by Ilya Gavrichenkov
11/25/2004 | 02:25 AM
The current year 2004 turned out the year when the major processor developers, AMD and Intel managed to master a more advanced 90nm production technology. Intel performed the transition to this finer manufacturing process in winter already, together with the launch of the new Prescott core for the Pentium 4 processor family. AMD is just starting the mass production of the Athlon 64 mobile processors manufactured with 90nm technology, and since September 15, 2004 we see the desktop Athlon 64 processors based on the new 90nm cores. This way, AMD managed to introduce the 90nm technology without any visible delays, according to their initial schedule.
When Athlon 64 processors with 90nm cores appeared, people pinned a lot of hopes and anticipations upon them as well as quite a few concerns. While some enthusiasts expected that with the introduction of the new 90nm production technology AMD would be able to increase the frequency potential of AMD Athlon 64 processor family and enrich the list of its features, the others were pretty concerned with it, believing that this technological advances would cause certain problems with the leakage current, just like Intel had during Prescott production.
Luckily, we managed to get hold of one of the mass Athlon 64 processors from AMD based on the new 90nm core. Thanks to this fact we got a beautiful opportunity to test all the features of the updated AMD processors ourselves. In other words, this whole article is going to be devoted to Athlon 64 processors for desktop systems manufactured with the new 90nm technological process.
At first let’s take a look at the AMD’s roadmap:
As we see, AMD is planning to introduce the new 90nm production technology in all market segments this year. So far they managed to complete 2/3 of this plan: 90nm Oakville and Winchester cores are already used in mass production of mobile and desktop CPUs respectively. Oakville and Winchester cores actually boast similar features that is why today we are going to take a closer look at Winchester core primarily, because it is used for the desktop Athlon 64 CPU family.
The major distinguishing feature of the Winchester core besides the 90nm production process is the 512 L2 cache and dual-channel memory controller supporting DDr400 SDRAM. This way, 90nm Winchester core can be considered an analogue of the 130nm NewCastle core used in Socket939 CPUs in the first place. Therefore, it is not at all surprising that the first Athlon 64 processors to use the new Winchester core appeared CPUs for Socket939 platform.
AMD engineers didn’t do any redesigns when working on the Winchester core. In fact, Winchester is the same NewCastle with that only difference that it is produced with a more advanced and up-to-date 90nm technology. However, we cannot deny that there were still a few minor architectural improvements made to it, which definitely had a positive effect on its performance compared with the performance of the predecessor, as we will see later during the benchmark results discussion. However, AMD doesn’t try to draw all the attention to the improvements of the 90nm core. The new D0 core stepping is remarkable primarily by the fact that it allows AMD to considerably reduce production costs.
Everything can be explained fairly simply. 130nm NewCastle core is 144sq.mm big. The die size of the new Winchester core shrank down to 84sq.mm due to the finer 90nm manufacturing technology. This way, the use of the 90nm process allowed AMD to design a 42% smaller processor core. All old NewCastle cores and new Winchester cores are currently manufactured in Dresden Fab30, which uses 200mm wafers. So, the new production technology allows obtaining 77% more cores per single wafer. In other words, the production cost of the Winchester core is ideally 44% lower than that of the least expensive 130nm core for the Athlon 64 processor family. However, this significant result can only be achieved if the yields per wafer are the same for 90nm and 130nm technologies. Of course, this is not yet the case. I really doubt that the new production technology can actually ensure the same high yields as the old 130nm process. In fact, AMD seems to be happy with the current yields for their new 90nm cores. The automated precision manufacturing methodology (APM) used on the Dresden Fab30 today allows polishing off the new technology within a very short period of time, so that already today the company can really benefit from selling the new slower CPU models manufactured with the 90nm process.
Since the current situation in the processor market doesn’t force AMD to increase the working frequencies of its processors that rapidly, AMD decided to start introducing 90nm Winchester cores into the slower processor models first, and then move over to the fastest ones. Especially, since there is a clear economical reason for this strategy. The first Winchester based processors are designed for Socket939 systems, support working frequencies from 1.8GHz to 2.2GHz and are rated as 3000+, 3200+ and 3500+. In other words, AMD managed to significantly enlarge the lower end of Athlon 64 processor family for Socket939 systems with the introduction of the new 90nm core. Now the least expensive CPUs from this product line cost around $150, which should make them much more popular.
Besides the reduction of production costs of the Athlon 64 processors, the launch of the 90nm technology allowed to improve their heat generation rate. While the 130nm NewCastle processors generated around 89W of heat, the 90nm Winchester solutions boast a much lower value of only 67W. This way, AMD managed to avoid leakage current issues, which caused Intel a lot of trouble during 90nm Prescott production: as you know the heat dissipation of Pentium 4 Prescott processors got notably lower than that of 130nm Pentium 4 Northwood CPUs.
This pretty impressive success of AMD Company, when the engineers managed to achieve a 25% heat dissipation drop by simply shifting to a finer 90nm production process has something to do with the nature of this technology. For the production of 90nm cores, just like to the production of 130nm cores, AMD uses SOI technology (silicon-on-insulator). As far as the strained silicon technology used for Intel’s 90nm processor is concerned, AMD refused to go with it yet. However the second-generation 90nm production process, which should be put into life next year, will include both: SOI and strained silicon. AMD engineers are very excited about this combined technology aka Strained Silicon Directly on Insulator (SSSDOI), since they expect it to open up new opportunities for further clock frequencies increase.
Moreover, in the upcoming E0 core stepping, which are due next year, AMD should introduce a few additional features. First of all, these processor dies will support SSE3 instructions, which have been implemented in Pentium 4 Prescott processors from the very beginning. Secondly, there will be 4 write combining buffers in the upcoming 90nm Athlon 64 cores instead of 2. And thirdly, AMD is planning to improve their power consumption. The today’s Athlon 64 based on Winchester D0 core stepping is similar to Athlon 64 on NewCastle CG core stepping.
In order to specify all the above discussed points, we would like to offer you a comparative table with the major features of the various K8 processors available in the today’s market:
As for the naming of the currently shipping AMD Athlon 64 processors, here is one more table summing up all the available models for your convenience. Below you can find a list of all OPN codes and the major specifications of the currently selling Athlon 64 processor models:
Although AMD is not yet using the new 90nm Winchester cores for the top Athlon 6 processor models, it doesn’t at all mean that the frequency potential of this core is lower than that of the 130nm NewCastle. Maybe the introduction of the new processor core in the low-end models of the family in the first place is justified by purely economical reasons described above. Therefore, when we managed to get hold of the Winchester based CPU we first of all decided to test its overclocking potential.
So, for our experiment we used Athlon 64 3500+ based on the 90nm core.
The diagnostic utilities launched when this CPU worked in the nominal mode provide the following report:
As we see, the nominal working frequency of this processor is 2.2GHz, and its default Vcore is 1.4V, which is 0.1V less than by Athlon 64 CPUs based on 130nm cores.
To test the overclockability of this CPU we assembled a test system including the following equipment:
Since we didn’t aim at revealing the maximum frequencies one can get from this CPU with the help of extreme cooling solutions, we used a regular boxed cooler throughout our experiments. We overclocked the processor by changing the clock frequency generator settings, without changing the CPU clock frequency multiplier.
The testbed included DDR500 memory and a mainboard based on NVIDIA nForce3 250, which knows to lock the AGP/PCI frequencies when the clock generator frequency gets higher. Therefore, we didn’t have to worry about the factors limiting the overclocking results other than the potential of our processor.
First of all, we decided to check what maximum working frequency our Athlon 64 3500+ Winchester can achieve if the Vcore remains the same. Without much effort we managed to speed up the clock generator to 230MHz. keeping in mind that the nominal clock frequency multiplier of Athlon 64 3500+ processor equals 11x,, the overall CPU frequency grew up to 2.53GHz. Further system overclocking by continuous increase of the clock generator frequency resulted into unstable functioning of the system under workload that is why we had to increase the CPU Vcore, in order to be able to proceed.
The second part of our overclocking experiments was performed with the Vcore increased by 8.3%: up to 1.52V. This pretty insignificant voltage growth has had a great positive effect on the overclocking results, so that we could continue increasing the clock generator frequency without any stability losses. Step-by-step we managed to set a new stability limit: when the clock generator frequency reached 238MHz, the CPU remained stable, while any further frequency increase appeared impossible. This way, the maximum frequency we managed to achieve by overclocking AMD Athlon 64 3500+ on Winchester core appeared 2.618GHz.
In other words, we can state that the ultimate result is the practical ability of one of the first Athlon 64 processors on Winchester core to exceed the 2.6GHz bar with a minimal Vcore increase. Is that a lot or not? This is a good question. On the one hand, we didn’t get that greatly impressed with the frequency potential of the new 90nm Winchester core. Of course, regular Athlon 64 CPUs based on the 130nm CG core stepping cannot always exceed 2.6GHz, however, their maximum supported frequencies lie really close to this value. Moreover, Athlon 64 FX-55 processor manufactured with 130nm strained silicon technology already works at the nominal 2.6GHz. So, AMD can actually achieve the frequency of 2.6GHz without the 90nm process. On the other hand, despite the fact that the maximum working frequency of the today’s Winchester based Athlon 64 CPUs is only 2.2GHz, these processors are clearly capable of working at much higher speeds. Moreover, keep in mind that 90nm production technology hasn’t yet been fully mastered by AMD and is still under certain development and enhancement, so that we can expect the processor cores manufactured with this new technology to acquire a considerably higher frequency potential in the future. So, Winchester definitely does boast a huge potential, which will fully reveal itself later.
From this viewpoint Athlon 64 3000+ processor for Socket 939 appears a pretty good choice for an overclocker today. This CPU based on Winchester core after successful overclocking will be able to work at up to 2.6GHz frequency, which will make its performance competitive with that of top Athlon 64 models. At the same time, its price is now about $150. And the fact that it is designed for Socket 939 will allow overclockers to start assembling a very promising platform, which will be supported for a considerable while unlike Socket 754.
In fact, the transition of the Athlon 64 processor family to the new 90nm Winchester core doesn’t lead to any features changes or working frequency increase for this CPU family. However, it would be incorrect to believe that the ordinary users, who are not that much into overclocking, cannot benefit from that at all. The thing is that the new AMD processors on Winchester core feature lower power consumption and hence lower heat dissipation, which are pretty noticeable and important parameters I should say. According to the official specifications revealed by AMD, the CPUs on Winchester core generate maximum 67W of heat, which is 25% lower than the top heat dissipation of Athlon 64 processors on 130nm cores. However, what is its actual practical value? In order to answer this question we tested the new Athlon 64 CPU and compared its thermal parameters with those of the similar Athlon 64 processors but based on the older NewCastle core.
The testing was carried out on the same platform as in the previous case (during overclocking experiments). During the tests we compared the temperature of Socket 939 Athlon 64 processors on Winchester core and Athlon 64 processors on NewCastle core working at the frequencies from 1.8GHz to 2.4GHz. Note that we set the 1.8GHz, 2.0GHz and 2.2GHz frequencies with the corresponding frequency multiplier, i.e. as 9x200MHz, 10x200MHz and 11x200MHz respectively, while the 2.4GHz core clock was achieved as a result of CPU overclocking by raising the FSB frequency: 11x219MHz. The processor Vcore remained nominal in all cases, and we used the same boxed cooler for all our tests. The temperature values were measured with the help of the thermal diode built into the processor core. The core temperature was measured in two modes: idle mode and under maximum CPU workload created by S&M utility version 1.0.0 alpha.
In the idle mode the 90nm Winchester core is considerably cooler than the NewCastle core. However, under 100% CPU workload these two temperatures appear really close. At first sight this is a very unexpected result, I should say, because the 25% reduction of the maximum heat dissipation by Athlon 64 processors on Winchester core doesn’t seem to be affecting anything. However, there still is a logical explanation to this fact. The thing is that Winchester die not only started to generate less heat, but also got smaller in size. As a result, it has become a little bit more complicated to arrange proper heat dissipation from the die surface, than in case of a larger NewCastle core. This is exactly the reason why the core temperature does not characterize the core heat generation that much, but indicated the density of the heat dissipation streams, which appears even higher by Winchester compared with the predecessor. Therefore, Athlon 64 processors on the new Winchester core cannot boast a significantly lower temperature during active work.
The lower heat dissipation of the Athlon 64 CPU on the new 90nm Winchester core actually has a different effect. Namely, you can feel its influence by the cooler temperature. In Athlon 64 NewCastle based platforms the cooler warms up a lot and is warm to the touch. However, in case of Athlon 64 Winchester the cooler temperature gets significantly lower.
However, there are more facts proving that the new Winchester core of the AMD processors boasts reduced heat generation. Namely, we managed to measure the processor power consumption with the help of a multimeter. To get the value of the processor power consumption, which is the same as the processor heat dissipation, according to the energy conservation law, we actually measured the current going through the 12V processor circuit. In fact, this method is not very precise, because it doesn’t take into account the performance index of the processor voltage regulator circuit, however, it suits quite well for a preliminary rough comparison. Just like in case of temperatures measurements, we considered two situations: the idle state and the maximum CPU workload created by a special S&M utility version 1.0.0 alpha. The results of our experiments are given below:
The numbers speak for themselves. The new 90nm core consumes considerably less power than the older 130nm core in both: idle state and under heavy processor workload. Note also that the 50W power consumption demonstrated by the Winchester core at 2.4GHz core clock is a very low value for contemporary processors. This fact gives us some reason to hope that the frequency potential of the Winchester core will turn out pretty significant. For a more illustrative comparison we also measured the power consumption of the Pentium 4 processors based on Northwood and Prescott cores and working at 3.4GHz core clock. The results turned out simply impressive: under maximum workload Pentium 4 processor on Northwood core consumed about 100W of power, while the Prescott based CPU (with a C0 core stepping) required about 132W. This way, we have every right to call all Athlon 64 processors very economical solutions, and certainly this first of all refers to the new CPUs based on 90nm Winchester core.
The current test session aimed at measuring the performance of the Socket 939 processors from AMD manufactured with the newer 90nm production technology and comparing that with the performance of the already existing 130nm solutions. The results of our benchmarks will be given for all CPUs on Winchester cores and we will also compare them with the results obtained for Socket 754 solutions and Socket 939 solutions with analogous ratings. Also, we added the performance results for the competing Pentium 4 processors, so that you could get a better vision of the market situation.
The testbeds involved into our testing session were configured as follows:
The tests were run in MS Windows XP SP2 with installed DirectX 9.0c. The testbeds were adjusted for maximum performance. Note that we increased the Cycle Time (Tras) timing of Athlon 64 to 10, because according to our experience, the memory controller of Athlon 64 processor works more efficiently in this mode, than in case this setting equals 5.
According to a popular SuperPi benchmark, processors on the new Winchester core perform somewhat faster than NewCastle based ones when working at the same frequency. This is very likely to be the result of those minor optimizations in the micro-architecture of Athlon 64 core that AMD was talking about.
Similar tendency can be observed when we look at the results obtained in PCMark04. According to this benchmark, Athlon 64 3500+ on Winchester core is again faster than Athlon 64 3500+ with similar formal specifications but based on a NewCastle core. As for the performance of slower processors based on these two processor dies, we can say that Athlon 64 3200+ (Winchester) outperforms Athlon 64 3200+ for Socket 754 platform due to its dual-channel memory controller, even despite smaller L2 cache memory. The new Athlon 64 3000+ yields to a similar Socket 754 processor because of lower clock frequency.
The memory subsystem benchmark from the PCMark04 test set shows that the enhancements of the memory controller introduced in the new Winchester core are quite insignificant. The 90nm and 130nm processors are differ very slightly in performance here, with a tiny advantage demonstrated by the new Winchester core.
The results obtained in 3DMark2001 SE test are quite expected, and the new core demonstrates a slight advantage over the old one. However, at the same time I would like to note that both modifications of Athlon 64 3500+ fall behind Athlon 64 3400+ on ClawHammer core and designed for Socket 754 platform. Of course, 1MB L2 cache of the Athlon 64 3400+ processor ensures much higher 3DMark2001 SE score of this CPU, than the dual-channel memory controller of Athlon 64 3500+.
The results of 3DMark03 are very similar to what we have just seen. However, the overall score for Winchester and NewCastle cores appears about the same.
In fact, we see that Athlon 64 with the newWinchester core is not always faster than a similar solution on the older NewCastle core. For instance, in 3DMark05 test Athlon 64 3500+ manufactured with 90nm process is defeated by the older Athlon 64 3500+ NewCastle.
Well, in the new 3DMark05 test the situation is definitely not in favor of the new core, which proved slower than the good old NewCastle. At the same time I would like to point out that new Socket 939 processors with 3000+ and 3200+ ratings based on the 90nm Winchester core appear very fast in 3DMark05, outperforming considerably the Socket 754 solutions.
Athlon 64 processor family has always been very successful in games. The new Winchester core doesn’t spoil this picture. Only in FarCry its performance is somewhat lower than that of NewCastle. However, the performance difference between different processor cores in this game is quite small. In all other games, Socket 939 processors on Winchester core outperform their 130nm counterparts for both: Socket 939 and Socket 754 platforms. The only exception here is the Socket 939 Athlon 64 3000+, which is very often slower than the Socket 754 Athlon 64 3000+ because of the 200MHz lower actual core clock.
Audio files encoding in Lame is a purely computational task. Therefore, among the processors with similar architecture the laurels are won by the CPUs working at higher clock frequency. Our results prove this point hundred percent.
The same picture can be observed when we encode video in MPEG-2 format. However, in this test the new Winchester core still manages to outperform the older NewCastle.
And when we perform MPEG-4 encoding with a DivX codec the older NewCastle ensures somewhat higher performance. Therefore, all CPUs based on 90nm core, namely Athlon 64 3500+ and Athlon 64 3000+ yield to their predecessors with the same performance rating. The new Socket 939 Athlon 64 3200+, however, outperforms the Socket 754 processor with the same rating.
In case of the XviD codec, the situation turns out completely different. Now the NewCastle core runs slower than Winchester. As a result, the performance of the new Athlon 64 3500+ is higher than that of its 130nm brother.
When we test the new processors by one of the most widely spread archiving utilities, the results appear quite disappointing. Although the new 90nm Winchester core is slightly faster than the older NewCastle one (the old Athlon 64 3500+ yields to the new Athlon 64 3500+), both these Socket 939 CPUs are completely defeated by the Socket 754 Athlon 64 3400+ on ClawHammer core. The similar thing happens to Athlon 64 3200+ for Socket 939, which yields to its Socket 754 counterpart because of the smaller L2 cache. The new Athlon 64 3000+ is defeated by its Socket 754 analogue because of the lower working frequency.
The situation is definitely much better for Winchester during antivirus performance check. However, Athlon 64 3000+ for Socket 939 is again not as fast as we wish it could be.
The new 90nm core manages to raise the performance level of Athlon 64 processors in Adobe Photoshop. Here I would like to draw your attention to the fact that the new Athlon 64 3000+ is not as fast as the old Athlon 64 3000+ for Socket 754 platforms because of its lower clock frequency.
The same picture can be seen when we measure the final rendering performance of the testing participants. In the tasks of this kind the processor clock rate is very important, and neither architectural enhancements, nor dual-channel memory controller can actually make up for it.
The new Winchester core used for the new Athlon 64 processors should be considered a success. AMD hasn’t faced any serious problems when shifting to finer 90nm production technology. The new Athlon 64 processors based on Winchester core boast pretty good frequency potential and extremely low heat dissipation and power consumption. Moreover, by shifting to the new core AMD manages to reduce the production cost of its 64-bit CPUs. This way we are about to see more and more CPUs on Winchester core in the market, and these solutions are most likely to turn into a better choice for the end users from the consumer point of view.
At the same time, you should keep in mind that the launching of the new Winchester core from AMD didn’t improve the performance of the Athlon 64 processor family that much, and didn’t enrich their functionality as well. Summing up the results of our today’s test session, we would like to offer you the following diagram.
The performance advantages ensured by the new 90nm Winchester core compared with the predecessor, 130nm NewCastle core, are minimal and sometimes the CPUs based on the new core are even slower than the predecessors. Since the K8 core design remained the same despite the use of finer 90nm production technology, the major trump of the new Winchester based CPUs is not their performance but their low heat dissipation.
At the same time, the reduction of the die size resulting form the use of finer production process allowed AMD to reduce the CPU production costs and release Athlon 64 3000+ and 3200+ for Socket 939 platform. This way, the mass production of Winchester based processors should automatically increase the popularity of the entire Socket 939 platform.
Unfortunately, the frequency potential of the new processor core doesn’t allow basing the top Athlon 64 models on it. However, AMD is planning to start using strained silicon technology for its CPUs manufacturing, which should make it simpler for the company engineers to increase the clock frequency potential of the Athlon 64 family. So far, the new 90nm core will be used only for the slower CPU models, which can freely overclock up to 2.6GHz. Also, AMD engineers are now working on the new Winchester core stepping with the SSE3 instructions support. So, the upcoming Athlon 64 processors should not only become faster, but also more functional.
In conclusion to our today’s article we would like to offer you a diagram with the overall performance coefficients for the CPUs tested. Let it be your helpful guide in the contemporary CPU market, if you are considering a platform upgrade: