3 ms·

Generally raising a processor's clock frequency will decrease its IPC as it spends more time waiting for main memory, so it isn't surprising that the overall IP

Generally raising a processor's clock frequency will decrease its IPC as it spends more time waiting for main memory, so it isn't surprising that the overall IPC only went up 3% despite the various cache optimizations. I wouldn't have actually been surprised if it had stayed the same. But I'm frankly sort of amazed that they've managed to go to having consistently lower cache latencies than Intel has. I always thought that Intel's greater willingness to throw engineer-hours at optimizing layouts would mean they'd always take less clock cycles to hit their caches.

sp332 9y ago
I don't think IPC tests involve main memory I/O. Edit: I missed that the benchmark methodology is on a different page from the results.
- dragontamer 9y ago
  An IPC test is basically running a benchmark while locking the processor's Frequency (through overclocking tools). That way, you get to test the architecture without dealing with all the Turbo-boost stuff on modern chips. Memory latency / bandwidth is a piece of any benchmark. Some benchmarks are purely CPU (Linpack), but others are better overall tests of the system (ex: Cinebench or Blender)
- Symmetry 9y ago
  IPC is always going to be a number that is dependent on your particular workload. Searching through a huge linked list laid out randomly in memory is going to measure just your memory latency. An optimized media codec might tend to just measure your execution resources. But really you want to measure all sorts of additional things like the out of order engine depth, the strength of the branch predictor and prefetcher, and the latencies and sizes of all the different cache levels. And it's hard to get a good measure of all the stuff within a chip and especially stress the branch predictor and prefetcher without having a working set that sometimes spills over into main memory. And really you want to include the efficiency of your memory controller in your rating of the chip too. But if you have a particular workload in mind some of these aspects might not important to you and you'd be best served by just benchmarking your particular application on a given chip.

sp332 9y ago

I don't think IPC tests involve main memory I/O. Edit: I missed that the benchmark methodology is on a different page from the results.

dragontamer 9y ago
An IPC test is basically running a benchmark while locking the processor's Frequency (through overclocking tools). That way, you get to test the architecture without dealing with all the Turbo-boost stuff on modern chips. Memory latency / bandwidth is a piece of any benchmark. Some benchmarks are purely CPU (Linpack), but others are better overall tests of the system (ex: Cinebench or Blender)
Symmetry 9y ago
IPC is always going to be a number that is dependent on your particular workload. Searching through a huge linked list laid out randomly in memory is going to measure just your memory latency. An optimized media codec might tend to just measure your execution resources. But really you want to measure all sorts of additional things like the out of order engine depth, the strength of the branch predictor and prefetcher, and the latencies and sizes of all the different cache levels. And it's hard to get a good measure of all the stuff within a chip and especially stress the branch predictor and prefetcher without having a working set that sometimes spills over into main memory. And really you want to include the efficiency of your memory controller in your rating of the chip too. But if you have a particular workload in mind some of these aspects might not important to you and you'd be best served by just benchmarking your particular application on a given chip.