- Dec 16, 2004
- 997
- 0
- 18,980
Earlier this month Tom's Hardware Guide released a review of Intel's upcoming Cedar Mill and Presler processors. This generated a rash of comments that doubted the ability of these processors to perform properly, especially at high clock rates, due to thermal issues. However, the latest review by AnandTech proves otherwise.
http://www.anandtech.com/cpuchipsets/showdoc.aspx?i=2578&p=3
Anandtech found similar power savings as Tom's Hardware in the new Cedar Mill. They found that a 3.6 GHZ 660 Prescott uses 21% more power than a 3.6GHz 661 Cedar Mill. What is more interesting is that the 3.4GHz 950 Presler uses as much power as a 2.8GHz 820 Smithfield. This would mean that the 65nm process yielded a 600 MHz or 21.4% increase in processor speed and a 1 MB increase in L2 cache (which is one of the more power consuming components) without additional power usage. This is a definite increase in performance per watt.
Previously, I mentioned that Intel should release Cedar Mills clocked at 4.2 GHz. The reply from people was that the 3.8GHz 671 would never be released due power inefficiency and that a 4.2GHz Cedar Mill was "a pipedream". Once again I beg to differ. The major concern was that for every increase in clock speed the power consumption and loss increases exponentially, and in order to achieve stability Vcore increases would be needed further worsening the situation. It was also mentioned that the "Vcore has remained basically unchanged in spite of the shrink". In fact the Vcore has decreased from 1.4V in Prescott to 1.3V in Cedar Mill. In the end, a lower starting Vcore and the 21% reduction in power consumption yield ideal conditions for higher clock speeds. Anandtech proves this by overclocking a early sample of the 661 from 3.6GHz to 4.25GHz. While a 625MHz or 18.1% increase in speed may or may not sound impressive depending on your pessimism, the important thing to note is that this was down without increasing Vcore and with Intel’s Reference Cooler. Only standard overclocking, no special parlour tricks. Anandtech saids that “any serious overclocker to reach speeds greater than 4.0GHz effortlessly”. This clearly indicates that a 3.8GHz Cedar Mill is a legitimate release processor, and as such appears on Intel’s Processor Roadmap. It also indicates that a 4.2 GHz Cedar Mill is not “a pipedream”, but rather is something that Intel can reach “effortlessly”.
In fact, Cedar Mills were found to run stably at 4.5GHz on the stock cooler although the Vcore had to be increased to 1.4125V. It is also interesting to note that these overclocks were done using a 3.6GHz 661. A higher quality, higher binned part such as the 3.8GHz 671 that Intel is producing will likely still have slightly more room for speed increases.
Of course, now the concern is whether the power consumption will make such a chip unreasonable. Once again we need to look at the figures. A 950 Presler has the same power consumption as a 820 Smithfield. This is analogous to a 651 Cedar Mill consuming the same amount of power as a 521. We ask ourselves, what is the power consumption of a 3.8GHz 671? Assuming the 65nm process gives 600MHz and 1 MB of L2 cache worth of additional power consumption room, the 3.8GHz 671 would consume more power than the 3.2GHz 541, since power consumption increases non-linearly as clock speeds increase, but less than the 3.4GHz 551. Similarly a 4.2GHz 691 would consume more power than a 3.8GHz 571. Now it would appear the 4.2GHz Cedar Mill has broke the bank, so to speak, since it consumes more power than a 3.8GHz Prescott. However, several things need to be taken into account. First, Intel motherboards are designed for higher power consumption than the 3.8GHz 572 ever since the advent of the dual-core Pentium Extreme Edition. Secondly, my estimates are decidedly conservative. I mentioned above that Cedar Mill runs stably at 4.5GHz. As it turns out, Anandtech proved that the power consumption of a Cedar Mill at 4.5 GHz is less than the power consumption of a 3.6GHz 660. Therefore, a 4.2GHz 691 should have no problems remaining far below Intel’s current maximum limits.
The situation is similar in the dual core segment. If a 3.4GHz 950 can maintain the power levels of a 2.8GHz 820, Intel’s fastest dual core, the 3.46GHz 955 should remain in the power levels of the 3.0GHz 830 despite the higher 1066MHz FSB. This means that Intel still has clocking room for Presler while still remaining within the power levels of the 3.2GHz 840EE. Anandtech themselves clocked Presler to run stably at 4.25GHz.
The remaining concern is Intel’s new marketing strategy of performance per watt. In this case, significantly higher clocks can be reached while maintaining the same power usage(watt). A 4.5GHz Cedar Mill provides a 18.4% clock increase compared to the 3.8GHz 670 yet remains within the power consumption of the 3.6GHz 660. Therefore, increasing clock speeds does not necessarily mean decreasing performance per watt. From a marketing standpoint, winning is still the most important. Conroe is scheduled for H2 2006, but won’t likely be released en masse until Q4 2006 when all the kinks are worked out. That leaves Intel 1 year with no speed increases. During this time AMD is planning on releasing the X2 5000+ and the FX-60. In order for Intel to remain competitive they must release faster processors regardless of any slight decreases in performance per watt. Yonah’s weak FPU performance means it’s in no shape to fight the AMD64 despite its Digital Media Boost.
As such, Intel should release a 4.0GHz 681, and the 4.2GHz 691 that I’ve been arguing for all along. In order to properly combat the FX-60, a new single core Extreme Edition needs to be introduced. A 4.26GHz 966 Cedar Mill Extreme Edition could be released. Its power consumption would still be within Intel’s current maximum, and its high clock speed will mean it can finally benefit from a 1066MHz FSB. A 3.6GHz 960 Presler is already rumoured to be in the works if required and should also be released. The ultimate Intel processor, would then be a 3.73GHz 965 Presler Extreme Edition. This part may be just over the Intel’s current power comsumption limits but would certainly be worth it. Where $999 currently buys you 1 3.73GHz Extreme Edition, the same could get you 2 in dual core configuration.
Intel could easily build these processors with little cost. As Anandtech mentioned, clocks over 4.0GHz are effortless, and the current reference cooler is sufficient and does not require redesign. The 4.0GHz 681 and 4.2GHz 691 are easily reached with processors binned as 3.6GHz 661s and higher. As the 65nm process continues to improve before Cedar Mill introduction in Q1 2006, the yields will improve. A 3.6GHz 960 Presler is easily produced by slapping 2 661s together and is Intel’s current backup plan for a trump card to compete with faster AMD processors. While Cedar Mills with 1066MHz FSBs are not currently on the Roadmap, they are being produced to make the 3.46GHz 955 Presler. Producing a 3.73GHz 965 or a 4.26GHz 696 is simply a matter of improving yields and binning higher. No major retooling is necessary.
Clearly, the Netburst architecture still had some life left with the 65nm process. I know many people would disagree with me but the ability to run a 4.5GHz Cedar Mill for the same power consumption as a 3.6GHz Prescott proves that power issues can be solved or at least reduced to more acceptable measures. Sadly Intel is obsessed with their next-generation Conro, Meron, and Woodcrest architecture. What they should have done is invested more into the development of Cedar Mill and Presler. I’m not talking anything revolutionary. Just making the necessary changes to allow SpeedStep to run at lower clock speeds would have solved most of the power issues since processors are rarely at full load anyways. If Intel was afraid of a fully Enhanced SpeedStep capable Desktop processor cutting into Pentium M sales, they could introduce a SpeedStep feature into the Cedar Mill that is closer to AMD’s Cool’n’Quiet. Instead of enabling all multipliers, just enable a few at set levels such as 6, 10, and 14 to yield steps of 1.2GHz, 2GHz, and 2.8GHz. Another feature that could have improved Cedar Mill’s performance without a major redesign is fixing the L2 cache latency problem. The high L2 latency of the Prescott 2M was mainly due to Intel simply sticking an additional 1MB of L2 onto the processor. They could achieve better performance by tweaking the routine that the processor uses to address and prefetch the cache to better take into account that 2MB of L2 is available instead of 1MB. These two improvements, along with higher clock speeds, and VT implementation (the 6x3s were cancelled) would have kept the Pentium 4 alive, and profitable for the rest of the year until Conroe. It would have also given Intel some fallback room in the event that Conroe is delayed. As it is the best Intel can do right now is to continue to improve there manufacturing process to further reduce processor power consumption, and increase yield allowing for greater clock speeds.
Just my opinion on Intel's strategy.
http://www.anandtech.com/cpuchipsets/showdoc.aspx?i=2578&p=3
Anandtech found similar power savings as Tom's Hardware in the new Cedar Mill. They found that a 3.6 GHZ 660 Prescott uses 21% more power than a 3.6GHz 661 Cedar Mill. What is more interesting is that the 3.4GHz 950 Presler uses as much power as a 2.8GHz 820 Smithfield. This would mean that the 65nm process yielded a 600 MHz or 21.4% increase in processor speed and a 1 MB increase in L2 cache (which is one of the more power consuming components) without additional power usage. This is a definite increase in performance per watt.
Previously, I mentioned that Intel should release Cedar Mills clocked at 4.2 GHz. The reply from people was that the 3.8GHz 671 would never be released due power inefficiency and that a 4.2GHz Cedar Mill was "a pipedream". Once again I beg to differ. The major concern was that for every increase in clock speed the power consumption and loss increases exponentially, and in order to achieve stability Vcore increases would be needed further worsening the situation. It was also mentioned that the "Vcore has remained basically unchanged in spite of the shrink". In fact the Vcore has decreased from 1.4V in Prescott to 1.3V in Cedar Mill. In the end, a lower starting Vcore and the 21% reduction in power consumption yield ideal conditions for higher clock speeds. Anandtech proves this by overclocking a early sample of the 661 from 3.6GHz to 4.25GHz. While a 625MHz or 18.1% increase in speed may or may not sound impressive depending on your pessimism, the important thing to note is that this was down without increasing Vcore and with Intel’s Reference Cooler. Only standard overclocking, no special parlour tricks. Anandtech saids that “any serious overclocker to reach speeds greater than 4.0GHz effortlessly”. This clearly indicates that a 3.8GHz Cedar Mill is a legitimate release processor, and as such appears on Intel’s Processor Roadmap. It also indicates that a 4.2 GHz Cedar Mill is not “a pipedream”, but rather is something that Intel can reach “effortlessly”.
In fact, Cedar Mills were found to run stably at 4.5GHz on the stock cooler although the Vcore had to be increased to 1.4125V. It is also interesting to note that these overclocks were done using a 3.6GHz 661. A higher quality, higher binned part such as the 3.8GHz 671 that Intel is producing will likely still have slightly more room for speed increases.
Of course, now the concern is whether the power consumption will make such a chip unreasonable. Once again we need to look at the figures. A 950 Presler has the same power consumption as a 820 Smithfield. This is analogous to a 651 Cedar Mill consuming the same amount of power as a 521. We ask ourselves, what is the power consumption of a 3.8GHz 671? Assuming the 65nm process gives 600MHz and 1 MB of L2 cache worth of additional power consumption room, the 3.8GHz 671 would consume more power than the 3.2GHz 541, since power consumption increases non-linearly as clock speeds increase, but less than the 3.4GHz 551. Similarly a 4.2GHz 691 would consume more power than a 3.8GHz 571. Now it would appear the 4.2GHz Cedar Mill has broke the bank, so to speak, since it consumes more power than a 3.8GHz Prescott. However, several things need to be taken into account. First, Intel motherboards are designed for higher power consumption than the 3.8GHz 572 ever since the advent of the dual-core Pentium Extreme Edition. Secondly, my estimates are decidedly conservative. I mentioned above that Cedar Mill runs stably at 4.5GHz. As it turns out, Anandtech proved that the power consumption of a Cedar Mill at 4.5 GHz is less than the power consumption of a 3.6GHz 660. Therefore, a 4.2GHz 691 should have no problems remaining far below Intel’s current maximum limits.
The situation is similar in the dual core segment. If a 3.4GHz 950 can maintain the power levels of a 2.8GHz 820, Intel’s fastest dual core, the 3.46GHz 955 should remain in the power levels of the 3.0GHz 830 despite the higher 1066MHz FSB. This means that Intel still has clocking room for Presler while still remaining within the power levels of the 3.2GHz 840EE. Anandtech themselves clocked Presler to run stably at 4.25GHz.
The remaining concern is Intel’s new marketing strategy of performance per watt. In this case, significantly higher clocks can be reached while maintaining the same power usage(watt). A 4.5GHz Cedar Mill provides a 18.4% clock increase compared to the 3.8GHz 670 yet remains within the power consumption of the 3.6GHz 660. Therefore, increasing clock speeds does not necessarily mean decreasing performance per watt. From a marketing standpoint, winning is still the most important. Conroe is scheduled for H2 2006, but won’t likely be released en masse until Q4 2006 when all the kinks are worked out. That leaves Intel 1 year with no speed increases. During this time AMD is planning on releasing the X2 5000+ and the FX-60. In order for Intel to remain competitive they must release faster processors regardless of any slight decreases in performance per watt. Yonah’s weak FPU performance means it’s in no shape to fight the AMD64 despite its Digital Media Boost.
As such, Intel should release a 4.0GHz 681, and the 4.2GHz 691 that I’ve been arguing for all along. In order to properly combat the FX-60, a new single core Extreme Edition needs to be introduced. A 4.26GHz 966 Cedar Mill Extreme Edition could be released. Its power consumption would still be within Intel’s current maximum, and its high clock speed will mean it can finally benefit from a 1066MHz FSB. A 3.6GHz 960 Presler is already rumoured to be in the works if required and should also be released. The ultimate Intel processor, would then be a 3.73GHz 965 Presler Extreme Edition. This part may be just over the Intel’s current power comsumption limits but would certainly be worth it. Where $999 currently buys you 1 3.73GHz Extreme Edition, the same could get you 2 in dual core configuration.
Intel could easily build these processors with little cost. As Anandtech mentioned, clocks over 4.0GHz are effortless, and the current reference cooler is sufficient and does not require redesign. The 4.0GHz 681 and 4.2GHz 691 are easily reached with processors binned as 3.6GHz 661s and higher. As the 65nm process continues to improve before Cedar Mill introduction in Q1 2006, the yields will improve. A 3.6GHz 960 Presler is easily produced by slapping 2 661s together and is Intel’s current backup plan for a trump card to compete with faster AMD processors. While Cedar Mills with 1066MHz FSBs are not currently on the Roadmap, they are being produced to make the 3.46GHz 955 Presler. Producing a 3.73GHz 965 or a 4.26GHz 696 is simply a matter of improving yields and binning higher. No major retooling is necessary.
Clearly, the Netburst architecture still had some life left with the 65nm process. I know many people would disagree with me but the ability to run a 4.5GHz Cedar Mill for the same power consumption as a 3.6GHz Prescott proves that power issues can be solved or at least reduced to more acceptable measures. Sadly Intel is obsessed with their next-generation Conro, Meron, and Woodcrest architecture. What they should have done is invested more into the development of Cedar Mill and Presler. I’m not talking anything revolutionary. Just making the necessary changes to allow SpeedStep to run at lower clock speeds would have solved most of the power issues since processors are rarely at full load anyways. If Intel was afraid of a fully Enhanced SpeedStep capable Desktop processor cutting into Pentium M sales, they could introduce a SpeedStep feature into the Cedar Mill that is closer to AMD’s Cool’n’Quiet. Instead of enabling all multipliers, just enable a few at set levels such as 6, 10, and 14 to yield steps of 1.2GHz, 2GHz, and 2.8GHz. Another feature that could have improved Cedar Mill’s performance without a major redesign is fixing the L2 cache latency problem. The high L2 latency of the Prescott 2M was mainly due to Intel simply sticking an additional 1MB of L2 onto the processor. They could achieve better performance by tweaking the routine that the processor uses to address and prefetch the cache to better take into account that 2MB of L2 is available instead of 1MB. These two improvements, along with higher clock speeds, and VT implementation (the 6x3s were cancelled) would have kept the Pentium 4 alive, and profitable for the rest of the year until Conroe. It would have also given Intel some fallback room in the event that Conroe is delayed. As it is the best Intel can do right now is to continue to improve there manufacturing process to further reduce processor power consumption, and increase yield allowing for greater clock speeds.
Just my opinion on Intel's strategy.
یί∫υєг ρђœŋίχ 
Facebook
Twitter
Instagram