|AMD FX-8350 Vishera Desktop Processor|
|Reviews - Featured Reviews: Processors|
|Written by David Ramsey|
|Tuesday, 23 October 2012|
Page 12 of 15
Piledriver Core Performance
As we saw in our original FX-8150 review, and indeed in this review, the FX CPU's weakness is in its per-core performance; or, more succinctly, instructions per clock cycle. Think of how you'd make a car faster: you could add more horsepower, or make it lighter, the idea in both cases being to increase the horsepower-per-pound ratio. Adding horsepower is easy: just make the engine bigger or add a turbo or supercharger. Making the car lighter, on the other hand (while staying within safety regulations), is both complex and expensive.
It's the same with processors. You can increase performance by increasing the instructions a core can execute per clock cycle, or by increasing the clock frequency and adding more cores, and the idea in either case is to increase the number of instructions executed per unit of time. Modern superscalar processors are insanely complex: x86 instructions must be broken down into microcode; pipelines must be managed, along with on-chip caches; branch prediction and speculative execution must be optimized, and there's even more that only processor architects could hope to understand. The point is that increasing instructions per clock is hard, but simply ramping up the clock speed (and adding cores) is a lot easier. This is the route AMD has taken with its FX series CPUs. Still, AMD says that they've made a number of architectural improvements in the Piledriver core over the previous Bulldozer core:
To test these improvements without considering core count or clock speed, I locked both my FX-8150 and FX-8350 at 3.6GHz and disabled turbo mode. I then ran some benchmarks to compare per-core performance at the same clock speed. Handbrake and Blender were both run only on a single core.
As you can see, the Piledriver cores are faster than the Bulldozer cores. If the percent improvement doesn't seem that dramatic, keep in mind that it's about the same level of improvement that we saw in the transition from Sandy Bridge to Ivy Bridge in the Intel world. Of course, Bulldozer is starting from a much lower base: as we noted in our FX-8150 review, the Bulldozer cores provide about the same performance overall as the much older Thuban cores in AMD's Phenom II X6-1100T processor. So while AMD has made progress, they still have a long way to go.
The Piledriver cores in the FX-8350 benefit from the same new instructions we saw introduced in the FX-8150. Specifically, these are:
The problem with introducing new instructions is that they don't benefit existing programs: new programs must be specifically coded to use them. And as our tests with the FX-8150 and modified versions of the x264HD benchmark showed, substantial performance increases are possible if programs are coded to use these new instructions. The FMA (fused multiply-add) instructions in particular can help video transcoding applications, but these new instructions have yet to achieve any broad penetration in the program market.
Each Bulldozer/Piledriver module shares L2 cache between its two integer cores, so if a multi-threaded process can be spread among modules so that (for example) two threads each execute on separate modules rather than in the same module, each will get twice as much L2 cache as it would otherwise, potentially increasing performance. Microsoft has released hotfixes for Windows 7 that supposedly addressed this issue, but they didn't seem to have any measurable effect on performance. Windows 8 may improve this...we'll see.