Detecting CPU Speed

Jump to navigation Jump to search

What is CPU Speed

"CPU speed" has several different definitions:

  1. How quickly the processor can execute code (e.g. instructions per second)
  2. How fast the processor's clock is running (e.g. cycles per second)

How quickly a CPU can execute code is important for determining the CPU's performance. How fast a CPU's clock is running is only useful for specific cases (e.g. calibrating the CPU's TSC to use for measuring time).

There are also several different measurements for these different "CPU speeds":

  1. Best case
  2. Nominal case
  3. Average case
  4. Current case
  5. Worst case

For example, look at a modern Intel Core i7 CPU (with turbo-boost, power management and hyper-threading). The best case instructions per second would occur when:

  1. There is no throttling/power saving at all.
  2. Only one logical CPU is running (turbo-boost activated and hyper-threading not being used)
  3. Simple instructions with no dependencies in a loop that fits in the CPU's "loop buffer" are being executed
  4. There are no branch mispredictions
  5. There are no accesses to memory (no data being transferred to/from caches or RAM)

The worst case instructions per second would be the exact opposite and may be several orders of magnitude worse (e.g. a best case of 4 billion instructions per second and a worst case of 100 million instructions per second). The nominal instructions per second (commonly referred to as "nominal cycles per second") is an estimation of the normal average instructions per second a developer would expect. All of these things are fixed values - a specific CPU always has the same best case, worst case and nominal case, and these values don't change depending on CPU load, which instructions are/were executed, etc.

The current instructions per second is the instructions per second at a specific instant in time and must be somewhere between the best and worst cases. It can't be measured exactly, but can be estimated by finding the average instructions per second for a very short period of time. The average case is something that has to be measured. Both the current instructions per second and the average instructions per second depend heavily on the code that was running. For example, the average instructions per second for a series of NOP instructions may be much higher that the average instructions per second for a series of DIV instructions.

General Method

In order to tell what the CPU speed is, we need two things:

  1. Being able to tell that a given (precise) amount of time has elapsed.
  2. Being able to know how many 'clock cycles' a portion of code took.

Once these two sub-problems are solved, one can easily tell the CPU speed using the following :

prepare_a_timer(X milliseconds ahead);
while (timer has not fired) {
   inc iterations_counter;
cpuspeed_mhz = (iteration_counter * clock_cycles_per_iteration)/1000;

Note that except for very special cases, using a busy-loop (even calibrated) to introduce delays is a bad idea and that it should be kept for very small delays (nano or micro seconds) that must be complied with when programming hardware only.

Also note that PC emulators (like BOCHS, for instance) are rarely realtime, resulting in the clock appearing to run faster than expected.

Waiting for a given amount of time

There are two circuits in a PC to deal with time: the PIT (Programmable Interval Timer - Intel 8253 and 8254 are common PITs) and the RTC (Real Time Clock). The PIT is probably the better of the two for this task.

The PIT has two operating modes that can be useful for telling the CPU speed:

  1. the periodic interrupt mode (0x36), in which a signal is emitted to the interrupt controller at a fixed frequency. This is especially interesting on PIT channel 0 which is bound to IRQ0 on a PC.
  2. the one shot mode (0x34), in which the PIT will decrease a counter at its top speed (1.19318 MHz) until the counter reaches zero.

Whether or not an IRQ is fired by channel0 in 0x34 mode should be checked.

Note that theoretically, one shot mode could be used with a polling approach, reading the current count on the channel's data port, but I/O bus cycles have unpredictable latency and it is up to the programmer to make sure the timestamp counter is not affected by this approach.

Knowing how many cycles your loop takes

This step depends on the CPU. On 286, 386 and 486, each instruction takes a well-known and deterministic amount of clock cycles to execute. This allows the programmer to tell exactly how many cycles a loop iteration took by looking up the timing of each instruction and then summing them up.

Since the multi-pipelined architecture of the Pentium; however, such numbers are no longer communicated (for a major part because the same instruction could have variable timings depending on its surrounding, which makes the timing almost useless).

It is possible to create code which is exceptionally pipeline hostile such as:

xor eax,edx
xor edx,eax
xor eax,edx
xor edx,eax

A simple xor instruction takes one cycle, which guarantees that the processor cannot pipeline this code as the current instructions operands depend on the results from the last calculation. One can check that, for a small count (tested from 16 to 64), RDTSC will show the instruction count is almost exactly (sometimes off by one) the cycles count. Unfortunately, when making the chain longer, code cache misses will occur, ruining the whole process.

E.g. looping on a chain of 1550 XORs may require a hundred of iterations before it stabilizes around 1575 clock cycles on a AMDx86-64 and will take an exceedingly long time to stabilize on a Pentium3.

Despite this inaccuracy, it gives relatively good results across the whole processor generation given a reasonably accurate timer. If very accurate measurements are needed the next method should prove more useful.

A Pentium developer has a much better tool to tell timings: the Time Stamp Counter: an internal counter that can be read using RDTSC special instruction

rdtscpm1.pdf explains how that feature can be used for performance monitoring and should provide the necessary information on how to access the TSC on a Pentium.

RDTSC Instruction Access

The presence of the Time Stamp Counter (and thus the availability of RDTSC instruction) can be detected through the [CPUID] instruction. Calling CPUID with eax=1, will put feature flags in edx. TSC is bit four of that field.

Note that prior to using the CPUID instruction, make sure the processor supports it by testing the 'ID' bit in eflags. (This is 0x200000 and is modifiable only when CPUID instruction is supported. For systems that doesn't support CPUID, writing a '1' at that place will have no effect.)

In the case of a processor that does not support CPUID, you'll have to use more eflags-based tests to tell if a 486, 386, etc. is running and then pick up one of the 'calibrated loops' for that architecture (8086 through 80486 may have variable instruction timings).

Working Example Code

A Real Mode Intel-copyrighted example is present in the above-mentioned application note. Another code snippet submitted by user DennisCGC that will give the total measured frequency of a Pentium processor follows.

Some notes:

  • irq0_count is a variable, which increases each time when the timer interrupt is called.
  • In this code it's assumed that the [PIT] is programmed to 100 Hz (The formula to calculate a desired Hz is provided in the snippet.)
  • The code assumes that the command CPUID is supported.
   ;first do a cpuid command, with eax=1
   mov  eax,1
   test edx,byte 0x10      ; test bit #4. Do we have TSC ?
   jz   detect_end         ; no ?, go to detect_end
   ;wait until the timer interrupt has been called.
   mov  ebx, ~[irq0_count]


   cmp  ebx, ~[irq0_count]
   jz   wait_irq0
   rdtsc                   ; read time stamp counter
   mov  ~[tscLoDword], eax
   mov  ~[tscHiDword], edx
   add  ebx, 2             ; Set time delay value ticks.
   ; remember: so far ebx = ~[irq0]-1, so the next tick is
   ; two steps ahead of the current ebx ;)


   cmp  ebx, ~[irq0_count] ; Have we hit the delay?
   jnz  wait_for_elapsed_ticks
   sub eax, ~[tscLoDword]  ; Calculate TSC
   sbb edx, ~[tscHiDword]
   ; f(total_ticks_per_Second) =  (1 / total_ticks_per_Second) * 1,000,000
   ; This adjusts for MHz.
   ; so for this: f(100) = (1/100) * 1,000,000 = 10000
   mov ebx, 10000
   div ebx
   ; ax contains measured speed in MHz
   mov ~[mhz], ax

See the Intel manual (see links) for more information.

- bugs report are welcome. IM to DennisCGC

Without Interrupts

This article's tone or style may not reflect the encyclopedic tone used throughout the wiki.
See Wikipedia's article on tone for suggestions.

I'd be tempted to say 'yes', though I haven't gave it a test nor heard of it elsewhere so far. Here is the trick:

disable()     // disable interrupts (if still not done)
outb(0x43,0x34);   // set PIT channel 0 to single-shot mode
outb(0x40,0);      // program the counter will be 0x10000 - n after n ticks
long stsc=CPU::readTimeStamp();
for (int i=0x1000;i>0;i--);
long etsc=CPU::readTimeStamp();
outb(0x43,0x04);   // read PIT counter command ??
byte lo=inb(0x40);
byte hi=inb(0x40);

Now, we know that

  • ticks=(0x10000 - (hi*256+lo)) periods of 1/1193180 seconds have elapsed at least and no more than ticks+1.
  • etsc-stsc clock cycles have elapsed during the same time.

Thus (etsc-stsc)*1193180 / ticks should be your CPU speed in Hz ...

As far as i can say, 0x1000 iterations lead to 10 PIT ticks on a 1GHz CPU and a bit less than 0x8000 ticks on the same CPU running BOCHS. This certainly means that on very high speed systems, the discovered speed may not be accurate at all, or worse, less than 1 tick could occur ...

This technique is currently under evaluation in [the forum|Forum:5849]

- hope you like my technique /PypeClicker

Asking the SMBios for CPU speed

Main article: SMBIOS

The SMBIOS (System Management BIOS) Specification addresses how motherboard and system vendors present management information about their products in a standard format by extending the BIOS interface on Intel architecture systems. The information is intended to allow generic instrumentation to deliver this information to management applications that use DMI, CIM or direct access, eliminating the need for error prone operations like probing system hardware for presence detection.

Do note that SMBIOS was never intended for initialization purposes. It was intended to provide information for asset management systems to quickly determine what computers contained what hardware. Unfortunately this means that it might be quite unreliable, especially on cheap/home systems. So ultimately it may not be the best way to determine CPU speed.

SMBios Processor Information

A Processor information (type 4) structure describes features of the CPU as detected by the SMBios. The exact structure is depicted in section 3.3.5 (p 39) of the standard. Within that information you will find the processor type, family, manufacturer etc. and also:

  • the External Clock (bus) frequency, which is a word at offset 0x12,
  • the Maximum CPU speed in MHz, which is a word at offset 0x14 (e.g. 0xe9 is a 233MHz processor),
  • the Current CPU speed in MHz, (word at offset 0x16).

Getting the SMBIOS Structure

SMBios provide a _Get SMBIOS Information_ function that tells you how many structures exists. You can then use _Get SMBIOS Structure_ function to read processor information.

As an alternative, you can locate the _SMBIOS Entry Point_ and then traverse manually the SMBIOS structure table, looking for type 4.

All this is depicted in 'Accessing SMBIOS Information' structure of the standard (p 11).


Related threads in the forum

Other resources

especially section 12: "Operating Frequency" on page 29 of 24161815.pdf

Regarding SMBIOS