x86, X64 Instruction Latency, Memory Latency and Cpuid Dumps | Hacker News Reader

x86, X64 Instruction Latency, Memory Latency and Cpuid Dumps

instlatx64.atw.hu

38 pointsarunc9 comments

What about perf counters for these things? Some did exist. My i7-4770K has many not present on newer models, especially in regards to uncore dram control.

Did they get removed or are they just not populated in the perf binding for the kernel?

Can someone explain how to read the readings from Instruction Latency(InstLatX86).

For example how to read the following line,

  201 X86   :INC r8                      L:  13.36ns=  1.0c  T:   6.68ns=  0.50c

This instruction is incrementing the value of r8 general purpose register. But I am not sure how to parse L: and T: section.

Also, does this number change proportionally if it is in low power mode or if the CPU is undervolted?

There are some explanation in Comment page:

  In this dump latency means the time that it takes for the next dependent same-type instruction to start. Throughput means the time that it takes for the next independent same-type instruction to start:

L is the latency of the instruction, so 1 clock cycle after the INC r8 instruction another instruction can operate on the result in r8. T is the reciprocal throughput, meaning in this case that the CPU can sustain 2 INC instructions per clock cycle, if there are no data dependencies of course.

In general these numbers will stay the same when counted in clock cycles, regardless of frequency.

To add: with pipelined processors, it doesn’t make sense to list “clocks per instruction” as a single number. Since the advent of superscalar processors, it makes even less sense. Hence why there’s “latency” and “throughput”.

Any link to the text version of those charts?

What's "GenuineIotel"? Did an actual processor report this ID?

Yes, one of the CPUID arguments returns a string. Normally, the values are either "GenuineIntel" or "AuthenticAMD" in ASCII, marshalled as 3 32-bit registers so 12 bytes in total.

I know that, but this one ("Iotel") either has a bit-flip or is some kind of fake?

Reply on news.ycombinator.com