If you?re an Intel enthusiast, the last eleven and a half months haven?t been especially bright, especially when compared to the heady days of 2002 and 2003, when Northwood hadn?t met a clockspeed it didn?t like, and Canterwood could do no wrong.  The retail market?s pre-launch anticipation of what was widely seen as Pentium 4?s pre-emptive strike against the Athlon 64 was badly deflated by the higher temperatures, decreased efficiency, and mediocre scaling Prescott actually delivered. 

The launch of Alderwood and Grantsdale in June hasn?t done much to change the situation.  While these chipsets are fine, stable products in and of themselves, they?re Prescott-only unless you?re one of the rare individuals who can afford an Extreme Edition.  For users replacing older systems the chipsets themselves could be attractive options, but they don?t offer much in the way of compelling features to the enthusiast market, and they?re saddled with Prescott?s thermal baggage.

Prescott may be as attractive as the idea of belly flopping into a concrete-filled swimming pool, but another star has been steadily rising over Santa Clara.  Intel first debuted the Pentium M (codenamed Banias) when they launched the Centrino mobile platform back in 2003, but enthusiasts didn?t really begin giving it a serious eye until Banias transitioned to 90nm, picked up an extra meg of L2 cache, and became Dothan.  Its easy to see why consumers turned off by Prescott?s thermal signature and power requirements might be attracted to the ultra-cool, low-wattage Dothan?but does Intel?s mobile processor have what it takes to compete on the desktop market?

The Pentium M Architecture

Many of the Pentium M?s design specifications aren?t publicly available; Intel has been relatively close-mouthed about the optimizations and alterations it introduced to first Banias and then Dothan. If you're looking for a good in-depth article on Centrino architecture, check Ars Technica here.   General information on the processor is available, however, and is summarized as follows:

 
Original graphic property of Ars Technica LLC.  Used here with permission.

• Based on the P6 Core:  The Pentium M?s architecture is based on the same P6 core that powered the Pentium Pro, Pentium II, and Pentium III families.  Intel?s success at adapting and extending this core (first launched in 1995) says a great deal about both the initial architecture and the skill of the later engineering teams tasked with updating it.
  
   
• Extended Pipeline:  The exact length isn?t known, but Pentium M?s pipeline is longer than the 10 stage P3 pipe, while shorter than the 20 stage P4.
  
   
• Optimized for Low Power:  Prior to the Pentium M, both Intel and AMD offered mobile processors that were only slightly different than their desktop counterparts.  These chips might run at slower clockspeeds or have lower operating voltages, but their fundamental architectures were identical on laptop or desktop.   Pentium M, in contrast, was designed from the beginning as a      low power architecture.
  
   
• Significantly Improved Branch Prediction:  One of the key differences between the Pentium M and the Pentium 3 lies in their branch prediction capabilities.  The P-M uses the same advanced branch predictor as the Northwood P4, but introduces two new features, namely the Loop Detector and the Indirect Branch Predictor. 

The Loop Detector evaluates a loop, detects the exit point, and notes it, thus preventing the ?exit? condition from catching a CPU by surprise.  This, in turn, prevents the CPU from stalling out and having to refill its pipeline.

The Indirect Branch Predictor is used to store information about the specific addresses an indirect branch is likely to target under any given set of circumstances.  Because indirect branches have a range of addresses they might use depending on what address value is loaded from a register, they are notoriously difficult to predict.  By storing and accessing the ?favorite? addresses of a branch, the Indirect Branch Predictor can significantly improve the likelihood of correct prediction?and hence reduce the CPU?s pipeline stalls. 

• Micro-Op Fusion:  Modern x86 processors don?t actually operate on x86 instructions at all.  The Pentium 3, AthlonXP, Athlon 64, Pentium 4, and Pentium M all translate x86 instructions into smaller instruction sets called micro-ops (uops) by Intel and Macro-ops by AMD.  As you might expect, different x86 instructions translate into different numbers of uops.  Typically, one x86 instruction translates to one uop, a minority translate into two or three uops, and a *very* small group translate into an extended sequence of uops.

x86 decoding was one potential bottleneck in the original P3?s design; although it has three x86 decoders, only one of them is capable of handling x86 instructions that translate into more than one uop.  Since only a minority of instructions translate into more than one uop, this isn?t usually a problem, but a burst of multi-uop activity leaves one decoder backed up with work while two others sit idle.  Micro-op fusion corrects this problem by fusing the store-address and load-op instructions into one operation.  According to Intel, this boosts integer code performance by approximately 5% and floating point code by 9%.  
 

• Stack Execution Unit:  The Stack Execution Unit is used to update the stack pointer register.  Previously, tracking and updating the register was handled by a uop generated by the decoder; this engine removes the need for such an action and thereby increases decoder efficiency and decreases the number of instructions that must be kept in flight.
   
• Supports SSE / SSE2:  Dothan offers full support for Intel?s SSE and SSE2 instruction sets. 

 

(1 of 18) - next >

---

• The Pentium M Architecture (1 of 18)
• The Board and Chipset: Looks Can Be Deceiving (2 of 18)
• Board Layout (3 of 18)
• How Pentium M Compensates for 855, Overclocking (4 of 18)
• Test Setup: (5 of 18)
• Synthetic Tests: Sandra, ScienceMark, StarsCFD, Winrar. (6 of 18)
• Synthetic Tests (2): DIEP and Cachemem. (7 of 18)
• MP3 and DVD Encoding (8 of 18)
• 3D Rendering: Cinebench, 3DS Max, Lightwave (9 of 18)
• Gaming: How We Test (10 of 18)
• Game Performance: 3DMark 2001, 2003, 2005 (11 of 18)
• Gaming: Final Fantasy, Quake 3, UT2K3 & 2K4. (12 of 18)
• Gaming: Far Cry, Doom 3, Half Life 2 (13 of 18)
• Dothan Performance Torpedoes Industry Upgrade Myths (14 of 18)
• Performance Analysis: Dothan Strengths and Weaknesses (15 of 18)
• Thermals and Power. (16 of 18)
• Is Dothan Worth the Money? What About Alviso? (17 of 18)
• Why Intel Isn't Driving Dothan to Desktop, Conclusion. (18 of 18)