AMD Ships 12-16 Core Opteron Processors Based on Bulldozer Design. It's Depressing

AMD has announced the immediate availability of new 12-16 core 32nm Opteron processors based on the company's new Bulldozer architecture. According to AMD, these new chips will provide better performance, virtualization scaling, and efficient economics.

“Our industry is at a new juncture; virtualization has provided a new level of reliable consolidation and businesses are now looking to the cloud for even more agility and efficiency. We designed the new AMD Opteron processor for this precise moment,” said Paul Struhsaker, corporate vice president and general manager, Commercial Business, AMD. “The wait for the most anticipated new product and architecture for servers is over. Leading OEMs are now offering cloud, enterprise and HPC customers a full suite of solutions based on the industry’s most comprehensive server processor portfolio, the new AMD Opteron family of processors which deliver an inspired balance of performance, scalability and efficiency.

Unfortunately, preliminary data—even the one-sided information AMD makes available in its own press release—raises serious questions about the accuracy of the company's claims.

In its announcement, AMD claims an "89 percent greater performance* than the most popular server processor* from the competition at the same price*." The asterix refer to the company's original qualifying statements and it's a bad sign when you need three of them in the same sentence.

AMD's Interlagos. The observant will note that it looks just like two Bulldozer's, glued together

AMD's competitive claim rests on Linpack performance, where a brace of Opteron 6276 processors (32 cores in total) are 89 percent faster than a pair of Xeon 5670's. Hyper-Threading was disabled in this test (Linpack's performance tends to drop with HT enabled), which means we're comparing apples-to-apples as far as core counts. 32 Bulldozer cores, in other words, are just 89 percent faster than 12 Westmere cores. That's an ugly leg to stand on, particularly when we compare the clock speeds and power envelopes of the new Opteron 6200 and 4200 series against the current Magny-Cours/Lisbon parts.

The old Magny-Cours 12-core chips topped out with the Opteron 6180 SE, with a clock speed of 2.5GHz, an integrated memory controller at 1.8GHz, and a 105W TDP. The new Opteron 6282 SE is a 16-core chip clocked at 2.5GHz, a 2GHz IMC, and a blistering 140W TDP.

The increased core counts would effectively balance the new products' increased TDP if Interlagos was as efficient as its predecessor—but it isn't. Research has shown that an eight-core Bulldozer scales more like a six-core Thuban; the chip takes a performance hit of 15-20 percent from sharing so much core logic. Toss in Bulldozer's IPC loss and weaker single-threaded performance, and the 16-core Interlagos chips will likely have real problems distinguishing themselves from 12-core Magny-Cours.

What's worse, an apples-to-apples core count comparison shows AMD was unable to hold TDP levels constant between Magny-Cours and Interlagos. A 12-core Opteron 6176 at 2.3GHz has an 80W TDP. A 12-core Opteron 6234 at 2.4GHz has a 115W TDP. Even in ideal server workloads and best-case scenarios, the Opteron 6234 is clocked four percent faster than its predecessor and can draw up to 43 percent more power. While it's true that TDPs tend to improve over time, it's unlikely they'll improve enough to equalize between the two chips.

AMD's positioning efforts are also badly undercut by the realities of software licensing. A great deal of server/workstation software is licensed per core, and vendors rarely distinguish between AMD and Intel as far as setting rates. Oracle, for example, charges a 0.5x multi-core factor for both Intel and AMD CPUs. If we assume a flat $1000 per-core license price (to keep the math simple), it would cost a company $3000 per year to license a six-core Westmere server and $8000 per 16-core Interlagos server.

Based on AMD's quoted (and presumably best-case) FLOPs results, the two systems would offer nearly equal performance—but one costs far more per year as far as licensing expenses.AMD can try to offset that expense by offering chips at much lower prices than Intel, but only to a limited degree. Over 3-4 years, license costs will eventually dwarf the one-time expense of acquiring new hardware.

Sunnyvale has already announced a handful of design wins and Interlagos deployments, and it'll undoubtedly score a few more, but the combination of high software licensing costs, high TDPs, and low IPC virtually guarantee that the company's new server chip will do very little to improve its competitive positioning against Intel.