On September 3rd, SiSoftware published a review of AMD’s new Zambezi processor, based on the new Bulldozer architecture.
After only about 4-5 hours, SiSoftware had to take it down again, but what goes on the internet, stays on the internet, it was already cached by Google: http://webcache.googleusercontent.com/search?q=cache:QLn9rR2rn5oJ:www.sisoftware.net/?d=qa&f=cpu_amd_bulldozer&l=en&a=+sisoftware+bulldozer&cd=2&hl=en&ct=clnk&gl=en
AMD, in the person of John Fruehe, was quick to dismiss the benchmarks as fake… But how fake are they?
- AMD started shipping Bulldozer CPUs on the 7th of September, only a few days after these benchmarks were published. So while AMD claims that all benchmarks are done with pre-production silicon, BIOS, microcode and more… how pre-production are they? (this late in the process, I would guess that pre-production systems would be within about 5% of the final product).
- SiSoftware is not just any random site that leaks some random benchmarks. SiSoftware has been developing benchmarks for many years, and has been a popular benchmarking tool for many high-profile review sites for a long time. What reason would SiSoftware have to leak false benchmark figures? It would only hurt their core business (which is selling benchmark software, not getting website hits).
- It seems too elaborate to be a total fake… If it is not actually SiSoftware who published this page, then it would be a pretty impressive hack to get such a page online, and in the Google cache. Also, SiSoftware has not denied that this page is real.
- The figures, as well as the accompanying analysis seem to make a lot of sense. This is not just some random guessing. It is written by someone who at least understands the Bulldozer architecture and the performance characteristics. It makes perfect sense that it is slower than Sandy Bridge in most tasks, and peaks at the crypto benchmark with ALU-only. After all, the Bulldozer has 8 integer cores, with 2 ALUs each, where Sandy Bridge has only 4 cores, with 3 ALUs each. So that is the one benchmark where you would expect Bulldozer to excel. Everything else seems to be exactly how you expect things to be between Intel and AMD, historically: Intel has the better memory controller, cache and SIMD performance. Those aren’t gaps that I would expect AMD to close in just a single generation.
Below is a capture of the entire page, in case the Google cache gets invalidated in the future.