Can you convert clock speed into bandwidth?

[ixelion]

Does it make sense to convert measurment of Mhz into MB/s?

For example a 3.0 Ghz processor can process a certain amount of data within a single clock cycle, can this be measured in terms of MB/s or GB/s?

I am afraid my knowledge is extremely limited, I do not even know how these measurements even reflect performance.

 

[PsYcHoCoW]

It can be done when speaking about memory and bus speeds... DDR400 memory [400MHz] is 3200Mbps (not sure about the units, would have to check..)

as for bus speeds and bandwidths, they are mostly documented. (PCI, AGP, PCI-Express are all "buses")

You take the bus' frequency and multiply by the number of bits (or bytes) which are transferred on every clock (1 clock = 1 transfer usually, now if there are 400 million transfers a second of 64 bits, this makes 25,600 millions of bits transferred a second... or 3200megabytes a second)





As for the processor's work, I don't think it could be transferred into megabytes per second because it's not regular/linear enough (too complex). Each operation done by the processor is 32-bit, but a single operation goes through several steps so I don't know how it could be counted.

 

[OdiN]

Originally posted by: PsYcHoCoW
It can be done when speaking about memory and bus speeds... DDR400 memory [400MHz] is 3200Mbps (not sure about the units, would have to check..)

as for bus speeds and bandwidths, they are mostly documented. (PCI, AGP, PCI-Express are all "buses")

You take the bus' frequency and multiply by the number of bits (or bytes) which are transferred on every clock (1 clock = 1 transfer usually, now if there are 400 million transfers a second of 64 bits, this makes 25,600 millions of bits transferred a second... or 3200megabytes a second)

Uhm....25,600 Mb/s != 3.2GB/s

 

[PsYcHoCoW]

Originally posted by: OdiN
Uhm....25,600 Mb/s != 3.2GB/s

Thanks for correcting me. (oh wait)

I typed this out without too much attention, I'll be glad if it's the only mistake that's found.

 

[PsYcHoCoW]

Actually I'm trying to find what kind of mistake I made in these numbers...

http://www.anandtech.com/printarticle.aspx?i=2223

PC3200:
3200 MB/s * 8 bits = 25600 Mb/s

 

[Continuity28]

Originally posted by: OdiN

Originally posted by: PsYcHoCoW
It can be done when speaking about memory and bus speeds... DDR400 memory [400MHz] is 3200Mbps (not sure about the units, would have to check..)

as for bus speeds and bandwidths, they are mostly documented. (PCI, AGP, PCI-Express are all "buses")

You take the bus' frequency and multiply by the number of bits (or bytes) which are transferred on every clock (1 clock = 1 transfer usually, now if there are 400 million transfers a second of 64 bits, this makes 25,600 millions of bits transferred a second... or 3200megabytes a second)

Uhm....25,600 Mb/s != 3.2GB/s

It depends on what we're talking about of course.

Bits are not always 8 to the byte, some use 9 bits, some use 10 bits, 12, etc. Some use 8 data bits with a bit or two for other purposes, making it really 10 bits per byte but only 8 are really countable.

At any rate, a processors frequency cannot be directly compared to data rates.

 

[OdiN]

Originally posted by: PsYcHoCoW
Actually I'm trying to find what kind of mistake I made in these numbers...

http://www.anandtech.com/printarticle.aspx?i=2223

PC3200:
3200 MB/s * 8 bits = 25600 Mb/s

...must have miss typed a number in the old calc.exe

 

[PsYcHoCoW]

Originally posted by: Continuity28
It depends on what we're talking about of course.

Bits are not always 8 to the byte, some use 9 bits, some use 10 bits, 12, etc. Some use 8 data bits with a bit or two for other purposes, making it really 10 bits per byte but only 8 are really countable.

At any rate, a processors frequency cannot be directly compared to data rates.

Thanks Continuity28 for a worthwhile informative forum post. OdiN, double check in the future, and explain in your post instead of just saying that is wrong. :heart:

 

[ixelion]

You take the bus' frequency and multiply by the number of bits (or bytes) which are transferred on every clock (1 clock = 1 transfer usually, now if there are 400 million transfers a second of 64 bits, this makes 25,600 millions of bits transferred a second... or 3200megabytes a second)

Can you not do the same with core clock speed? I.e. can you take the core clock frequency and multiply it by the number of bits that are transfered to every clock and then find out how many bits transfered in a second?

 

[Continuity28]

Originally posted by: ixelion

You take the bus' frequency and multiply by the number of bits (or bytes) which are transferred on every clock (1 clock = 1 transfer usually, now if there are 400 million transfers a second of 64 bits, this makes 25,600 millions of bits transferred a second... or 3200megabytes a second)

Can you not do the same with core clock speed? I.e. can you take the core clock frequency and multiply it by the number of bits that are transfered to every clock and then find out how many bits transfered in a second?

Well keep in mind, when talking about memory, you're going to be dealing with writing data anyways, in that case the frequency can be converted into a data rate.

With a processor it's different, as its doing computational adds, subtracts, multiplies etc. Not necessarily writing data to anywhere but its own cache. Many other factors such as the length of the pipeline have an influence on thouroughput as well, even if you were only using the processor to write the same data, differing architectures with the same frequency would write that data at different speeds.

In other words, even if you could make a direct comparison, it would not be true all the time because what your processor is used for is so varied. It makes making a comparison almost pointless in this case.

 

[ixelion]

That makes a lot of sense, I also did not consider that FSB, although messured in frequency is still a bus that a certain amount of data passes through at one point like 64 bit or 32 bit, but a processor is not a bus and data does not pass through in the same manner i dont think.

It seems to me mhz seems to mean less and less, so I was wondering if there was some other means of measuring a CPU's performance.

 

[Stas]

I was wondering if there was some other means of measuring a CPU's performance.
flops

 

[Stas]

Floating Point Operations Per Second

 

[Stas]

check this one out, too: top500.org You might find it interesting.

 

[biostud]

nm

 

[biostud]

Originally posted by: ixelion

It seems to me mhz seems to mean less and less, so I was wondering if there was some other means of measuring a CPU's performance.

The reason why benchmark programs exists.