For a relatively long period of time AMD, which tries to compete with Intel not only in the performance CPU market, but also in the value PC sector, couldn't represent any real threat for the rival even despite its progressive architectures. The failures in the value PC market had to do basically with the absence of Socket A chipsets, which could allow building really low-cost systems. However, it never meant that AMD didn't try to win a part of the market from Intel Celeron processors. Right on the contrary, AMD tried several different methods here. If they couldn't boast the lowest cost of the systems on Duron processors, then the performance was their trump. The company repositioned their value CPUs as the best low-cost solution for performance systems. That is why from the very beginning AMD Duron family had always been ahead of Intel Celeron in terms of core clock frequencies as well as in terms of performance, defeating the competitor working at the same clock frequency. It is due to this particular tactics as well as to low processor costs that AMD Duron managed to win a significant part of the market.
Time passed and AMD's problems with the absence of low-cost Socket A chipsets started getting solved little by little. Duron turned more and more popular. Now it was the time for Intel to express concerns. As a result, Intel made two significant moves in order to increase the rating and competitiveness of its Celeron processors. Firstly, Celeron CPUs acquired a faster 100MHz bus instead of the 66MHz one with relatively low bandwidth. Secondly, these CPUs have recently got a new core, which put them very close to Pentium III family. Now the new Celeron processors with 1.2GHz frequency feature not a 128KB L2 cache but a 256KB one. At the same time, Intel has significantly increased the working frequencies of its low-cost processor family, so that they have even surpassed AMD Duron in terms of clock frequency by now. However, as far as performance goes, all these improvements didn't have any tangible effect and Duron remained the leader.
Nevertheless, AMD also didn't keep its hands in pockets. Even though Duron CPUs based on Spitfire core are still quite competitive, all the processors starting from 1GHz core clock are based on a new more progressive Morgan core. What forced AMD to replace the processor core in the Duron family? It is most likely to be the aspiration for unification. The shift to a new core in the value processor family took place together with the shift to a new Palomino core in the Athlon family. Bearing in mind that Morgan and Palomino, like Spitfire and Thunderbird, do not have many architectural differences (only the L2 cache size is different), this move doesn't seem so unjustified any more. Especially since Morgan, just like Spitfire, is manufactured with 0.18micron aluminum interconnect technology, and hence the new core doesn't require any production lines re-equipment.
Now that we have found out what pushed AMD to the change of the Duron core, let's see what the advantages of the innovation will be. At first, let's take a look at the specification of the new Duron on Morgan core:
- The core is known as Morgan. It is manufactured with 0.18micron aluminum interconnect technology in Fab 25 (Austin).
- 128KB L1 cache (64KB for data and 64KB for instructions); 64KB on-die L2 cache working at the full processor frequency. L2 cache is exclusive.
- 1.0GHz and 1.1GHz models available.
- EV6 system bus working at 200MHz. Socket A physical interface.
- Supports 3DNow! Professional instructions set (107 SIMD instructions).
- 106sq.mm die size comprising 25.2 million transistors.
You can see the Morgan based CPU on the left and the one on Spitfire core on the right photo.
It is basically the L2 cache size that allows to distinguish between the new Morgan core and the Palomino one, which is used in AMD new performance CPUs (Athlon XP for desktops, Athlon MP for dual-processor servers, and Athlon 4 for mobile PCs). Processors based on Morgan core feature 64KB L2 cache, while Palomino based ones boast 256KB L2 cache. Moreover, unlike Intel, which used the same die for its Pentium III and Celeron processors on Coppermine core and simply disabled a half of L2 cache in Celerons, AMD uses (and used) two physically different dies for its value and performance solutions. That is why the die size of Morgan based Duron processors is 30% smaller than that of Palomino based Athlons.
The second difference between Morgan and Palomino, which seems to come from pure marketing, is the different system bus frequency. Palomino processors work with 266MHz bus providing 2.1GB/sec, while Morgan works with 200MHz bus featuring 1.6GB/sec. However, this is mainly a "paper difference", as there is nothing in Morgan's architecture, which could prevent it from working with 266MHz bus.
As a result, the innovations introduced in Morgan core compared with the older Spitfire used in Duron CPUs, do not actually differ from the innovations introduced in the new Palomino compared with the older Thunderbird. There is only one exception, which we are going to discuss a bit later. As for the changes made to the new Palomino, we have already discussed them in great detail in reference to Athlon MP in our review of Dual-Processor Socket A Systems on AMD 760MP Chipset. That is why here we will simply repeat the key things.
- Morgan core supports 3DNow! Professional instructions. The previous Duron models on the older Spitfire core supported a smaller 3DNow! set. A new bigger set of SIMD-instructions called "Professional" includes 52 new instructions providing compatibility between 3DNow! Professional and SSE set supported by Pentium III, and now by Athlon XP as well. Unfortunately, AMD didn't include in its CPUs the SSE2 unit implemented in Pentium 4 processors, however, the future AMD products have every chance to feature SSE2 support.
- Morgan core is provided with Data Prefetch Mechanism. The main idea of this mechanism is quite simple: the CPU tries to predict in advance what data from the main memory it may need in the near future and saves this data in the cache then. If the mechanism works correctly, then the future data processing may take considerably less time. This way, data prefetch mechanism allows Morgan to use the processor bus and the memory bus more evenly, smoothing the peaks and increasing the workload during idle periods.
- Morgan based processors feature larger Translation Look-aside Buffers (TLB). It serves to cache the translated physical memory addresses. Translation procedure takes place every time the CPU addresses the data stored in the main memory that is why caching these addresses helps to significantly reduce the delay before the requested data is given out to the CPU.
- Morgan core features a built-in thermal diode. It allows monitoring the processor's physical state and protect it against overheating.
Well, if this review were devoted to Palomino, we would have to mention one more advantage of this core over the predecessor, namely the lower heat dissipation, which let achieve higher working frequencies. Although Morgan is nearly the same Palomino, its heat dissipation has hardly become any lower compared to the predecessor. For example, the maximum heat dissipation shown by Duron 950MHz based on the older Spitfire core makes 41.5W, while the maximum heat dissipation demonstrated by Duron 1.0GHz on the new Morgan core makes 46.1W. The same ratio is valid for the typical heat dissipation values, which makes 37.2W for Duron 950MHz and 41.3W - for Duron 1.0GHz.
So what's the matter? Though the whole thing seems pretty strange at first sight, the explanation is very simple. The Morgan yields appeared somewhat lower than those of Spitfire. Therefore, AMD decided to improve this parameter at the expense of some hidden potential. Namely, taking into account that Spitfire Vcore was equal to 1.6V, while Athlon processors on Thunderbird and Palomino use 1.75V Vcore, AMD decided that they could possibly increase Morgan's Vcore up to the same 1.75V. As a result, the 20% reduction in heat dissipation (with the same 1.6V Vcore), which took place due to the core redesign, was compensated by the following increase in Vcore by 0.15V (up to 1.75V). We should also point out that high heat dissipation is not such a serious problem for Duron CPUs, as it is for Athlon processors. So, Duron Vcore increase is quite justified due to smaller die size.
This way, Morgan can boast two improvements compared to Spitfire, namely: larger TLB and Data Prefetch Mechanism, which should provide a significant performance increase compared with the older core even at similar working frequencies. In fact, AMD is not going to produce Duron processors based on different cores but supporting the same working frequencies, as they are planning to use Spitfire core for the CPUs working at the frequencies below 1.0GHz and Morgan core for the frequencies above 1.0GHz. Nevertheless, it is still of certain theoretical interest to compare the performance of the two cores working at the same clock frequency. We made this kind of comparison having overclocked our Duron 950MHz on Spitfire core up to 1GHz and compared its performance with that shown by a "real" Duron 1.GHz on Morgan core:
|Spitfire 1.0GHz||Morgan 1.0GHz||Performance gain|
|Business Winstone 2001||39.5||40.5||2.5%|
|Content Creation Winstone 2001||51||52.9||5.7%|
|Quake3 Arena (four) |
|Unreal Tournament |
As it follows from the table above, the performance difference between the two cores is not that significant and makes 3%-5% depending on the applications. The Morgan's advantage over Spitfire is most noticeable in those tasks where large blocks of sequential data are processed. This effect most likely results from the work of the Data Prefetch Mechanism, which prepares the necessary data in the processor cache well in advance. However, you shouldn't disregard the fact that in those applications, which use SSE instructions (and do not support 3DNow!) the new core may appear much more efficient, because Morgan boasts an SSE unit, absent by Spitfire.
Summing up our story about Morgan and its innovative features, we would like to point out that Duron processors on this core can be used in the same mainboards as their predecessors on Spitfire core. The only thing required to ensure smooth functioning of the new Duron CPUs is an updated BIOS version. Note that almost all the mainboard manufacturers have already made the corresponding BIOS updates.
Testbed and Methods
First of all, we would like to point out that the comparative testing of all processors for low-cost systems provided within this article was carried out only on platforms supporting PC133 SDRAM. As the systems equipped with this memory type do not fall too far behind those with DDR SDRAM in performance, and the cost of the DDR memory is still almost twice as high as that of PC133 SDRAM, we consider it quite logical that low-cost systems should be equipped with cheaper PC133 memory.
That is why we tested AMD Duron CPUs on VIA KT133A based mainboards and Intel Celeron CPUs - on i815 based mainboards. Moreover, to get a complete picture of the low-cost processors performance we also included some results for Intel Pentium III and AMD Athlon CPUs working at the same core clock frequency. In order to let you compare the performance of the new Morgan core with that of the predecessors, we included the results for Duron (Spitfire) 950MHz, which is the eldest currently available model on the older core.
Besides, there is one more system included into the testing set. As you remember, we wrote in our Pentium 4 Chipsets Comparison:
This way, building an i845 based system makes sense only if you are hunting for a low-cost Pentium 4 based system, and its performance is not a determinative… In games i845 is out and away the slowest chipset of all, so if there is anything equal to i845, these are Celeron and Duron based platforms.
Well, we believe it is the right time now to prove this statement on practice. So, we decided to include the results for the today's youngest Socket478 Pentium 4 processor working at 1.5GHz. Especially, since this configuration costs about the same as systems built on the today's fastest Intel Celeron processors, according to Intel's products positioning scheme.
And now we would like to say a few words about the mainboards we selected for our testbeds. For an i815 based mainboard we chose ABIT ST6 built on i815 B-stepping chipset. Another board from the same manufacturer, ABIT BL7, represented i845 based solution. However, as for VIA KT133A, we failed to get a mainboard from ABIT here. The tests run in our laboratory showed that most mainboards on this Socket A chipset from VIA supporting PC133 memory cannot work stably with the new NVIDIA Detonator XP drivers. Basically, this leads to unusually low performance in 3D applications of NVIDIA GeForce3 based graphics cards installed into this kind of systems on VIA KT133A. Unfortunately, neither the new Detonator 21.85 drivers, nor the new VIA Service Pack 4.34v helped to solve this kind of problem. The only mainboard on VIA KT133A of the bunch we tried, which appeared free from this trouble was ASUS A7V133. So, we decided to take it for our investigation.
So, here are the configurations assembled for our tests:
|AMD Duron||AMD Athlon||Intel Celeron||Intel Pentium III||Intel Pentium 4|
|CPU||Duron 950MHz |
|Athlon 1.0GHz||Celeron 1.0GHz |
|Pentium III 1.0GHz||Pentium 4 1.5GHz|
|Mainboard||ASUS A7V133-C (VIA KT133A)||ABIT ST6 (i815 B-step)||ABIT BL7 (i845)|
|Memory||256MB PC133 CL2 SDRAM|
|Graphics Card||Gigabyte GV-GF3000DF (NVIDIA GeForce3)|
|HDD||IBM DTLA 307015|
All the test systems featured Microsoft Windows 98 SE.
As you can notice, the CPUs tested do not include the recently announced Celeron 1.2GHz, which is the first CPU in the family based on the new Tualatin core. To our great disappointment, these CPUs weren't available at the time we ran all these tests. However, the new Celeron undoubtedly deserves our attention, so we will very soon offer you another review, don't miss it! :-)
Actually, office applications should represent the main field for low-cost processors. However, they have now become so fast that any of these applications run quite OK on systems with all the value processors. So, this test has purely theoretical value.
New AMD Duron processors proved more than worthy here. They managed not only to beat their competitors from Intel family, but also to outperform Intel Pentium 4 1.5GHz with PC133 SDRAM. We should also mention that high results shown by Duron processor are not limited by the seemingly small 64KB L2 cache. Nevertheless, you should understand that due to exclusive L2 cache structure in AMD Athlon and Duron CPUs, the data stored in L1 cache is not duplicated in L2 cache. In other words, with the 64KB intended for data in L1 cache, the efficient caching works for 128KB of data. It is exactly as much as Intel Celeron can cache, even though its L2 cache is twice as big as that of AMD Duron. However, unlike the competitor, it has to duplicate the L1 contents.
In content creation applications AMD Duron proved even better. It is especially impressive that Duron 1.1GHz managed to outperform Athlon 1GHz. This fact can be explained partially by the Data Prefetch Mechanism of the new Duron CPUs, which is absent by the regular Athlon processors. Due to this mechanism, the system uses its memory bus more efficiently, which can't help telling on the performance of applications requiring high memory bus bandwidth. The tasks included into Content Creation Winstone 2001 package belong to this type of tasks, of course.
SYSmark2001 is more loyal to Intel processors. That is why the situation here is totally different from what we saw in the previous cases. For instance, Duron CPUs here fall behind Celerons working at the same core clock frequency. Moreover, even the new Morgan core doesn't help to improve the state of things. Why so? Well, take a look at the next diagram to understand the reasons:
Duron processors owe their failure in the previous benchmark to that part of SYSmark2001, which deals with Internet Content Creation. To be more exact, to Windows Media Encoder 7.0 included into this benchmark. This application works with Intel processors much faster than with AMD ones as it is optimized for SSE and SSE-2 instructions. However, even though the new Durons on Morgan core do support SSE instructions set, Media Encoder 7.0 doesn't use the SSE unit when working on Duron based systems, because of some bug in the recognition procedure of the supported SIMD-instructions. There already appeared a special patch for this test, which allows forcing the Media Encoder to use SSE instructions with the new AMD CPUs. However, this patch can't eliminate the problem completely, because it corrects the work of Media Encoder 7.0 application only within the benchmark. In real conditions, this application still doesn't involve SSE unit when working in systems built on AMD CPUs. So, the best thing to do is to wait for the new Media Encoder version to come out. In one of our upcoming reviews we will see how greatly Morgan's performance in SYSmark2001 will rise once it is forced to use SSE instructions.
As for the office applications, there are no hidden obstacles for Duron processors that is why they are as fast as usual and easily surpass Celeron competitors.
To make the office coverage complete, we measured how fast the systems could archive the data packs of big size (directories with installed Unreal Tournament) with the maximum compression, which loads the CPU quite heavily. For this purpose we used a widely-spread WinZIP archiving utility. So, the smaller value on the diagram denotes higher performance. As our experiment showed, new Duron CPUs prove simply excellent here, falling only behind AMD Athlon with four times as large L2 cache and Pentium 4 with high-speed ALU working at twice the CPU frequency.
We have also taken the trouble to test the performance during the encoding of DVD graphics data stream into DivX MPEG-4 format. And again thanks to Data Prefetch Mechanism, the new Duron (Morgan) performs very well. Moreover, the preliminary caching of the selected data allowed Duron 1GHz to leave behind even Athlon 1GHz processor.
In Quake3 Arena the laurels were won by the CPUs with larger cache size. However, Duron (Morgan) remains at the head of all the value processors.
In Unreal Tournament Duron 1.1GHz manages to outperform not only the entire Celeron family but even Pentium 4 1.5GHz working in i845 based system.
We believe you don't have any more doubts about AMD Duron processors being faster in games than Intel Celeron. But nevertheless, let's take a quick look at the results obtained in a popular DirectX 8.0 benchmark - 3DMark2001.
In 3DMark2001 the performance difference in various tests wasn't constant. However, Duron managed to beat Celeron everywhere and in Dragothic benchmark new Duron CPUs turned out even faster than Pentium III 1GHz.
Disabling hardware T&L unit results into the fact that all the geometry and lighting calculations are carried out by the CPU using SIMD-instructions. Moreover, 3DMark2001 recognizes SSE support in the new Duron 1GHz and 1.1GHz correctly and uses the corresponding SSE unit for the calculations. So, new Duron processors start looking even more attractive leaving behind not only all Celeron processors but also Pentium III 1GHz and sometimes even Pentium 4 1.5GHz as well as Athlon 1GHz.
In conclusion, we decided to add one more benchmark, which allows estimating the CPU performance in real scientific tasks. Science Mark V1.0 we used for this purpose measures the processor speed when solving real math1ematics modeling tasks.
Science Mark V1.0 measures the time each CPU requires to solve three real physical tasks: to calculate the general energy of the H20 molecule with the help of the Monte-Carlo method, to solve Shredinger equation for all the 61 electrons of Promethium, and to model all 216 Argon atoms at 140K. Besides, the test also includes some linear algebra tasks with the vectors and matrixes of large size. Now wonder that AMD processors look much more attractive in this test. One of AMD Athlon and Duron true advantages is floating point operations, which allows Duron to rush ahead of all the Intel's competitors.
The diagram above shows the time each CPU required for the particular tasks from the Science Mark V1.0 benchmark set. So, smaller value stands for higher performance. These results show that larger L2 cache may not matter at all for some scientific tasks. That is why using Duron processors with the powerful FPU may turn out a good choice for scientific researchers.
Every time a new CPU on a new yet unknown core comes out, overclockers hold their breath and pin a lot of hopes upon it. It is quite natural, actually. Every core has some potential, which allows it to support clock frequencies raised up to a certain limit, determined by the manufacturing technology and design features. When the processor family based on the new core is just starting to spread over the market, the potential of the new core is not yet used by the manufacturer to the full extent, leaving more opportunities for overclockers. That is why the youngest CPU models are usually the best choice for overclocking needs, as their core frequency can be increased really significantly. For example, the most overclockable Duron processors on the older Spitfire core could reach up to 1GHz core clock frequency. As a result, taking the youngest 600MHz models some guys managed to get an over 70% frequency increase or even more. Elder CPU models usually boast the same top frequency or maybe a slightly higher one that is why overclocking them is not as efficient.
Of course, a lot of people looked forward to the new Duron on Morgan core, because these CPUs were supposed to restore the fame of the Duron family as of the "overclocker's dream". Moreover, the new core is quite of interest because Athlon XP processors featuring similar architecture work at some 1.5GHz. However, in this case one very important issue appears disregarded. New Duron CPUs, just like the older ones, are made on Fab 25, where no copper is used. That is why their heat dissipation and top core clock cannot be compared to those of the "copper" Palomino. So, Morgan is a core, which cannot be used at the frequencies as high as those supported by the Palomino core, because of aluminum interconnect technology involved in its production.
According to the current roadmap, AMD is going to launch just a few Duron (Morgan) processor models. The maximum core frequency of the last model is to make 1.2GHz, and then they are planning to move to 0.13micron Appaloosa core. That is why please don't expect your new Duron (Morgan) CPU to overclock well over 1.2GHz.
Well, now let's have some practice here. We tried overclocking AMD Duron 1.0GHz AHHAA series produced during week 30 of this year. To find out what is the maximum core frequency this CPU can work at, we increased the clock frequency multiplier. We should point out that both: older Duron processors on Spitfire core and the new ones on Morgan core come with the locked clock multiplier. However, the good old methods of unlocking it still work even though the Golden Bridges are located a bit differently now. Having closed L1 bridges (see this article for the ways to unlock CPU clock multiplier) and increased Vcore up to 1.85V, we managed to set the multiplier to 11.5x. Further multiplier increase resulted into unstable performance of the system. After that we managed to additionally increase the FSB frequency up to 103MHz, so that the maximum frequency our Duron 1.0GHz managed to work at equaled 1184MHz.
Well, new Duron processors on Morgan core are another evolutionary step forward of the value processor family by AMD. In fact, there seem to be no revolutionary innovations besides SSE support, which could speed up the newcomer significantly. Of course, increased TLB and Data Prefetch Mechanism do affect the Morgan's performance positively. However, this increase doesn't exceed 5%. Nevertheless, due to more progressive architecture, which has been its trump from the very first CPU models, Duron is still much faster than Celeron, Intel's value processor. Moreover, in some cases elder Duron processors work even faster than Pentium 4 1.5GHz in i845 based systems. And this is not the top yet. Further reduction of DDR SDRAM prices should inevitably let us use this memory in low-cost systems one day, so that Duron performance will grow even more. So, in the meanwhile, AMD has no causes for concern in the value market sector.
In conclusion we would like to say that to overclockers' disappointment Morgan's potential appeared quite low, which makes us consider the new core an intermediate stage between the older Spitfire and the new 0.13micron Appaloosa, due in mid 2002.