DDR2: Front And Center, Finally

Last week, AMD unveiled its DDR2 SDRAM-based mobile CPUs which, notably, also added dual-core capability. This week, hot on the heels of Dell's admission that it planned to add AMD to its server repertoire by year end, it's AMD's desktop processors' turn at the spotlight. Specifically, today AMD announced DDR2-and virtualization-cogniscent versions of many of its existing single- and dual-core CPUs, plus three new products:

• Athlon 64 FX-62: 2.8 GHz w/1 MByte L2 cache per core ($1,031, 1 ku)
• Athlon 64 X2 5000+: 2.6 GHz w/512 KByte L2 cache per core ($696), and
• Athlon 64 X2 4000+: 2 GHz w/1 MByte L2 cache per core ($328)

A bit over a month ago, I gave you the heads-up that AMD's DDR-to-DDR2 transition, exemplified by the new AM2 socket, was nearing lift-off. In then-confidential meetings, my AMD contacts were highly critical of the AnandTech benchmarking that I pointed to in my mid-April blog post. They suggested that Anand had employed immature silicon, motherboards and firmware that weren't reflective of final product performance.

My AM2 system is reportedly sitting in Sacramento, awaiting my return from WinHEC, so for now I'll need to rely on others' conclusions. And those conclusions support both Anand's earlier work and AMD's final statement on DDR-vs-DDR2, which I received on Monday, "A fair expectation for performance gain from 939-pin to AM2 is about 1% or more across various application-based benchmarks. That assumes equal model numbers for processors and an equal configuration. This also assumes premium memory is used for each configuration."

Take a look, for example, at page 5 of AnandTech's report published earlier today. Most of the DDR-400 vs DDR2-800 benchmarks run on the Athlon 64 X2 4800+ reveal a 0-to-2% performance increase for AM2. To be fair, DDR2-800 benchmark improvements ranged as high as 27%, and the results on one CPU can't necessarily be auto-extrapolated to all CPUs migrated from socket-939 to socket-AM2 (other analyses come from ExtremeTech and the sources listed in this Slashdot discussion).

Results will also be dependent on which core logic chipset you mate with the CPU. But generally speaking, Hannibal's words of wisdom from mid-April (which I've also inferred in my past print and online writeups) still ring true; AMD's "processors aren't yet bottlenecked by memory bandwidth". The studies I've read today do show a performance boost for the Athlon 64 FX-62 and Athlon 64 X2 5000+ versus their Athlon 64 FX-60 and Athlon 64 X2 4800+ predecessors, but it almost exclusively seems to be driven by clock-speed boosts.

And how do AMD's new CPUs stack up against the coming-soon Intel Conroe? Ironically, Intel's in very much the same situation that AMD was in a month and a half ago, with press folks getting their hands on immature silicon leaked by motherboard suppliers and other sources. One of my Intel contacts sent me a memo today that, among other things, clarified that "ANY REVIEWS YOU ARE SEEING NOW ARE USING NON-INTEL PROVIDED/SUPPORTED SAMPLES!"

However, unlike AMD a month and a half ago, Intel's Conroe is looking very strong in its infancy. I earlier shared with you data released at the earlier March Intel Developer Forum. Earlier today, Hexus leaked benchmark results that compared Conroe against both prior-generation (Extreme Edition 965) Intel silicon and current (FX-62)- and prior (FX-60)-generation AMD chips. After you've perused them, next consider the high likelihood that Conroe will support DDR2-800 memory and the rumours that Conroe's 'Extreme' variant will come with 1200 MHz FSB and 3 GHz core clock frequencies. Methinks AMD's got its work cut out for it. Stay tuned for my hands-on perspectives after I get hardware in front of me.

At the beginning of my late April feature article, I mentioned that "designing an embedded system based on PC-industry building blocks is like dancing with the devil. The chips and subsystems, including add-in cards, hard-disk drives, optical drives, and power supplies, are low-cost and abundant, thanks to the high-volume-manufacturing efficiencies of the PC market. However, although you measure your design's anticipated production life span in years or even decades, the fickle fortunes and fast evolution of the PC industry drive rapid obsolescence of your raw materials."

That perspective defines what I see as the real impact of today's AMD announcement, for the majority of EDN's readers. Two years ago, with the 915 and 925X chipsets, Intel drove DDR2 SDRAM into the PC platform for the first time. Now, AMD has also fully embraced DDR2, and with the now-united support of the top two x86 CPU suppliers, the DDR-to-DDR2 transition is at full stride. If you're about to start a new embedded system design, I don't recommend you base it on now-out-of-date DDR SDRAM. And what if your already-underway or -completed design employs DDR? In that case, I hope you've followed the advice in the subsequent few sentences of my late-April writeup, "Design smartly, planning the ability to later upgrade, and you'll be able to nimbly sidestep any supply-chain potholes. Failure to plan for future substitutions and advancements, on the other hand, means you'll soon—and perhaps repeatedly—redesign."