AMD product rollouts, pricing moves send mixed messages

Mid-June, I chastised AMD for pricing itself out of a substantial chunk of the dual-core consumer PC market with the expensive Athlon 64 X2 versus Intel's Pentium D, a situation made all the more ironic by AMD's subsequent (end of June) antitrust lawsuit against its much bigger competitor. I'm happy to report that the pricing on AMD's latest Athlon 64 X2 variant, the 3800+ (2 GHz core clock speed, with 512 Kbytes of L2 cache per core), announced at the beginning of this month, is much more reasonable; $354 for 1,000 units. For all intents and purposes, it's an Athlon 64 X2 4200+ with loosened test specs to maximize manufacturing yield. It's still over $100 more than Intel's several-month-old published pricing for the 2.8 GHz Pentium D, but at least it's not more than twice the price of the Intel alternative (as the Athlon 64 X2 4200+ was). And as the AnandTech and ExtremeTech benchmarking suggests, it's a capable performer (one that, not surprisingly, is also reasonably overclock-friendly). There's more discussion at Slashdot.

Meanwhile, a few weeks ago while I was at Siggraph, AMD briefed me on its newly announced variants of the Opteron 100 Series CPUs supporting non-registered (aka un-buffered) DRAM. A quick review; AMD intends for Athlon processors to find use in conventional desktop and notebook PCs, with Opteron chips going in workstations and servers (akin to Intel's positioning of Pentium versus Xeon). Opteron 100 Series CPUs support single-processor system configurations, Opteron 200 chips enable two-CPU designs, and Opteron 800 processors go into four-way systems. Un-buffered memory modules dispense with the incremental chip cost of registers, the freed-up DIMM area can potentially be used to squeeze more DRAMs onto the module, and by eliminating the signal propagation delay of the registers they tend to have faster random access latencies. Conversely, because the lack of intermediary buffers means that multiple DRAMs present multiple input loads to the microprocessor's integrated memory controller outputs, there's a finite limit to how many DRAM chips you can put on an un-buffered module.

Focus in particular on one phrase in the previous paragraph; "Un-buffered memory modules dispense with the incremental chip cost of registers". Although the correlation between cost and price is at best tenuous, that statement implies that all other factors being equal, an un-buffered DIMM of given capacity should be cheaper than a buffered alternative. You might extrapolate this fact to its logical conclusion, that systems based on un-buffered DIMMs are cheaper than ones based on buffered DIMMs. And you might be right. Or not. For as a writeup in The Register points out, AMD substantially raised the prices on all of its Opteron 100 series chips in conjunction with the un-buffered version introduction. A reasonable person might conclude that AMD's conveniantly decided to shift the percentage of system cost previously devoted to DIMM registers over to itself, improving its CPUs' profit margin in the process. Or at least that's what I've concluded.

AMD's Opteron 100 chips are single-core for the moment; dual-core CPUs (already represented, as the first paragraph points out, in the Athlon line) are promised by month-end. Regardless, AMD's got a real positioning problem on its hands; from a feature set standpoint, Athlon 64 and un-buffered Opteron 100 chips are near-identical. I don't care how much nebulous hand-waving the company does about the 'incremental testing' an Opteron chip gets over its Athlon counterpart; if I was locked in an entry-level workstation or server pricing battle with a big Intel customer like Dell, I'd be sorely tempted to develop my AMD-based design around an economical Athlon 64 or Athlon 64 X2 CPU, versus a much more expensive Opteron 100 device. How about you?