% increase in clock speed, meaning that you'll see an even smaller increase in performance scores.
By far the most attractive points about the Thoroughbred are its smaller core, cooler operation and increased headroom. The last point we'll address after the benchmarking section in our overclocking tests.
Motherboard Compatibility
The 0.13-micron Athlon XPs should work just fine in existing Socket-A motherboard platforms that already supported the Palomino; the main requirement for proper operation with these motherboards is BIOS support. As we're just about to publish our KT333 roundup we ran some quick compatibility tests on all of the motherboards in the roundup to point out which manufacturers need to hurry up and release BIOS updates in order to properly support Thoroughbred. Thankfully only a select few gave us any issues, so here's the info:
|
Motherboard
Compatibility
|
|||
| Motherboard |
BIOS
Revision
|
Date
|
Works
with Thoroughbred?
|
| ABIT AT7 |
4/22/2002
|
No
|
|
ABIT KX7-333 |
kx77m
|
4/22/2002
|
No
|
ASUS A7V333 |
latest
|
Yes
|
|
| ECS K7VTA3 |
VTA30404
|
4/04/2002
|
Yes
|
| EPoX 8K3A+ |
8k3a2328
|
3/28/2002
|
Yes
|
Gigabyte GA-7VRXP |
7vrx_f2
|
5/9/2002
|
Yes
|
| MSI KT3 Ultra |
6380ev531
|
4/10/2002
|
Yes
|
| Shuttle AK35GT2/R |
ak35s20c
|
5/8/2002
|
No
|
| Soyo KT333 Dragon Ultra |
KVXB2AA1
|
4/26/2002
|
Yes
|
Keep in mind that the motherboards listed above that currently do not support the Thoroughbred core should be able to offer support through a simple BIOS update.
The Test
|
Windows
XP Professional Test Bed
|
|
|
Hardware
Configuration
|
|
| CPU |
AMD
Athlon XP 2200+ (1.80GHz)
AMD Athlon XP 2100+ (1.73GHz) AMD Athlon XP 2000+ (1.67GHz) AMD Athlon XP 1800+ (1.53GHz) AMD Athlon XP 1600+ (1.40GHz) Intel Pentium 4 2.53GHz Intel Pentium 4 2.40B GHz Intel Pentium 4 2.40GHz Intel Pentium 4 2.20GHz Intel Pentium 4 2.0A GHz Intel Pentium 4 2.0GHz Intel Pentium 4 1.8GHz Intel Pentium 4 1.6GHz |
| Motherboard |
EPoX
8K3A+ - VIA KT333 Chipset
ABIT TH7-II RAID - Intel 850 Chipset Intel D850EMV2 - Intel 850E Chipset |
| RAM |
1
x 256MB DDR333 CAS2 Corsair XMS3000 DIMM
2 x 128MB PC800 Samsung RIMMs 2 x 128MB PC1066 Samsung RIMMs |
| Sound |
None
|
| Hard Drive |
80GB
Maxtor D740X
|
| Video Cards (Drivers) |
NVIDIA GeForce4 Ti 4600 (28.32) |
Internet Content Creation & General Usage Performance
With this review we continue to use SYSMark 2002; SYSMark 2002 can be considered to be a much more memory bandwidth intensive version of the Winstone tests. The benchmark is split into two parts, Internet Content Creation which deals with content creation applications (Photoshop, Dreamweaver, etc...) and Office Productivity which is more general usage oriented (Word, Excel, Netscape, Anti-Virus, etc...).
The 2002 update changes things around a bit; first of all the benchmark's total scores are arrived at differently than in the 2001 benchmark. Windows Media Encoder no longer accounts for close to half of the Internet Content Creation test, rather only about 10%. There is also no need for a special Athlon XP SSE patch as the 2002 suite uses a version of the encoding dll that properly detects SSE support on all Palomino cores as well as Pentium 4 cores.
The rest of the benchmark is much more evenly distributed and it is much more memory bandwidth intensive than the old benchmark. The Internet Content Creation tests on average use about 600MB/s of bandwidth vs 300MB in SYSMark 2001. The Office Productivity tests are still stuck at around 580MB/s of memory bandwidth.
For more information on the tests and the applications used consult this whitepaper provided by BAPCo.
|
178
169
155This is quite possibly the most important benchmark when it comes to measuring the way the majority of users interact with their PCs on a daily basis. Differences greater than 10% are generally noticeably in the real world and taking that into account you'll see that the Athlon XP 2200+ actually falls just outside that marker when compared to Intel's fastest. The performance is competitive but remember that we're not dealing with a huge speed bump since we last investigated the Athlon XP's performance.
As you can see from the CPU scaling chart, it will simply take a higher clocked Athlon XP to continue to be competitive with the Pentium 4 under most general use applications. This will change once Prescott arrives with its 1MB L2 cache but by then the highest end Athlon processor will have the benefit of an on-die memory controller which should help AMD get by with a smaller cache.
|
SYSMark's Internet Content Creation suite is more of a niche test focusing on those users who are constantly using their PC to produce HTML, Flash, Video and other similar types of content. The performance here is greatly skewed towards the Pentium 4 which definitely excels at this type of usage.
A relatively flat scaling curve from the Athlon XP indicates that we'll need to see some architectural changes before the processor family can catch up to Intel here. A 512KB L2 cache Athlon XP would help but what's truly necessary here is Hammer.
Media Encoding Performance
What once was a very CPU intensive task is now fairly trivial. Because of the streaming nature of MP3 encoding, having a larger cache doesn't necessarily result in a tangible increase in performance. The reason we continue to stress MP3 encoding as a CPU benchmark is mainly because of the fact that MP3 encoding usually does play a role in larger projects such as MPEG-4 video encoding where you're ripping audio as well as video.
|
Media encoding of any sort is generally a very CPU intensive task that scales very well with clock speed. The performance standings are fairly self explanatory here, it's the CPU scaling chart that does the majority of the talking however:
We can see the Athlon XP illustrate the beginnings of leveling off in performance, although we'll need another speed grade to truly be sure. A larger L2 cache won't help much here as we're dealing with streaming data from main memory to the CPU and rarely making use of repeated accesses of the same bits of data.
Video Effects Rendering Performance
We added two benchmarks to our suite when we reviewed the Pentium 4 2.53GHz: Adobe After Effects 5.5 and NewTek's Lightwave 7.5. These two make good examples of what heavy SSE2 optimizations can bring to the Pentium 4. You'll remember from the original discussions of the Pentium 4's architecture, many criticized Intel's decision to move to an essentially weaker x87 FP execution setup in favor of putting great faith in the adoption of SSE2. The adoption of the instruction set has been going well but as you can tell by most of our 3D rendering and other FP intensive benchmarks, the Pentium 4 is only now becoming competitive because of its high clock speeds.
With AMD's Opteron and the next-generation Athlon scheduled to receive support for Intel's SSE2 instructions as well, the assimilation of SSE2 optimizations into as many applications as possible is in the best interests of both CPU giants. If history is any indication however, it will take quite a bit of time to see significant optimizations in place.
Adobe After Effects is one application that has received a high level of SSE2 optimizations as the type of video manipulation the program allows is perfectly suited for SSE2. Let's have a look at the results:
|
Although the Pentium 4 has a significantly weaker FPU than the Athlon XP, once SSE2 is thrown into the equation the picture changes dramatically. Intel has done a great job of making sure that applications are taking advantage of the Pentium 4's SSE2 instruction set where applicable; this is also good for AMD as it won't be long before the Athlon (Hammer) can take advantage of these optimizations as well.
Again we see that CPU scaling is almost linear for all of the CPUs, with the gap between the AMD and Intel curves existing solely because of SSE2 optimizations. AMD's SSE2 support in Hammer will close that gap but it will also put the Hammer's SSE2 core to the ultimate test. It is also worth noting that once again, due to the streaming nature of the test, L2 cache size has little effect on the overall performance data.
3D Rendering Performance
Next we have our usual two 3D rendering tests. We'll start off with rendering the first frame of the Waterfall.max scene (provided on the 3DSMAX CD) at 1024x768
|
The Pentium 4 has been able to overcome its architectural weaknesses by quickly scaling to higher clock speeds which allow it to still be at the top of the charts even under 3D Studio MAX; this time around Intel has to share that spot with the Athlon XP 2200+.
AMD has little to worry about in the x87 intensive workstation market; higher clock speeds will give them all they need to remain competitive.
|
Although a different application, Maya's performance characteristics don't differ at all from 3D Studio MAX. Remember that it's not the application that dictates performance, it's the workload that determines everything.
3D Rendering Performance using SSE2
While 3D Studio MAX is SSE2 optimized, the level of optimization is nowhere near what NewTek reported with Lightwave upon releasing version 7.0b. The performance improvements offered by the new SSE2 optimized version were all above 20% using NewTek's supplied benchmarking scenes.
We chose two benchmarks to use, the least SSE2 optimized one and another that is more optimized just to get an idea of the potential that lies for Pentium 4 users running heavily optimized applications
|
Again we see SSE2 flex its muscle and the Pentium 4 skyrocket to the top. CPU scaling is similar among the two tests so we just stuck to one chart:
|
3D Gaming Performance
When it comes to most 3D games there's generally very little performance to be found by heavily optimizing for SSE2 or 3DNow! on either of these processors and thus the performance is mostly dependent on the overall platform (e.g. FPU capabilities, chipset, memory latency/bandwidth, cache latency/bandwidth, etc...).
We'll start off with our favorite 3D gaming benchmark - the Unreal Performance Test 2002. For an explanation of what this test is and why it is so significant, be sure to read our 15-way GPU Shootout that we used to introduce the test. In short, the benchmark uses the current build of the Unreal Engine (that will power games such as UnrealTournament 2003 and Unreal II) and serves as a great indication for future performance in games that use the engine.
|
The Athlon XP seems to be fine for next-generation 3D game engines as it is able to come within 0.3 fps of the 2.53GHz Pentium 4 here. It will be interesting to see how Doom 3 performs on the two processor families but the workload should be quite similar to what we're seeing here. The biggest determinant of performance will end up being the GPU in these next-generation games as the CPU's role will be diminished to mostly physics and AI.
|
Current games are much less limited by GPUs and are thus more CPU dependent. With NVIDIA's GeForce4 Ti 4600 already able to run almost all of today's games at the highest possible settings flawlessly, it isn't too surprising that CPU performance has a much bigger impact on the overall performance picture.
|
Not a good Overclocker
One of the reasons we usually get excited with the launch of a new core is to test its overclocking potential, unfortunately the Thoroughbred core doesn't succeed in impressing us. We had a sample of four Thoroughbred CPUs that are pictured below:
The CPUs seen above include an Athlon XP 2200+, 1900+, 1800+ and 1700+; all of the CPUs used the new Thoroughbred core.
Unlocking these processors for overclocking was quite simple since the L1 bridges didn't have little pits dug into them as you can see from the following picture:
The L1 bridges just need a little bit of conductive ink to unlock this CPU
One of the best ways to see how much headroom a manufacturing process has is to see how far you can overclock a chip with normal cooling and at default core voltages, so we did just that on these four CPUs. The results are as follows:
Athlon XP 2200+ (1.80GHz) - The highest we could overclock this core at default core voltage with AMD's recommended heatsink/fan was 1822.5MHz using a 135MHz FSB.
Athlon XP 1900+ (1.60GHz) - We managed to increase the clock multiplier and the FSB to 12.5x and 140MHz respectively, resulting in a 1750MHz overclocked frequency.
Athlon XP 1800+ (1.53GHz) - The 1800 made it to 1.60GHz, the speed of a 1900+ without any effort.
Athlon XP 1700+ (1.47GHz) - The 1700 made it to 1.53GHz, the speed of an 1800+.
It seems as if the Thoroughbred core needs some time to mature at this point, let's see how far we took the CPUs at voltages as high as 1.850V. Again, we stuck to stock cooling:
Athlon XP 2200+ (1.80GHz) - The Athlon XP 2200+ wouldn't go any further, 1822.5MHz was its peak.
Athlon XP 1900+ (1.60GHz) - We managed to increase the clock multiplier and the FSB to 12.5x and 150MHz respectively, resulting in a 1875MHz overclocked frequency.
Athlon XP 1800+ (1.53GHz) - Our best overclock out of the batch took us to a 12.5x multiplier with a 145MHz FSB which left us with a 1813MHz processor.
Athlon XP 1700+ (1.47GHz) - The 1700+ made its way up to 1.67GHz, the speed of a XP 2000+.
Once again, nothing stellar out of the cores although we did get a little more out of them.
Final Words
When you have an extremely successful launch like what AMD achieved with Palomino, it's very difficult to one-up yourself and in this case it doesn't seem that AMD has. The Thoroughbred core does run cooler and cost less than its predecessor but there's simply not enough headroom in the current process to really get excited about it. Obviously if you're buying an Athlon XP today then you'll definitely want to make sure you get a Thoroughbred, but current Palomino owners should have no burning desire (no pun intended) to upgrade just yet.
The lack of overclocking headroom in these 0.13-micron parts make us question exactly how well AMD's 0.13-micron process is doing at this point. If yields are currently low and set to improve over time then AMD's 0.13-micron process is at fault currently, however if the process is sound and 1.8GHz is the max we're getting out of these cores then that points to architectural limitations that are holding the Athlon XP back. We will be keeping a very close eye on how AMD is able to scale their 0.13-micron processors over the coming months, if they are unable to hit above 1.8GHz today it may be a struggle between now and the end of the year to remain competitive with the Pentium 4.
On the other hand it seems as if AMD's sole purpose right now is to ensure that Hammer gets off on the right foot. A couple design wins in the server market next year for Opteron could spell more success for AMD than any Athlon has been able to do on the desktop side. In talking to VIA at Computex last week we came across a very important point about AMD's strategy with K8 vs. K7; if AMD can succeed with the K8 in the enterprise market then it opens new doors in the corporate market that were never penetrated by the current Athlon.
This may mean that the Athlon XP family is put on hold for a while as Intel gains a strong foothold in the desktop segment for the sake of making sure that Hammer is all that it could possibly be by the end of the year. We may see a few more 66MHz bumps between now and then but don't expect to get too excited about them unless AMD does decide to go ahead with Barton...
Facebook
Twitter
RSS
