Traduction du test du Barton
http://www.flickerdown.com/phpBB2/viewtopic.php?t=2301
AMD Barton 2500+ arrives ? the true power of a 512K L2 cache is revealed
After several core revisions, the Athlon XP will move to a more interesting core variant, the long-awaited Barton. To be honest, the Barton has been delayed for a considerable length of time. Rumours of its impending launch have been rife since the release of the Thoroughbred, but what enthusiasts saw was the release of the TBred-B, then a shift to a 166MHz FSB. The road to the Barton with 512K L2 cache has indeed been a long one...
AMD Barton 2500+ arrives
Although we?ve waited so long, this time we weren?t disappointed: Barton has finally emerged. Let?s have a look:
(Picture of Barton core)
This is the real face of Barton. It doesn?t look all that different to a TBred-B, but the die size has increased slightly due to the onboard 512K L2 cache. The bottom of the CPU remains relatively clean and continues to sport 462 pins. Unlike the Athlon 64 (which will use Socket 754), the Barton will be usable on Socket A motherboards with a simple BIOS flash. However, the motherboard must support a 166MHz or even 200MHz FSB.
(Picture showing underside: ?The Barton?s underside is as clean as the existing Athlon XP?)
(Picture showing top: ?The die looks slightly wider?)
Let?s first examine the die: It looks quite like a rotated Duron (Morgan) die and is quite a bit wider than the previous TBred-B. This is especially clear when the two CPUs are placed next to each other. Although the two are aesthetically similar and both use a 0.13 micron process, in terms of die size the two are slightly different: the Barton?s die looks more elongated than that of the TBred-B and is obviously larger due to the increased cache. Like the TBred-B, the Barton uses OPGA packaging and is SOI-free. Also, the Athlon XP 2200+ (TBred) has an actual clock speed of 1800MHz, but the Barton 2500+?s actual clock speed is 1833MHz. It seems that the combination of a 166MHz FSB and 512K of L2 cache pays significant dividends ? or so AMD believes. We?ll test this using benchmarks later on...
(Picture of TBred-B and Barton: ?The TBred-B and Barton look similar, but their die sizes differ slightly?)
Currently the Athlon XP market is going through a core change: the TBred-A and ?B revisions are now emerging on the market, including the 2400+, 2600+ and the 166MHz FSB 2700+ and 2800+. Yet even now there are Athlon XPs based on the Palomino core on sale, and new PCs continue to be based on it. If we compare the three it is easy to see the changes in die size and core revision. If you are interested in buying an Athlon XP, make sure your hard-earned money goes into buying the right core.
(Picture comparing the Palomino, TBred and Barton)
But back to Barton. Apart from die size, the CPU markings give a clearer answer. ?AMD Athlon? is still clearly marked. The ?0243? marking shows just how new this Barton is: a Week 43, 2002 product. The ?ES? behind it needs little explanation of its significance: Engineering Samples are multiplier-unlocked.
As for the first row ?AXDA1833DKT4D?, this shows the Barton?s clockspeed and other infromation, and is worth explaining further:
We can split this row into eight sections to identify its OPN:
AXD: As the name suggests, this is an acronym for ?Athlon XP Desktop?. As the AMD Datasheet says: AXD=AMD Athlon XP Processor Model 8 with QuantiSpeed Architecture for Desktop Products
A: Desktop Processor
1833: Originally this area would show the Athlon XP?s QS rating (e.g. 2600 for 2600+), but it shows the actual clock speed (1833MHz) here. However, according to our research this is a 2500+ processor. We?ll explain how the QS ratings are calculated later.
D: Packaging type ? ?D? stands for an OPGA package.
K: Vcore voltage. ?K? stands for 1.65V, ?L? stands for 1.50V, ?U? stands for 1.60V.
T: Maximum die temperature. ?T? stands for 90oC, ?V? stands for 85oC.
4: Represents L2 cache size. ?4? means 512K, ?3? means 256K.
D: Front Side Bus speed. ?B? was the old 200MHz FSB, ?C? represents a 266MHz FSB and ?D? represents the new 333MHz FSB.
This much is clear: the Barton?s OPN is defined differently, but its Vcore is still 1.65V and its die temperature limits are again the original 90oC. Further, the packaging number is also different to the TBred-B?s: The Barton?s is 27488 whereas the TBred-B?s is 27291:
(Picture of ?AMD Athlon? label: ?The printed markings clearly show the actual clock speed?)
(Picture of OPGA packaging: ?Even the OPGA package version is different?)
One more detail to note: Since the Barton looks so similar to the TBred-B, it?s worth noting multiplier and MP details and the L3 and L5 bridge settings:
On the left is an enlargement of the Barton bridges. Most importantly, the L1 and L3 bridges (11X multiplier) are uncut. The enlargement of the TBred-B?s bridges at the right also show uncut L1 bridges, but the L3 bridges (13.5X multiplier) are cut. The L5 (MP identifier) bridges tell a similar story (i.e. they are also cut), but connecting the last bridge will also allow the operation of dual Bartons. Also note the L12 bridges to the left of L5: On the L12 bridges, cutting the second bridge from the right (i.e. cutting that bridge and the rightmost bridge) enables a 133MHz FSB, but the Barton?s L12 bridges are intact except for the rightmost bridge. The same L12 configuration can be found on the 166MHz FSB XP2700+ and 2800+. It may well be possible that connecting the second-rightmost bridge will enable a 166MHz FSB on the Athlon XP. The performance increases from this will be explained later.
(Picture of Barton?s bridges: ?Barton?s L1 bridges are uncut ? its multiplier is unlocked. The L3 bridges show a 11X multiplier?)
(Picture of TBred-B?s bridges: ?The TBred-B?s bridges are cut for a 13.5X multiplier?)
A large 640K cache
We?ve already carefully examined the external dimensions of Barton. Although the CPU markings show the presence of a 512K L2 cache and a 166MHz FSB, ultimately it?s more practical to test this using publicly available CPU identifying software (in this case WCPUID 3.10).
(WCPUID error message: ?Unknown CPUID Data?)
As you can see, we get an ?Unknown CPUID Data? message from WCPUID. Of course, if it was possible, we could email the CPUID data to H.Oda (creator of WCPUID), and the next revisiou would have full Barton support. However, it is still possible to examine the CPU through WCPUID ? albeit without proper identification of the CPU.
(WCPUID screen)
From WCPUID we can find that Barton?s CPU ID is 6-10-0, rather than the TBred-B?s 6-8-1. Also note the lower multiplier of 11X ? future Bartons will use this multiplier as a baseline. From our examination we can see that this Barton?s actual clock speed is 1833MHz, that its multiplier is 11X (as mentioned earlier), and that it has 512K of L2 cache, double that of the original Athlon XP. In conjunction with the 128K of L1 cache, Barton has a total of 512+128 = 640K. AMD?s supporters can at long last claim that even the Pentium 4 has less cache than the Athlon XP: (The Pentium 4 has 32K of L1 cache and 512K of L2 cache, totalling 544K.)
What about other relevant information?
In WCPUID we can see a great deal of cache information. The Barton?s main difference in this regard is the larger L2 cache. Comparing the information screens, one can see that the only difference between the TBred-B and the Barton is the L2 cache size.
(Picture of Barton?s cache info in WCPUID)
(Picture of TBred-B?s cache info in WCPUID)
This cache information is corroborated by ZDNet?s ZDBenchmark:
(Picture of ZDBenchmark showing Barton cache amount)
New QuantiSpeed rating definitions
From Barton we can see a great deal of new features, including 512K L2 and a 166MHz FSB ? but apart from that, the QuantiSpeed rating has become more bizarre:
The earlier Athlon XPs were named based on 133/2 = 66MHz increments (i.e. a 0.5X jump in multipliers). This meant that a 12.5X Athlon XP was dubbed a 2000+, and a 13X Athlon XP a 2100+. This was (at the time) a very clear rating system, but this changed with the arrival of the 2400+ and 2600+: the QS jump went from the 2200+ (1800MHz) to 2400+ (2000MHz), missing two QS ratings while the actual clock speed increased by 200MHz ? according to the old method, the 2400+ would run at 1933MHz rather than 2000MHz. In other words, 2000MHz is now being used as a new baseline for the old method of QS calculation ? the 2400+? actual clock speed is 2133MHz.
The 2400+ and 2600+ still use a 133MHz FSB but are based on the Thoroughbred core rather than the original Palomino, and use a 0.13 micron process ? perhaps AMD is planning to release a 2800+ and 3000+. This, however, is not the case: AMD increased the FSB to 166MHz and released two new CPUs: the 2700+ and 2800+ (at 2167MHz and 2250MHz respectively), meaning that 166/2 = 83MHz and a 0.5X multiplier increment separate each QS rating. These appear to be stopgap measures before AMD can release its Athlon64 and Opteron processors, and require relatively few changes to implement. However, this may lead to a strategic problem when competing with the Pentium 4: with Hyper-Threading technology, clock speeds of up to 3.06GHz and constantly decreasing prices leading to increased user adoption, it will be very hard for AMD to compete ? especially when the Pentium 4 has almost completely migrated to a 533MHz FSB.
However, AMD isn?t only using this tactic on the 2800+, and 166MHz FSB version of slower processors are beginning to appear: the 2700+, 2600+ (at 2083MHz ? 83MHz and a 0.5X multiplier decrement behind the 2700), and more recently the 2400+ at 2GHz ? although the QS rating is 2400+ rather than 2500+.
Why this situation? Did AMD neglect the 2500+ rating, or is it being put to another use? The answer is in front of you: The Barton (at 1833MHz) has a QS rating of 2500+. If you increase the multiplier, the 2800+ and 3000+ QS ratings emerge. It is probable that these QS spaces will be occupied by future Bartons (although it is also possible that there will be faster TBred-Bs on the horizon). Nonetheless, is it possible to clarify AMD?s current QS situation?
(Table showing AMD?s QS ratings. Rows as follows
Model Number
Actual clock speed, MHz
FSB (DDR)
FSB
Multiplier
Footnotes:
*Has not been officially announced ? FSB may increase to 200MHz (400DDR).
?Two versions of this processor may exist simultaneously, with different amounts of cache.
Sneak Peek: Barton?s actual performance
Having said so much, it would still be difficult to imagine the Barton 2500+/1833MHz?s performance without actual tests. The test platform will be based on an Epox 8K9A2+, which should hopefully give more overclocking headroom. Corsair XMS3500 RAM will be used. Other specs as follows:
(Spec list)
Benchmark results may vary due to environmental or equipment factors ? your mileage may vary
RAM settings are by SPD, BIOS settings are predominantly the default factory settings.
We?ve tested the CPU with WCPUID. However, as the mother overclocks the FSB slightly by default, the aggregate clock speed will be a bit higher as well ? but this should not significantly affect the results. The reference CPUs were benchmarked on the same platform.
In our benchmarks we will not only explain Barton?s performance but compare it to a reference CPU: The Athlon XP2400+/166 (AMD?s next 166MHz FSB CPU, based on the TBred-B core). As the CPU is multiplier unlocked, it can be set to the Barton 2500+? actual clock speed of 166*11 = 1833MHz. In this way we can determine the cache?s affect on performance.
(Picture: ?Athlon XP 2500+? POST)
Let?s have a look at some benchmarks:
(Table: Benchmarks of Barton 2500+)
Overall performance is quite good, especially for a 1833MHz actual clock speed rather than a 2GHz version. We also discover from our tests that the operating temperature is not high (about 30 degrees Celsius) when using an Arkua 7228 HSF.
From SiSoft Sandra 2002?s comparisons we can see the Barton 2500+? performance level: The CPU is slightly better than the Athlon XP 2200+, but a battery of tests reveals that in Sandra, the extra 256K of cache does not correlate to a marked performance increase, something that will be shown again in our overclocking tests.
(SiSoft Sandra 2002 benchmarks)
Overclocking Barton to the limit
It would surely be a disappointment to only show Barton?s performance at its default settings ? it wasn?t easy to obtain this CPU, after all. It?s, therefore, well worth looking at Barton?s overclocking performance as well.
The testing platform was the same as that used for earlier benchmarking, except with a higher Vcore (up to 1.90V) and Vram (up to 2.8V) where necessary when overclocking. Some readers may wonder if further voltage boosts are possible for further overclocking ? indeed this is possible, but as Bartons are not for sale yet, I simply wouldn?t be able to pay for it. Besides, under air cooling 1.90V is far beyond the safety limit of 1.85V imposed by many motherboards ? which is why we have imposed these limits on our testing.
Also, at the very end we test the CPU with a 200MHz (400DDR) FSB. However, with the default 11X multiplier booting into Windows is impossible, so we have tested with a 10X multiplier. It?s worth mentioning that Barton was originally meant to be launched with this FSB speed along with the upcoming VIA KT400A and SiS 746FX. However, this preview showed a Barton with 166MHz FSB and its actual clock speed marked on the CPU. Does this represent a return to actual clock speeds for AMD, or is this limited to engineering samples only? This is more difficult to establish, but AMD simply wouldn?t be able to compete against Intel?s marketing of a 800MHz FSB. A 200MHz FSB Barton is an eventuality: with DDR400 boosting overall system performance and Intel?s 800MHz FSB waiting in the wings, AMD will be forced to release a faster Barton ? or release Athlon64 ? in order to survive. One shudders at the consequences should this not occur.
Let?s first look at the BIOS POST screens:
(2000MHz, 12X multiplier)
(2083MHz: ?2800+? should not appear as it has not been officially announced. The existing 2800+ runs at 2250MHz, so it is probable that this will be the next Barton speed increment)
(2166MHz, 13X multiplier ? the same as the current 2700+)
(An astounding ?3000+? QS rating at only 2250MHz, the actual clock speed of the existing 2800+. An Athlon XP 3000+ has not yet been announced.)
(According to the roadmap there should be Bartons at 200MHz: we will test the CPU?s stability at this FSB speed)
From the POST onwards we can see that motherboard makers have had the relevant information for quite a while ? we expect a response from them soon. In fact, AMD?s announced roadmaps quite clearly indicate upcoming CPU versions. However, due to delays this roadmap should only be used as a reference:
(AMD roadmap)
Let?s have a look at some test data:
SuperPi-1M (lower is better)
We can see that overclocking to 2250MHz/3000+ gives the best results ? the CPU blazes through the 1M test in a mere 51 seconds. The 2GHz CPUs are within a second of each other, despite their FSB speed difference. The 2GHz/200 and 2083MHz CPUs both finish the test in 54 seconds.
CPUMark99 (higher is better)
In CPUMark99 we find that the 200-point is broken only at 2166MHz and above, and at 2250MHz a score of 208 is obtained. The 200MHz FSB achieves better results, scoring 191 at 2000MHz (within striking distance of the 2083MHz test results).
3DMark 2001 SE Build 330 (higher is better)
Similarly, in 3DMark 2001 SE we can see that the 2250MHz clock speed gives the best results. Importantly, the 2000/200MHz clock speed places second, better than the 2166MHz setting. It appears that boosting FSB speed gives a tangible performance increase.
PCMark2002 Pro-CPU: (higher is better)
We?ll only discuss the CPU here: looking at the 2250, 2166, 2086 and 2000MHz clock speeds, the speed grades have a 4% performance difference, and the 2000/200 is between the 2083 and 2000/166 in terms of performance.
Content Creation Winstone 2002 (higher is better)
Even the 2000/166 clock speed gives a score of nearly 40, and clock speeds above that break the 40 barrier. The 2250MHz setting gives a score of 42.2, while the 2000/200 performs slightly better than the 2083.
Business Winstone 2001 (higher is better)
Above 2083MHz all the clock speeds yield scores above 70, and the 2000/200 clock speed gives a score of 70.7. The fastest 2250MHz setting gives a whopping score of 72.6, a 6% improvement over the initial score of 68.6.
N-Bench v2 (higher is better)
Only by overclocking to 2250MHz does the Barton?s score surpass 3000 in AMD?s benchmark suite. However, by 2166MHz the score approaches 3000. Compared to the 1833MHz stock clock speed, the overclocked CPU has an astounding performance boost of nearly 20%. The 2000/200 overclock gives a score similar to that of the 2083MHz overclock.
SiSoft Sandra 2002 SP1 (higher is better)
SiSoft Sandra features two main types of CPU tests, which are shown in the table. According to Sandra, the 2083MHz overclock gives a P-rating of over 3000. However, the difference between the 2000/200 and 2000/166 clock speeds is not apparent: there is a slight (almost negligible) score increase. The 2166MHz overclock breaks the 6000 and 3000-point barriers. Obviously, the 2250MHz clock speed yields the best results.
Having looked at the benchmark results, it?s clear to see that the 2250MHz overclock gives the biggest performance boost in most situations. Unfortunately, a 200MHz FSB required the multiplier to be set down to 10X. As mentioned earlier, when the CPU was set to an 11X multiplier it would crash in Windows, despite an increased Vcore. Further, there are ways of setting the multiplier to 14X beyond 2250MHz (allowing a peak clock speed of 2333MHz). However, at this setting the PC would not boot into Windows.
Despite these incidents, we can see the possibilities for Barton. This is only the beginning: will Socket 462 develop further in the future? Currently AMD can only rely on its 166MHz TBred-Bs and the release of Barton. Without a Barton variant using a 200MHz FSB, will Socket A continue? The answer lies in the market: facing Intel?s Pentium 4 price cuts and a move to DDR333, Socket A will become obsolete if no new features are introduced. Further threats come from the more feature-packed Socket 478 and from AMD?s own Athlon64/Opteron. The outcomes of these new competitors and new challenges remain unknown, but lost territory is not easily regained. I fervently wish AMD the best in these future challenges.
(Pictures of failed POSTs using 200*11 and 166*14)
The secret L12 bridge
We mentioned earlier that there differences between Barton and TBred-B in terms of their L12 bridges. The difference can also be seen with the Athlon XP 2700+ and 2800+, which share a 166MHz FSB. Would it be possible to go straight to a 166MHz FSB by changing the L12 bridges? We decided to test this with a TBred-B Athlon XP 2200+:
To reconnect the two severed L12 bridges, we will use the same method used to unlock the multiplier. Apply correction fluid to the gap in the second bridge from the right (the bridge we will attempt to connect), then tape the adjacent bridges to prevent short-circuiting. Then connect the bridges using a conductive silver pen.
After connecting the bridges, boot up the computer to test. One major worry is that the PC may not boot if it has a high multiplier setting. If at all possible, unlock the Athlon XP?s multipliers as well to prevent booting problems.
In this case the PC booted and the BIOS quite clearly shows a 166MHz FSB. A simple L12 connection to unlock 166MHz FSB speeds seems to be well worth the effort ? if you choose to unlock the multiplier, unlock this higher FSB setting as well to give a significant performance boost.
(POST Screen: ?Boots with 166MHz FSB, but fails to recognise the multiplier as it has been unlocked?)
(Adjusting the CPU multiplier to 11X in BIOS)
In our tests we changed the clock speed to 1833MHz. Given the Athlon XP 2200+? inherent performance and its TBred-B core, the CPU seems to have significant overclocking headroom and works with a 11.5X and even 12X multiplier. Beyond that ? well, that?s for you to try...
Message édité par gliterr le 23-12-2002 à 09:08:39
![[:yamusha]](https://forum-images.hardware.fr/images/perso/yamusha.gif "[:yamusha]")
bon je vais me recoucher moi je raconte trop de connerie. C'est npas de ma faute je suis désepéré à force d'attendre ce putin de proco.
![[:mmmfff]](https://forum-images.hardware.fr/images/perso/mmmfff.gif "[:mmmfff]")
![[:mlc]](https://forum-images.hardware.fr/images/perso/mlc.gif "[:mlc]")
![[:xfred4]](https://forum-images.hardware.fr/images/perso/xfred4.gif "[:xfred4]")
. Mais, je me suis dit que si les Bartons sont compatibles avec toutes les CM suffisantes pour les supporté, un p'tit coup de pouce sur les AMD-760 ne serait pas de trop. Je ne sais pas du tout ou cela en est pour le support des derniers XP (officiel ou pas), mais si c'est possible, ça ne fera que du bien.
