AMD's Energy Efficient and EE SFF CPUs

30 Comments

View All Comments

• safan80 - Wednesday, August 2, 2006 - link

  I can't find the 35watt cpus anywhere. I brought an 89 watt x2 3800 for my file server, but I'd like to buy a 35 watt part.

• VooDooAddict - Monday, July 31, 2006 - link

  I'd definetly like some more information on cost and availability of the EEs. Anandtech has resorces and contacts available that other sights won't. Many sites will review the EEs on the basis of heat, power, and silence ... does us no good if we don't know how to get them or or if we can't plan the budget for them.
  
  Large PC makers only?
  
  OEM only? Retail?
  
  Projected cost?
  
  

• mooshi - Sunday, July 30, 2006 - link

  "Performance per watt" sounds like a good concept, but its definition is vague.
  
  Energy measures should reflect the resources consumed to accomplish a given task. There is, of course, the time involved to accomplish the task (and related the number of clocks), but also the instantaneous max/min powers, integrated power over time, and don't forget the spatial extent--how much physical area is in use, how many gates, how much L2 cache.
  
  The reason CPUs of today are faster than before is because we have traded space for time. They are more energy efficient because of length scales getting smaller, so distance traveled and energy diffused is lessened, which helps to reduce voltage, which in turn reduces energy at a squared rate.
  
  All energy metrics should use the completion of a standard real world task as its invariant. Integrate instant power over the time it takes to accomplish this task to come up with an "energy to complete" metric. For instance, measure the watt-hours (or watt-seconds) consumed to boot into Windows, calculate some millions of digits of pi, compress a large wav, encrypt some file, whatever. But choose a few memorable accomplishments, and associate an energy consumption with it.
  
  In this way, time is an integral part of the performance metric. It may be that some tasks are performed with the same overall energy, but with substantial time differences. For those considering large installations of systems for computational tasks, the time it takes to accomplish a task may be less important than the total energy consumed.
  
  Introducing a concise way to tradeoff your resources would be a great boon for the computing industry: Time, Instant Power, Total Work Done. Ideally, chips could be pushed up to extremely high clock rates (greater than "stable overclocks" even) for very brief intervals of time to accommodate strict latency constraints, and then clocked back down again, perhaps with a "penalty" period of "no-overclocking" where the excess heat is guaranteed to be transported away.
  
  Keep in mind that the most efficient CPU could be defined as one that is always at 100% utilization, and meeting all of its scheduled task completion deadlines.
  
  Consider a representative from the VIA EDEN series to compare with, and maybe an older Pentium PRO 200. With the ever descending core voltages, and noting energy goes as a function of voltage squared, the performance of today's CPUs may be put into perspective with those of old, with a more equal ground energy comparision (as opposed to simple timing statements or FLOPS).
  
  A metric that compared space and time would be cool, too. How much space-time (measured in transistor-clocks) does it take to boot windows? (Pentium Pro has 5.5 Million core and 15.5 million 512k cache transistors, so 21 million x 200 seconds @ 200MHz = 8.4 E17 transistor-clock operations to boot, perhaps.) Compared to the Pentium D at 376Million transistors, and gigahertz clock rates, how does the old Pro fare? How does the Core 2 fare?
  
  Looking at useful efficiency metrics in computing is still in its infancy. But keeping in mind that energy metrics may be wide ranging, if they may all be related to a common reference (completing a real world task) then more intelligent, more pertinent information can be gleaned from the numbers to guide purchasing decisions.

• mooshi - Sunday, July 30, 2006 - link

  "Performance per watt" sounds like a good concept, but be careful in how it gets defined. Consider choosing

• pablo906 - Wednesday, July 26, 2006 - link

  I've always been curious the difference that extreme cooling can make on low watt cpu's. I think some chilled water or sub zero would be interesting in that for sure.

• Jjoshua2 - Tuesday, July 25, 2006 - link

  Are there socket 939 EE? I have an Opteron 144 socket 939 currently, and am thinking about upgrading to X2 3800+ EE or something with the cheaper prices.

• Sgraffite - Monday, July 24, 2006 - link

  I'd like to see how much you can undervolt them but still retain their rated clock speed, as well as power consumption when undervolted, etc. Also how do these compare to TCaseMax 35w TDP Opteron processors as far as power consumption and performance per watt?

• eRacer - Saturday, July 22, 2006 - link

  Performance per watt will be very dependent on what chipset is used. ATI and NVIDIA chipsets should be compared. Power use at idle, under CPU load and under CPU and GPU load would be interesting. The nForce 590 SLI is overkill for SFF and media center systems and is going to run quite hot. I'd like to see if the NVIDIA 570 and/or NVIDIA 550 chipsets are better alternatives in terms of power consumption.

• Jjoshua2 - Sunday, July 23, 2006 - link

  What I want to see is given x $ is it better to get the EE or the non-EE of a higher model number.

• JumpingJack - Saturday, July 22, 2006 - link

  Anand,
  Viewing the specs for the EE vs non-EE they appear to by undervolting the EEs to achieve the lower power, and likely down-binning but it is not clear. Here is the comparision I would like to see:
  
  EE vs non-EE at identical clocks and Vcore across a matrix, say take a 3800+ do a matrix:
  
  2.0 GHz @ 1.20, 1.25, 1.30, 1.35
  2.2 GHz @ 1.20, 1.25, 1.30, 1.35
  etc.
  Put it under full load and measure power and temperature (use the same thermal solution at the same fanspeed RMP), disable cool and quiet.
  
  My suspicion is that they are doing more than simply undervolting and that there is a processing trick that is getting an extra 10%. For example, at 1.20 the dynamic power would only scale down about 15%, or roughtly 89 TDP to 75 TDP, another 10 watts needs to come from somewhere -- is it binned from frquency, is it a leakage trap/trick, could these be just super low yielding straing 90 SOI wafers?
  
  Doing the skew above on both a non-EE and EE would allow one to determine where within the matrix of bin's AMD may be choosing the EE processors and why we have such low availability.
  
  Thanks.