What makes a cpu "faster"

[xYowie]

I would assume clock speed is like a multiplier for how fast it works, but what kind off difference does hyper threading and cache make?

[SLAYR]

1 minute ago, xYowie said:

I would assume clock speed is like a multiplier for how fast it works, but what kind off difference does hyper threading and cache make?

There is also the IPC of the chip which is the base architecture of the chip.

 

If you look up TechQuickie (another LTT channel) they have videos explaining these topics.

[ThinkWithPortals]

5 minutes ago, xYowie said:

I would assume clock speed is like a multiplier for how fast it works, but what kind off difference does hyper threading and cache make?

IPC, frequency, amount of cores, amount of threads, amount of cache, speed of cache all play their roles.

[Spork829]

Clock speed and core count are the main largest factors. Hyper threading basically makes the OS think that one core is 2 (4 cores with hyper threading are essentially treated like 8), which helps with more multithreaded workloads. There are some other factors too like architectural improvements measured in IPC. Like SLAYR said all this stuff is on TechQuickie. 

[TidaLWaveZ]

A pretty good deal.  I don't have numbers nor do I know where you can get them besides maybe a side by side of two CPU's from the same generation with different cache sizes running in the same machines.

 

As for hyperthreading, it's definitely offers a decent percentage of cpu performance increase. 

[AresKrieger]

Frequency x ipc + software + data streamline (cache etc) = performance (this is just a basic idea not a realistic model)

 

Hyper threading can add a lot of performance if a task can use more threads (ideally a 50% boost but is usally less due to optimizations so 30-40% for multi threaded tasks)

 

As for cache it is required to give instruction to the CPU however that is more chip to chip but generally more cache is better if the higher level cache isn't excessive (since bigger cache can be slower

[TidaLWaveZ]

1 minute ago, Spork829 said:

Clock speed and core count are the main largest factors. Hyper threading basically makes the OS think that one core is 2 (4 cores with hyper threading are essentially treated like 8), which helps with more multithreaded workloads. There are some other factors too like architectural improvements measured in IPC. Like SLAYR said all this stuff is on TechQuickie. 

 

Gotta watch out for that clock speed and core count though, AMD taught us something about that in the past few years.

[Spork829]

1 minute ago, AresKrieger said:

Frequency + ipc + software + data streamline (cache etc)

Core count is huge as well.

[Spork829]

1 minute ago, TidaLWaveZ said:

 

Gotta watch out for that clock speed and core count though, AMD taught us something about that in the past few years.

Well not really. The AMD chips with higher clocks and core counts are worse than Intel chips with lower numbers there because of their way lower efficiency. Only thing is that sometimes they try and include GPU cores in that core count for marketing. 

[Spork829]

1 minute ago, TidaLWaveZ said:

 

Gotta watch out for that clock speed and core count though, AMD taught us something about that in the past few years.

Well not really. The AMD chips with higher clocks and core counts are worse than Intel chips with lower numbers there because of their way lower efficiency. Only thing is that sometimes they try and include GPU cores in that core count for marketing. 

[AresKrieger]

Just now, Spork829 said:

Core count is huge as well.

ipc is instruction be cycle (it's not per core per cycle that's a different spec from my understanding) so it's more relevant than core count, core count usually doesn't mean anything just look at fx line of cpus

 

For example if you have a single core that can do the same number of calculations as 1000 cores would the single core not be a better option

 

 

[TidaLWaveZ]

1 minute ago, Spork829 said:

Well not really. The AMD chips with higher clocks and core counts are worse than Intel chips with lower numbers there because of their way lower efficiency. Only thing is that sometimes they try and include GPU cores in that core count for marketing. 

 

That's the whole point right? The higher core count and clock speed doesn't mean the CPU performs better.

[Spork829]

2 minutes ago, AresKrieger said:

For example if you have a single core that can do the same number of calculations as 1000 cores would the single core not be a better option

Right, but realistically a 1000 core CPU would be used for an extremely multi threaded workload. So the 1 core CPU would have to be running a LOT faster and built on an architecture not even comparable to that of the 1000-core to get the same performance. Since most taxing programs can take advantage of 2, 4, or even more cores, if the architectures of the two CPUs being compared are within range of each other at all, core count is still important. Those AMD chips performed poorly in comparison because of relatively inefficient architectures and the fact that Intel had hyperthreading. 

[Spork829]

2 minutes ago, TidaLWaveZ said:

 

That's the whole point right? The higher core count and clock speed doesn't mean the CPU performs better.

Oh for sure, I didn't mean ONLY look at that. Architecture/IPC is perhaps the most important part, but once you narrow it down to one line of CPUs with one architecture, core clock and count are a good indicator of performance. 

[AresKrieger]

1 minute ago, Spork829 said:

Right, but realistically a 1000 core CPU would be used for an extremely multi threaded workload. So the 1 core CPU would have to be running a LOT faster and built on an architecture not even comparable to that of the 1000-core to get the same performance. Since most taxing programs can take advantage of 2, 4, or even more cores, if the architectures of the two CPUs being compared are within range of each other at all, core count is still important. Those AMD chips performed poorly in comparison because of relatively inefficient architectures and the fact that Intel had hyperthreading. 

What I was getting at is that core count is only a workaround to gain more ipc in workloads that can utilize it while if a single core had the same ipc and clock speed of a multi core cpu they would perform the same (assuming the program supported multiple cores)

[Spork829]

3 minutes ago, AresKrieger said:

What I was getting at is that core count is only a workaround to gain more ipc in workloads that can utilize it while if a single core had the same ipc and clock speed of a multi core cpu they would perform the same (assuming the program supported multiple cores)

Well I wouldn't call it a 'workaround' if it actually increases performance. It's much more expensive to make a single core versus 4 if you want them to perform the same. 

[AresKrieger]

5 minutes ago, Spork829 said:

Well I wouldn't call it a 'workaround' if it actually increases performance. It's much more expensive to make a single core versus 4 if you want them to perform the same. 

Yes hence why I call it a workaround, it's a cost saving measure in addition to a time saver, I'm not saying it's a bad thing or an unnecessary thing (it's both needed and good in my eyes) it's just a way to avoid the difficulties of trying to increase the performance drastically with only one core available

 

Workaround: a method for overcoming a problem or limitation in a program or system.