New Bulldozer Benchmarks Leak, Mostly Miss the Mark

With Bobcat and Llano launched, AMD has one more major product overhaul set for this year. The company's Bulldozer CPU will launch in the next few months, and after years of waiting, enthusiasts and IT industry analysts are both curious to see what AMD has in its high performance pipeline. A Turkish website, Donanim Haber, recently got its hands on what they claim is a Bulldozer FX processor--but while we're going to talk about the site's findings, we'd recommend readers take them with some caution.

The new octal-core processor is an AMD FX-8130P running at 3.2GHz base speed, what's reported as a 3.7GHz Turbo speed, and a 4.2GHz clock speed if only half the CPU's cores are in use. This implies that the new chip mixes TDP modes much more flexibly than previous six-core Thuban processors. AMD is therefore guaranteeing 3.2GHz operation, offering 3.6GHz (on all eight cores) if conditions are right, and implying that 4.2GHz might happen if four or fewer cores are in use.

The Turkish story also implies that we'll see FX return as an entire product brand affixed to Bulldozer processors as opposed to a top-end designation. There are some discrepancies between the clockspeeds and features of the CPUs listed as forthcoming in the accompanying chart vs. the FX-8130P running on the site's testbed, but here are the basics:  According to DH, we'll see 8xx, 6xx, and 4xx parts, with the first digit corresponding to the number of CPU cores. Clockspeeds will range from 3.6GHz base / 4.2GHz Turbo on the top-end FX-8150 to 3.6GHz base / 3.8GHz Turbo on the quad-core FX-4100. The amount of L3 cache will also vary, from 8MB on the 8xxx parts to 4MB on the 4xxx chips.

Here's where we have to interject a note of caution. DH ran a number of benchmarks, but virtually all of the results are worthless. That's a harsh thing to say, so let us explain our reasoning.

3DMark 11 is a graphics subsystem test. Based on the program's inherent statistical weighting, CPU differences can account for 25 percent of the total score, at most. PCMark 7 is designed to measure the performance of an entire system--it's great as part of a suite of tests, bad for examining architectural differences. This goes double when we have no idea what storage solution was installed or whether AHCI was enabled/disabled. Super Pi is a ridiculously outdated, inapplicable metric that equates to using triangle throughput as a metric for 3D purchases. The AIDA64, results, meanwhile, either mis-detected the CPUs clockspeed, or were run on an overclocked flavor, not the base model.

That leaves us with Cinebench and Fritz Chess. The DH team unfortunately used Cinebench R10, not 11.5. This is a major problem--CB R10 scales very badly above four cores, as we've previously demonstrated when comparing performance on both AMD and Intel processors. Regardless of architecture, the test plateaus. In our review of the Digital Storm Enix, we showed that the Maingear Shift (with a six-core processor) was just six percent faster than the quad-core Sandy Bridge in CB10, but was 16 percent faster in CB 11.5.

The Fritz Chess result of 14,197 suggests that a 3.2GHz eight-core Bulldozer is ~23 percent faster than a 3.2GHz six-core Thuban. Since an eight core chip has 33 percent more cores than a six-core chip, that's precisely the sort of scaling we'd expect to see from a chip with two additional cores bolted on.

This is why company's work hard to control how and when information is shared with the public. What looks like solid data is rendered useless due to poor benchmark choices, less-than-full disclosure, and an apparent lack of knowledge about newer tests that have replaced old standards due to a need to support multi-core processing.