by Ilya Gavrichenkov
04/03/2005 | 04:06 PM
The transition of the enthusiasts’ favorite Athlon 64 processors to the new cores manufactured with 90nm production technology started over half a year ago. The 90nm K8 Winchester core is now widely used in Socket 939 Athlon 64 processors with the 3000+, 3200+ and 3500+ performance ratings, and in Socket 745 Sempron processor family. However, Winchester didn’t manage to completely oust the 130nm cores from the previous generation Athlon 64 processors.
There are multiple reasons for that but the major one lies with the insufficient frequency potential of the Winchester core. Even though it boasts much lower power consumption and heat dissipation than its 130nm predecessor, the maximum actual working frequency of the processor based on Wi8nchester is only 2.2GHz. That is why the top Athlon 64 models as well as Athlon 64 FX-55 with 2.4GHz and 2.6GHz core clock rate are still based on the old Newcastle and ClawHammer cores manufactured with 0.13micron technology.
However, AMD is going to announce the discontinue program for its Athlon 64 processors based on the cores manufactured with outdated production technologies beginning this April. Since the time the first Winchester based processor came out, the company engineers have done great work. They designed a new 90nm core aka Venice (E3 revision), which should send old 130nm cores to the garbage heap of history. Big hopes pinned upon the new core are based on the fact that AMD starts introducing new production standards used specifically for the Venice core.
As a result, this new core should not only allow replacing the Winchester in slower Athlon 64 CPU models and adding new functionality to these processors, but also should come to take the place of the Newcastle and ClawHammer cores in the fastest modifications of this processor family. Moreover, the arrival of Venice sets green light to the release of even faster Athlon 64 processor models. In the near future AMD is expected to announce new Athlon 64 4200+ and Athlon 64 FX-57 processors based on Venice and San Diego cores (San Diego is an analog of Venice but with larger L2 cache).
So, even though this time AMD decided not to make a big event out of the new core release, we could not disregard this important fact, opening new horizons for the further development of the Athlon 64 processor family. Especially, since higher frequency potential is not the only advantage of the new core. Besides that it also provides Athlon 64 processors with SSE3 instructions support, which has been a prerogative of the competing CPUs from Intel since the launch of the Prescott core last year. That is why we decided to write this article devoted to all the details of the new Athlon 64 processor core.
Before we start talking about the new potential of the Athlon 64 processors in terms of clock rates, we have to pay due attention to those CPU features that made this frequency growth possible.
In the end of 2004 AMD and IBM announced another technological breakthrough in the transistor performance field. The technology jointly developed by the engineers from both companies called Dual Stress Liner allowed them to improve the response time of semiconductor transistors by 24%.
The idea behind this technology is fairly simple. In fact, Dual Stress Liner is very similar to strained silicon technology introduced by Intel together with their new 90nm production process. In other words, Dual Stress Liner implies the use of silicon with transformed atomic lattice so that the transistors made of it could boast faster response time and lower heat generation. In one case the silicon atoms are “stretched” and in another case – “squeezed” by moving them onto a nitride capping layer with either stretched or squeezed atomic lattice. Unlike strained silicon used by Intel, Dual Stress Liner from AMD and IBM can be used for both types of transistors: NMOS and PMOS (with n- and p-channels) without the use of exotic silicon-germanium layer, would add cost and possibly hurt the yield of the chips.
This twofold character of the Dual Stress Liner, makes it more efficient than Intel’s strained silicon. Dual Stress Liner allows improving the transistors speed by 24%, while the strained silicon provides the maximum improvement on only 15-20%. And what is also very important, the new technology from AMD and IBM doesn’t reduce the production yields and doesn’t affect the production cost per die.
The new Venice processor core appeared AMD’s first practical experience with Dual Stress Liner technology for desktop PC CPUs. This new technology used together with the successful SOI (Silicon-on-Insulator) technology allowed Venice based processors to reach higher working core clock frequencies. According to AMD engineers’ expectations, Dual Stress Liner and SOI together should ensure about 16% increase in the frequency potential of Athlon 64 processors. In other words, Venice based CPUs should have their nominal frequencies reach 2.8GHz.
Moving from the changes in the production process over to some more tangible things, we should first of all point out that Venice processor core acquired extended SIMD instructions support. Now new Athlon 64 on Venice core support SSE3 instructions, just like the Prescott based Pentium 4 processors. However, it would be good to remind you that SSE3 is not a complete set of instructions, but an addition to the SSE2 instructions set.
Thus, SSE3 instructions set introduced in Venice includes 11 new instructions:
Two more SSE3 instructions implemented in Pentium 4, MONITOR and MWAIT, are absent in the Venice core because they are intended only for work with Hyper-Threading technology.
As a result, new Athlon 64 processors on Venice core boast the today’s biggest SIMD instructions set including 3DNow!, SSE2 and SSE3. However, we can hardly expect SSE3 instructions set to boost up the performance of the new Athlon 64 processors that much. The list of applications using SSE3 instructions is unfortunately still too short and pretty specific.
Every time a new Athlon 64 processor core come out AMD engineers continue tuning up their integrated memory controller. Note that this is done not only for the sake of increasing the processor performance, but mostly in order to extend the compatibility of this memory controller with the different DIMM modules and their configurations. The previous 90nm Athlon 64 processor core aka Winchester had certain performance limitations when four DDR400 SDRAM modules were used. If there were four single-side DDR400 SDRAM DIMMs installed in a system with a Winchester based Athlon 64 CPU, they could only work with the 2T timing, which caused a few percent drop below the usual performance level. If there were four double-side memory modules installed, DDR400 SDRAM would not work at all and its working frequency would automatically drop down to 333MHz.
AMD engineers promised to eliminate this problem in the new Venice core and they did keep their word. Athlon 64 processors based on Venice can work with four single-side DDR400 SDRAM modules without any limitations, and if there are double-side DDR400 SDRAM DIMMs installed, they can work at 400MHz with 2T timing.
Besides wider compatibility of the Venice’s memory controller, it also got noticeably faster. Among the tweaks implemented in Venice we should definitely point out enhanced hardware data prefetch and more write combining buffers (4 instead of 2).
All these improvements should help Venice based Athlon 64 processor to outperform their counterparts based on the older core revisions but working at the same clock frequency. And the performance gain will be more noticeable if there are four memory modules installed in the system.
The new 90nm Venice core will let AMD update the entire Socket 939 Athlon 64 processor family. If the previous Winchester core was only used for CPUs working at up to 2.2GHz core clock, then Venice should definitely help improve this situation. AMD starts shipping Socket 939 Athlon 64 on Venice core with performance rating from 3000+ to 3800+ on April 4. Note that the 3000+, 3200+ and 3500+ CPU models will replace the corresponding models on Winchester core, and the new Athlon 64 3800+ will replace the corresponding solution on Newcastle core.
The total core replacement will also include Athlon 64 4000+, which is now based on ClawHammer core with 1MB L2 cache. On April 15 AMD will begin shipping new Athlon 64 4000+ processors on San Diego core, which will be the same as Venice but with larger L2 cache memory. Summing up everything we have just said, we would like to offer you a table listing all Socket 939 Athlon 64 processors based on the old and new processor cores, which are already available or will become available shortly:
As we see, Venice based CPU prepared one more surprise for us: variable Vcore. AMD resorted to the same trick Intel did when packaging their CPU dies. Venice based processors will not have Vcore indicated on the die package. Different CPUs may have different Vcore: 1.35V or 1.4V. And unfortunately, there will be no way to tell the nominal processor core voltage from its exterior. As for the thermal design power, the new Venice core seems to have the same heat dissipation characteristics as the previous cores at first glance. However, in reality things are not quite like that. The thing is that Venice has somewhat bigger frequency potential for the allowed heat dissipation limited by the thermal design power. The new Athlon 64 on Venice core working at 2.6GHz frequency will feature 89W typical heat dissipation, while the next stage of heat dissipation, 104W, will only be reached when the working frequency rises to 2.80GHz, if at all.
In conclusion I would like to mention that despite all these changes made to the Venice core, it retained the same die size and number of transistors as Winchester core.
We managed to get hold of one Venice based Athlon 64 processor – 3800+. This is the top model from the family of new Athlon 64 CPUs, which will start shipping before April 15.
The new Venice based CPU looks just like its predecessors:
You can tell the newcomer from the marking. The last two letters “BP” indicate that there is a new Venice core (E3 revision) under the processor package lid.
CPU-Z diagnostic utility provides us the following information about this processor.
As we see, the latest version of the CPU-Z available at the time the testing was done, version 1.28, has no idea who Venice is. But it detects SSE3 support and reveals the CPUID of the new processors: 00020FF0h.
Before we pass over to the actual testing of the new processor on Venice core, we decided to check the compatibility of this processor with different Socket 939 mainboards. The thing is that there have been quite a few rumors about the compatibility issues Venice based Athlon 64 processors might have. Luckily, and our test session proved it, these rumors were too exaggerated. We checked a few mainboards on NVIDIA nForce3 chipset: MSI Neo2 Platinum and EPoX 9NDA3J, and a few boards on NVIDIA nForce4 chipset: ASUS A8N-SLI, EPoX 9NPA+ Ultra, DFI NF4 Ultra-D and MSI Neo4 Platinum SLI. There wasn’t a single case when the new Venice based processor wouldn’t work any of these mainboards at all. However, we still revealed a few issues (even though we reflashed the latest BIOS versions to all the boards in advance). On most mainboards Cool’n’Quiet technology of the new Athlon 64 3800+ didn’t work. In fact, the only mainboard where this power saving technology showed some signs of life was MSI Neo4 Platinum SLI. With the help of this board we found out that during low power consumption Athlon 64 3800+ on Venice core drops its clock frequency to 1.0GHz and Vcore to 1.1V, i.e. behaves just like any of its predecessors.
Moreover, we have also discovered some other minor problems. For example, ASUS A8N-SLI mainboard set the processor Vcore incorrectly, making it 0.1V higher, and DFI NF4 Ultra-D refused to work with the 1T memory timing when we installed the new Venice processor. Actually, you can see that all these issues are very specific for each mainboard and they will undoubtedly be eliminated when the new BIOS versions come out.
Since AMD promises that Venice core will allow the company to raise Athlon 64 working frequencies without any modifications of the mainboards processor voltage regulator, we decided to measure the electrical parameters of the newcomer at different frequencies. We have also measured the working temperatures of the Venice based processor. These parameters were compared with the corresponding measurements taken for Newcastle and Winchester based CPUs. When we measured the thermal parameters we used the following testbed:
First of all let’s take a look at the temperature measurements taken for two working modes: idle mode and under maximum CPU workload created by a special S&M utility version 0.3.2, which is the best tool for heating up the CPUs available today. The temperatures were taken from the thermal diodes built into the CPU cores.
Well, this is a pretty typical picture. In idle mode as well as in burn mode the temperature ratio for the different CPUs is the same. 0.13micron Newcastle core heats up much more than its fellows manufactured with finer technology. The new Venice core is a little bit warmer than Winchester, which is actually reflected in its official specification: the maximum CPU case temperature claimed for Venice is higher.
This way, it would be incorrect to claim that Dual Stress Liner technology used for Venice processors reduced their heat dissipation level. The major gain is definitely obtained from the higher frequency potential, but in no way from lower heat dissipation.
However, it is still too early to draw any final conclusions yet. Let’s take a look at the results of our power consumption tests carried out for Newcastle, Winchester and Venice based processors. Just like in the previous case, these measurements were taken in two work modes: in idle mode and in case of maximum CPU workload created by the S&M utility. To estimate the average power consumption we used special clamp multimeter to detect the current going through the 12V circuit powering the CPU. In other words, the data given below doesn’t take into account the performance index of the CPU power converter that is why you may find these numbers a little bit higher than the actual processor power consumption values (about 10%). However, in this case this difference is negligible, because it is relatively small and repeats for all results.
The picture is just the same as in our temperature tests. The CPUs based on 90nm cores boast considerably lower power consumption than their predecessors on ClawHammer core. This tendency takes place for both: idle and burn modes.
As for the power consumptions of Venice and Winchester cores, we can see that the newcomer eats up a little bit more power. The power consumption of the processors with E3 core stepping is about 17% higher than that of processors with D0 core stepping. All this pushes us towards the conclusion that AMD didn’t focus on lowering the power consumption and heat dissipation of its new Venice based processors. In this case Dual Stress Liner technology should primarily ensure higher frequency potential.
However, despite this fact the maximum heat dissipation of the future Venice based Athlon 64 processor working at 2.6GHz core clock will not exceed the indicated threshold of 89W. I would even expect this maximum to remain valid for the 2.8GHz clock frequency, too.
So, from the power consumption point of view, Athlon 64 processors based on the new Venice core have every chance to reach relatively high working frequencies without any additional modifications of the processor voltage regulator on the corresponding Socket 939 mainboards.
In order to evaluate how greatly the new Venice sore affected the performance of Athlon 64 processors compared with the performance of their fellows on ClawHammer and Newcastle cores, we carried out a brief comparative test session for the Athlon 64 3800+ and Athlon 64 3500+ based on different cores. Of course, the SSE3 instructions support as well as the enhancements of the integrated memory controller improved the performance of the new Athlon 64 on E3 core stepping aka Venice. On the other hand, this performance gain will not be that noticeable, because the processor rating system remained the same. What I am trying to say is that the old relation between the clock frequency and performance rating is also true for the new Athlon 64 on Venice core.
There are systems that took part in our test session:
The main goal of this test session is certainly to compare the performance of Athlon 64 processors based on different cores. However, we have also included the results for the top Intel Pentium 4 CPUs based on Prescott and Prescott 2M cores. This will allow us to get a better idea of the SSE3 instructions support efficiency in AMD CPUs.
So, let’s get started.
The SuperPi benchmark, which is very popular among computer enthusiasts, shows clearly that Venice core is faster than Winchester. However, I have to stress that the performance increase we observe here is not that great and hardly exceeds 0.5%. Anyway, nobody promised anything significant here: the memory controller in the Venice core got just a little bit better, and the discussed SuperPi test doesn’t use SSE3 instructions at all.
Here I have to clarify that there is an unofficial patch for SuperPi benchmark, which involves FISTTP instruction from SSE3 set into its calculations. Once you run this patch, the calculation of 8M digits of the pi-value on Athlon 64 3800+ (Venice) will take 375 seconds instead of 385, which is about 2.5% faster than same calculation speed without SSE3 patch.
The results of the FutureMark PCMark04 allow us to draw similar conclusions. The advantage of the new Venice core over the Winchester doesn’t go beyond 1% all the time.
3DMark2001, which has long become more of a CPU test rather than graphics accelerator test, is also showing the same results as the previous benchmarks. Venice based CPUs are again faster than Winchester based ones, and again this advantage is relatively small.
But in the newer 3DMark05 Athlon 64 3500+ with D0 core stepping appears faster than the same CPU on the E3 core stepping. So, it looks like the improvements of the Venice memory controller may sometimes have reserves effect. In fact, the processor performance increased much more on the transition from Newcastle to Winchester core, than on the transition from Winchester to Venice. In other words, the enhancements of the Winchester memory controller were much more global than those of the current Venice memory controller. This is our judgment so far, because we haven’t yet discussed the performance in applications using SSE3 instructions.
CPU performance index from the same FutureMark 3DMark05 reveals more significant dependence of the results on the processor core optimizations. According to the benchmark numbers, Venice is about 4.7% faster than Winchester.
Now let take a look at the performance of new Athlon 64 processors on the E3 core stepping in real games.
The situation is almost the same in all gaming benchmarks. Athlon 64 on the new Venice core is undoubtedly faster than the Winchester based predecessor working at the same clock frequency. However, this advantage is not overwhelming and makes 1.5% on average. Against the background of 4% performance gain Winchester based processors demonstrated in gaming applications over Newcastle based ones, the results we are seeing now from the E3 core stepping are a little bit disappointing. And this happens even in the games supporting SSE3 instructions, such as Doom III. However, we should keep in mind that AMD developed this new core not for the sake of higher performance, but for the sake of bigger frequency potential.
Moreover, we shouldn’t forget that SSE3 support is the major marketing attraction of the new Athlon 64 processors. The best effect from these instructions is available today only in some audio/video codecs, because other applications supporting SSE3 are not that widely spread in the market yet. Therefore, we decided to pay special attention to the performance of Venice based processors in audio/video encoding and digital content creation applications.
Unfortunately, we did not reveal any qualitative changes in the audio/video encoding tasks by different codec. The SSE3 instructions support introduced in Venice based Athlon 64 processors failed to provide the anticipated improvement, despite all our expectations. Although two codec from the set of applications we used do support SSE3 instructions for sure (these are DivX and Mainconcept MPEG Encoder), we do not see any performance growth once we shift from Winchester to Venice core. So, we have to conclude that SSE3 instructions (or at least their implementation in the new Athlon 64 processors) are of no practical use. Although we have to keep in mind that there will be more applications coming out in the future, that will take better advantage of SSE3 instructions.
And here is one more very curious result. WinRAR is an application that is very sensitive to the memory subsystem performance. However, this benchmark provides practically the same result for different Athlon 64 processor cores: Newcastle, Winchester and Venice.
ScienceMark 2.0 benchmark using math1ematical modeling methods does support SSE3 instructions set. However, we see once again that this innovation in the Venice processor doesn’t have any positive effect on the score.
The final rendering speed measured in the 3ds max 7.0 generates even more suspicious result. The best performance here belongs to Athlon 64 processors based on the old 130nm ClawHammer and Newcastle cores. The newer core versions, Winchester and Venice, cannot boast the performance as good as that of their predecessor here.
In Photoshop CS Athlon 64 3500+ processor with E3 core stepping outpaces its predecessor, Athlon 64 3500+ on D0 core stepping, by 0.4%. This is a pretty logical result.
This way, we have to admit that the performance of the new Venice processor has grown up just a little bit compared with that of the predecessor on the Winchester core. The average advantage of the Athlon 64 3500+ with E3 core stepping over its counter part with the previous D0 core stepping is about 0.7%. This is considerably less than the performance growth we observed between the new Winchester core and the Newcastle core. So, we have to state that the memory controller tweaks implemented in the new processor core are really insignificant for the overall CPU performance. As for the SSE3 instructions support, it is hardly of any effect at all, and simply adds a new item into the list of Athlon 64 features.
All in all, the performance of the processors based on the new Venice core wasn’t that very surprising. Now let’s take a closer look at the frequency potential of this core. Maybe we will find out something special there?
Right now AMD is not manufacturing any CPUs on Venice core with the clock frequencies over 2.4GHz. However, this is just for now. AMD believes that the new core manufactured with Dual Stress Liner and SOI technologies will allow increasing the working frequencies of Athlon 64 processors up to 2.8GHz. So, I would expect the frequency potential of the new Venice based CPUs to be pretty high.
According to statistics, Athlon 64 processors based on Newcastle core can be overclocked up to 2.6GHz on average. The same CPUs on Winchester core overclock about the same, although if you are lucky enough you might find a better sample, which will be able to reach 2.65-2.7GHz frequencies. We expect the new Athlon 64 on Venice core to become the record-breakers today.
Well, for our overclocking experiments we took an Athlon 64 3800+ on Venice core marked as ADA3800DAA4BP. The nominal clock frequency of this processor equals 2.4GHz, and the core voltage is 1.4V.
The overclocking tests were done on the following testbed:
Since Venice based processors, just like their predecessors, support Cool’n’Quiet technology, the clock multiplier can be reduced below the nominal value. However, we carried out our first overclocking tests without resorting to this option.
First of all, we decided to check how far our Athlon 64 3800+ on the E3 core stepping could go without raising the Vcore. Here we managed to raise the clock generator frequency from the nominal 200MHz to 226MHz. In other words, the processor working frequency got 2,712MHz. The CPU overclocked up to this point worked stably and went through all the benchmarks including Prime95 and S&M stability tests. So, without much effort we managed to reach the frequency unattainable for most Athlon 64 processors on D0 and CG core revisions (without any additional cooling involved).
But we didn’t want to stop at this point, of course. We also wanted to see what height we could reach once we increase the processor Vcore. That is why the second overclocking attempt was undertaken once we increased the Vcore by 13% over the nominal value, i.e. up to 1.58V. In this case the top frequency limit our processor could work on stably got much higher. We managed to make our clock generator work at 240MHz, thus bringing the processor core clock up to 2,880MHz. At this frequency our CPU run through all the tests flawlessly again, but the frequency couldn’t be raised any further without losing stability.
The screenshot below proves how far we reached:
Note that if you do not run any stability tests for an overclocked system, you can obtain even more impressive results. For example, our CPU could still boot the OS when we overclocked it almost to 3GHz. We even managed to take a screenshot from the CPU-Z utility. Take a look:
However, in this case almost every CPU loading test results into immediate system freeze or crash.
We measured the performance of our Athlon 64 3800+ processor on the Venice core overclocked to 2.88GHz after we reduced its multiplier to 10x and raised the clock generator frequency to 288MHz.
This parameters combination allowed us to increase the memory subsystem performance since the memory was now clocked synchronously with the clock generator frequency. The memory (we used Corsair CMX512-4400C25) worked at 576MHz with the timings set to 3-4-4-8. We compared the performance of the Venice based processor overclocked to 2.88GHz with the performance of the today’s fastest AMD processor - Athlon 64 FX-55. The results of this comparative testing are given in the table below:
Venice 3800+ | Venice @ 2.88GHz | Athlon 64 FX-55 | |
Super PI 8M, sec | 385 | 333 | 368 |
PCMark04 | 4952 | 5605 | 5278 |
3DMark2001 SE, | 26129 | 28928 | 28733 |
3DMark05, Default | 5694 | 5790 | 5756 |
3DMark05, | 5293 | 5890 | 5111 |
Quake3 (four), 1024x768 | 481.5 | 553.4 | 510.3 |
Doom 3, Medium Quality, | 106.7 | 111.7 | 110.6 |
Half-Life 2 (d3_c17_02), | 81.55 | 90.4 | 88.72 |
MPEG-4 Encoding, | 43.14 | 47.88 | 45.19 |
3ds max 7.0, Final Rendering, | 237 | 199 | 219 |
The results indicate that the Athlon 64 CPU on E3 core overclocked to 2.88GHz turns out even faster than AMD Athlon 64 FX-55.
So, we can state that the frequency potential of the new Venice core from AMD really proved highly impressive. This core can easily work at frequencies beyond 2.8GHz, and this has been impossible for any other core so far. Therefore, we tend to believe with all certainty that the new E3 core will allow AMD to reach new speeds in the nearest future.
In conclusion I would only like to add that we hope slower models on the new Venice processor core will also boast the same impressive frequency potential. In this case they will undoubtedly become overclocker’s best choice.
Well, we have just introduced to you one of the first Athlon 64 processors on the new Venice core (E3 core revision). AMD engineers will most likely be using this core for their processors for a long time. That is why a Venice based processor will very soon become a regular component of an AMD Athlon 64 based system. Today Venice is still regarded as a new product boasting a few important advantages compared with the previous cores such as Winchester and Newcastle.
Among these advantages I would first of all like to mention SSE3 instructions support, which makes Athlon 64 look not any weaker than Pentium 4 Prescott from the SIMD instructions support point of view. Unfortunately, we failed to see any dramatic performance improvement from SSE3 support this time, however, we still hope for the best :) SSE3 support will most probably play its part to the advantage of Athlon 64 later, besides, it is not that bad to have something in reserve anyway.
The new Athlon 64 on Venice core also boast enhanced memory controller. Besides a slight performance improvement, Athlon 64 became more tolerant to various DIMM module configurations. Venice based CPUs have finally learned to work with four double-side memory modules supporting 400MHz.
But the most important advantage of the Athlon 64 processors based on Venice core is the significantly higher frequency potential. This opened new horizons for further AMD processor family development, at the same time giving overclockers got new opportunities for building high-performance systems on the far not the most expensive processors.
As for the drawbacks of the new Venice core, there are hardly any. In other words, if you are planning to buy a Socket 939 CPU, then you definitely have to go for the new core. That’s for sure.
And in conclusion I would like to offer you a nice graph showing the performance advantages of the Athlon 64 processor on Venice core over their fellows on the previous generation cores aka Newcastle and Winchester.
