When AMD launched the Athlon 64 last September its debut was marred by a great deal of uncertainty over what platforms would provide the CPU?s long-term future.  In contrast to the stability of Socket 462 which debuted in 2000 and has continued to the present day, AMD chose to launch a Socket 754-based Athlon 64 and a Socket 940-based Athlon FX (basically an Opteron of a different name).  At the same time, Sunnyvale announced its intention to migrate both CPU?s to a new, dual-channel Socket 939 platform.  This new platform would allow four-layer boards (rather than the current six) and would be the basis for volume Athlon 64 production.

The consequence of this, however, was that the Socket 754 platform was seen as an interim alternative similar to Intel?s Socket 423, despite AMD?s assurances that it would be supported after the introduction of Socket 939.  With second-generation chipsets from NVIDIA (and to a lesser extent, VIA) offering new features and abilities, the launch of Socket 939 was seen by many as the point where Athlon 64 would enter the mainstream market as the ?real? next-generation AMD platform.  Launch day is here?is
Socket 939 ready to fly?

NOTE:  Due to the severe storms that struck the Midwest on May 27th, I was without power from Thursday night to Saturday afternoon, and was unable to work most of Sunday due to continued severe weather.  This has impacted the number of systems we were able to test for the review.  Tornadoes tend to make it difficult to benchmark, though they could provide an interesting alternative for measuring just how fast these processors really are.

Athlon 64 ?Newcastle? and Athlon 64 FX-53:

The Socket 939 Athlon FX is identical to the 940-pin version in virtually every way.  It has 1 meg of L2 cache, runs at 2.4 GHz, and has a 193mm sq die.  Available memory bandwidth is 6.4 GB/s, just as with the 940.  There are only two  differences between the two:  is Hyper-Transport speed and the ability to use unbuffered memory:  Socket 940 uses an 800 MHz DDR connection (6.4 GB/s of HT bandwidth) while Socket 939 uses a full 1 GHz, for a total of 8 GB/s of HT bandwidth.  The lifting of the buffered memory restriction is exactly what it sounds like?unlike Socket 940 Athlon 64 FX parts, Socket 939 parts can use higher-speed, enthusiast-oriented unbuffered RAM. 

Unlike its big brother, the 939-pin version of the Athlon 64 is notably different than its 754-pin sibling.  When AMD introduced slower Athlon 64?s running with only 512K L2 cache earlier this year, several sites erroneously reported this as the quiet launch of AMD?s ?Newcastle? core.  This was incorrect.  The 754-pin chips with 512K cache were standard Athlon 64?s with half the cache disabled.  The 939 pin Athlon 64 AMD is launching today is the ?true? Newcastle, and has only 512K of L2 cache, period.  This has allowed AMD to shrink both processor count and die size from 105.9 to 68.5, and from 193 mm sq to 144 mm sq and in turn, will allow Sunnyvale to produce more CPU?s per wafer?a critical advance, given AMD?s limited production space.

In order to compensate for Athlon 64?s halved cache, AMD has boosted the CPU in three ways:

Increased Hyper Transport Bandwidth:  Like the 939 FX-53, the Athlon 64 now uses a 1 GHz DDR (2 GHz equiv.) Hyper Transport bus. 

Higher Clockspeed:  Previously, the highest-end Athlon 64 typically ran a speed-grade slower than the FX, save for brief periods of overlap.  This has now changed.  FX and A64 parts will now run at the same speed, with the difference between the two being the FX?s larger L2 cache.

Double-wide Memory:  The Athlon 64 is also the proud recipient of a double-wide trailer?err, memory bus.  This move may make the marketing whelps happy, but as we?ve repeatedly discussed, the real-world value of a 128-bit wide or dual-channel memory controller is minimal

The following chart should help with understanding AMD?s new speed grades and cache amounts:

We?ve no word on whether AMD intends to release 939 speed grades down past 3500+, but based on their current numbering convention, we?d guess that such parts would be odd-numbered in order to avoid confusion.  Its not impossible that we?ll see a 3300, 3100, 2900, etc.  The 2600+ isn?t formally announced yet, but is included as its widely expected to quietly appear. 

next >

---

• Does 3800+ Deserve the Title? (1 of 11)
• Test Configuration (2 of 11)
• Benchmark Methodology and Updates (3 of 11)
• Synthetic / Scientific / Mathematical Results (4 of 11)
• Scientific / Mathematic Benchmarks, Cont. (5 of 11)
• Office Performance (6 of 11)
• Multimedia Encoding (7 of 11)
• 3D Rendering (8 of 11)
• Gaming (9 of 11)
• Performance Analysis (10 of 11)
•  Conclusion:  An Achilles Heel the Size of an Elephant (11 of 11)