May 23

Written by: George Ou
5/23/2008 12:23 AM

AMD just released a "new" low-power 1.8 GHz quad-core model X4 9100e and dubbed it "the world's first 65W desktop processor".  But this claim seem to be dubious when you consider what AMD had to do to get to that power level and what they're up against.

• AMD uses very low clock speeds to achieve these power numbers.  The 9100e is clocked exceptionally low for a modern day processor which gives it horrible single- or dual-thread performance compared to a less expensive dual-core 2.5 GHz or 2.66 GHz Intel processor.  For the desktop market, it's hard to see where a low-clocked quad-core makes sense.
• The other problem is that this is a very qualified and questionable victory in actual power performance.  Anytime you drop the clock speed in a given processor, the power consumption will drop.  Intel for example offers their 50W 2.5 GHz 45nm Penryn-class chip called "Harpertown" in the lucrative server space but chooses not to offer this chip in the desktop space.
• Even Intel's 2.66 GHz 45nm quad-core "Yorkfield" model Q9450 would come very close to the AMD 1.8 GHz 9100e quad-core in power consumption.  This is fairly certain since TechReport, LostCircuits, and HotHardware all showed how close the QX9650 3.0 GHz 45nm came to 65W TDP dual-core CPU performance which is nothing short of amazing.  The only anomaly is this recent set of power benchmarks released by TomsHardware which strangely claims that a 45nm Q9450 2.66 consumes more power than an Intel Q6600 2.4 GHz 65nm processor which goes against every 45nm power benchmark I've ever seen.  As soon as the AMD 9100e quad-core becomes available, I'll try test it against an Intel 45nm quad-core to verify.
• But the biggest problem with this 9100e is that it is based on the buggy B2 stepping which has that infamous TLB (Translation Lookaside Buffer) bug that cripples performance when it's patched.  It would seem as if AMD is taking their leftover stockpiles of B2 stepping chips and repackaging them as energy efficient CPUs.  So we're talking about a low-clocked low-performance chip to begin with that has to take an additional TLB patch performance hit.

So why did AMD go with a buggy chip?  It would seem that getting rid of excess buggy B2 stepping CPUs isn't the only motivation.  I got this set of stunning data below from someone I know who we'll call "Jack" for the time being.   Jack has a PhD in Physics and he is also a huge CPU technology enthusiast.  Jack sent me the following power reading chart comparing two AMD X4 quad-core CPUs at identical voltage and clock speeds.  The AMD 9850BE is a B3 stepping CPU and the AMD 9600BE is a B2 stepping CPU.

The reason the 9850BE voltage and clock was dropped down to the level of the 9600BE was to provide an apples-to-apples comparison between the B2 and B3 process.  As you can see, there is still a 16 to 20 watt delta between idle and peak power consumption between these two chips.  What this seems to suggest is that the B3 stepping consumes more power than the B2 process as a tradeoff for higher clock speeds.  In fact, this may explain why AMD silently increased their TDP ratings in their quad-core Opteron CPUs which use the same die as the X4 desktop quad-core processors.

This explains why AMD would use the B2 stepping despite its TLB bug to produce their first 65W TDP desktop quad-core desktop processor where the market isn't as discriminating as the server market against processor bugs.  But why does AMD's B3 stepping consume more power than the B2 stepping?  Jack's theory is that the gates are thicker in B2 which avoids leakage.  His theory is based on the following observation:

Jack: In Nov. Semi International reverse engineered a Barcelona and quoted a gate thickness it was 15% HIGHER than what AMD disclosed at the IEDM in 2005."

Where Jack has the time to dig this kind of stuff up is beyond me, but it sounds like a reasonable theory that is backed up with hard numbers.

Here are some of Jack's disclosures for his test methodology

• The data was collected using the OCCT 2.0 tool, a tool produced by a Russian enthusiast site. The utility stresses the CPU and/or RAM to a high level, while monitoring core temperatures and voltages. The utility can be found here.
• The utility is configurable in terms of time it will stress, this data was collected by setting the utility to stress the CPU only, for a 20 minute duty cycle. The utility first idles for 1 minute, stresses the CPU for 15 minutes, then idles again for 4 minutes in this configuration.
• I used OCCT because it collects temperatures as well and is uses a steady load. Prime 95 mixes up the load and in some segments is lower power and in some cases higher power.
• Several experiments has shown this utility to stress the CPU to within a few Watts higher of a typical Prime95 run, the power was collected using a data logging power meter which was then plotted in Excel.
• The power meter used was a Watts Up Pro ES, and the calibration was verified with a 100, 200 and 300 Watt load and found to be within 1 W accurate.  The test system consists of an Asus M3A32-MVP Deluxe, an NVIDIA 8800GTX card, 2 gigs of Corsair 1066 DDR2 C5 ram, 3 case fans inside a Thermal Take Armor case, a WD 150 G Raptor drive, a Seagate 300 Meg drive, and a Lite-on DVD RW.
• The exact configuration can be found here.
• The ThermalTake 750 W Tough Power PSU has an efficiency rating between 81-83% in the range of 200 to 300 Watts according to the official 80 Plus numbers.

Conclusion on the AMD X4 9100e

• Performance at 1.8 GHz is low to begin with, worse when you have to factor in the TLB BIOS patch.
• The power-efficiency claim is not so compelling and it's a small and questionable victory on energy efficiency.
• At roughly $200, it is probably not worth the premium for such a low performance buggy processor.
• Maybe future bug-free B3 variants that have improved process with a more compelling price point will be worthwhile, but I can't recommend this CPU today.  You would be better off with any cheaper dual-core from Intel or AMD.

Tags: