A Handful of Salt: Analyzing Early Bulldozer Benchmarks
Leaked benchmarks courtesy of openbenchmarking.org have shed some light on what we might see when the processor launches. David Kanter of Real World Technologies has compared the Bulldozer results against Magny-Cours performance, but warns that it's unwise to draw conclusions based on engineering samples. There's a dearth of information regarding which compiler options were selected or if AVX was enabled. Finally, there's clockspeed—the Bulldozer ES processors were running at 1.8GHz. AMD has previously indicated that Bulldozer will launch at clockspeeds well above this figure; current Magny-Cours chips top out at 2.3GHz—nearly 30 percent faster. We also don't know how fast the integrated memory controller was running—as Kanter points out, AMD's snoop filter may not have been functioning properly when these benchmarks were run.
Check RWT's article if you want specific numbers, but we'll say this: The results are erratic. Bulldozer soundly beats Magny-Cours in some areas, only to fall well behind in others. This loosely matches our expectations given Bulldozer's shared FPU, but the size of the gap between MC and BD may well be due to other intervening variables.
We don't have enough information to draw any solid conclusions. The comparative data suggests that Bulldozer may have traded significant performance increases in certain areas for significant decreases in others. Without further information on compiler flags and the ES chips themselves, it's impossible to judge whether or not the bet paid off. AMD has always claimed that Bulldozer was designed to meet the core needs of the server market while conserving die space and minimizing production costs. It's not unreasonable to think that Sunnyvale is counting on higher clockspeeds and larger caches to achieve or surpass Magny-Cours in sub-optimal workloads.
If the performance gap we see here reflects real-world measurements, it'll be difficult for Bulldozer to match Sandy Bridge. AMD's attempt to compete with Intel by selling cores at a 2:1 ratio in the server market last year has yielded scant results. The implication is that Bulldozer may actually end up losing to Intel's Sandy Bridge-based Xeons by a wider margin than as present.
At the moment, Bulldozer's performance eerily resembles the P4—and that's not exactly a ringing endorsement. With any luck, proper compiler flags and the new AVX instructions will minimize any performance losses we might see when comparing next-generation processors. For now, we're cautiously optimistic. It's important to keep expectations in line with reality—Bulldozer, even if it's great, isn't likely to bound the gap between AMD and Intel in one amazing product generation.