You misunderstand the article. Bulldozer has 4 modules each of which has 2 tightly linked pipelines that share fetch and decode resources. It's a radical new approach towards multithreading called cmt. So Bulldozer essentially has the execution resources of 8 cores and the fetch/decode resources of 4. It will show up as 8 logical cores though. But yes dolphin will use 2 threads on bulldozer just like all other x86 cpus.
I'm curious about performance criteria of dolphin-emu. What is the most important aspect in the cpu that provides higher performance in dolphin-emu?
Is it raw clock speed or ipc or fpu? Thanks.
Raw clock and memory bandwidth.
And architecture (i7 processors have the best performance at this moment for dolphin).
Quote:Raw clock
I think I get what you mean to say...but it sounds like your talking about JUST the clock rate. Which would be untrue. After all a 3GHz athlon X2 will perform like shit next to a 2 GHz core 2 duo even if it has more memory bandwidth. Likewise xtreme2damax saw no difference in performance overclocking his e8400 (3GHz core 2 duo I believe) to 4GHz yet a massive gain when he upgraded to a 3GHz i7. Even when underclocked to 2 GHz and the ram underclocked to 1066MHz it still massively outperformed the 4GHz 1666 MHz FSB core 2 duo. Clock rate is only good for comparing performance of the same chip. So I think what you meant to say was cpu throughput such as FLOP/s (clock rate x operations per clock with bottlenecks factored in). And memory bandwidth seems to only make a difference when using efb to ram. Otherwise it seems to have no effect...just something to think about.
Quote:What is the most important aspect in the cpu that provides higher performance in dolphin-emu?
The overall architecture. Many factors play a role. Dolphin is a very floating point intensive application so FLOP/s would be a good measurement. But you must also take into account that dolphin will only use 2 threads (floating point synthetic benchmarks are multithreaded and use ALL physical/logical threads, simply divide the result to get the numbe rof physical threads down to 2, so divide by 2 if it;s a quad core 3 if it's hexacore etc.) and of course efb to ram is very taxing on memory bandwidth so if you're using that option memory bandwidth is also extremely important. Synthetic benchmarks obviously aren't a super accurate way to guage performance in an application like dolphin but they give you a general idea of where a cpu ranks in certain areas.
ok, thanks for the tip.
So, when evaluating dolphin-emu cpu performance
#1 FPU scores matters more
#2 clock speed matters when * comparing the same micro-architecture/model of cpu *
#3 memory bandwidth/channel matters when using option 'efb to ram'
IIRC, AMD Bulldozer seems to have a shared FP scheduler between two cores in the same module. What is the FP advantage of Intel Sandy Bridge's ring bus?
When Bulldozer debuts next year, it will compete with late Sandy Bridge models or early Ivy Bridge models. Note that Ivy Bridge entry models are all quad-cores.
Another question, how does cache levels/sizes and cpu instructions impact dolphin-emu performance? Any specific criteria? Thanks.
Quote:What is the FP advantage of Intel Sandy Bridge's ring bus?
What? The ring bus won't affect FLOP/s. It just affecst bandwidth between the cores/cache/memory controller/northbridge.
Quote:Note that Ivy Bridge entry models are all quad-cores
Ivy Bridge is not coming until 2012. Early next year the 2 and 4 core sandy bridge cpus using socket 1155 (replaces 1156) are coming out. Late next year 6 and 8 core sandy bridge cpus using socket 2133 (replaces 1366) are coming out. Bulldozer is coming out somewhere in between those two (middle of next year assuming they hold to the schedule, which they rarely do).
Quote:Another question, how does cache levels/sizes and cpu instructions impact dolphin-emu performance? Any specific criteria? Thanks.
More cache = SMALL improvement in performance per clock. As for cpu instructions I assume you're talking about extensions to the instruction set such as sse3/ssse3/sse4a/sse4.1/sse4.2/sse4.3/sse5/avx? Depends. A regular build of dolphin won't use the new instructions. Regular dolphin builds use sse2. Some builds are "optimized" by the compiler using newer instructions but the results seem to vary depending on settings, what game, and from person to person. Some people claim significant speedups while others see no difference at all, I personally have never seen any difference in performance between sse2/sse3/ssse3 builds on my machine but that doesn't mean I can speak for all users, for the people who do see and improvement in performance it is usually small.
(11-19-2010, 08:19 PM)pokipoki08 Wrote: [ -> ]Another question, how does cache levels/sizes and cpu instructions impact dolphin-emu performance? Any specific criteria? Thanks.
In theory, this should have a large impact on Dolphin's performance. In practice, the impact is quite small. The code that Dolphin generates to translate the Wii instructions into PC instructions always overflow the cache limiting the cache's effectiveness.