How to compare CPUs (for example Intel vs Apple)?

[brajesh]

TL;DR Summary

CPUs used to be easy to compare and now it seems much more complicated. What are some simple thumbrules to gauge CPU power?

Hi,

Back in the day, I remember my PC going from a x286 to a x386, to a x486, then pentium.

It was fairly easy to compare the changes, it was easy to notice the clock speed that was increasing, from 33Mhz for the x386 to the 300Mhz for the pentium.

In the recent years, we seem to be reaching limits of clocks speeds, around 3Ghz-5Ghz and I'm no longer able to easily compare CPUs.

There is so much more to consider, number of cores, architecture, gpus, on board ram, gaming frame rates, power usage, who knows what else.
I'm seeing all kinds of measurements being used by different web sites and reviews.

My question is, what's a ballpark way to compare modern CPUs?
This doesn't have to be perfect answer, but just to get an idea of the power of the processor.

Can I use gigaflops/teraflops (floating point operations)?
Or millions of instructions per second (MIPS)?

What are a few simple benchmarks I could use to compare different CPUs?

 

[anorlunda]

Is single-threaded speed your only criterion for comparison?

 

[jedishrfu]

There are several benchmark tests that can be run on a system to give you an idea of the speed of processing. The tests are a mix of cpu intensive, and disk io intensive tasks.

As an example, the X86 was superceded by the X286 in IBM's early offerings. The literature said the CPU change would provide a 6x increase in processing. However testing revealed that we only got a 3x boost. Investigation revealed that the disk I/O was doubled from the PCXT(had X86 cpu) to the PCAT(had X286 cpu) and so from that we realized the true speedup we got.

Nowadays, things have more complex architectures that affect the speeds and so benchmark tests help determine the likely speed up based on the type of work you plan to do.

This article discuss some of the best benchmark tests to consider:

https://geekflare.com/pc-benchmark-software/

 

[DrJohn]

You could try googling cpu benchmark ...

 

[brajesh]

Is single-threaded speed your only criterion for comparison?

No, how can I compare any two CPUs?

 

[phinds]

No, how can I compare any two CPUs?

You could try googling cpu benchmark ...

 

[Algr]

When determining the speed of two CPUs, you have to start by deciding what tasks you'll be doing more often, and other environment. For example, if you will be running a laptop on battery power, Apple will crush anything running windows. But if you are fine with a PC drawing 1000 watts out of an electrical outlet, then a PC with a beefy graphics card will win out.

 

[Rive]

My question is, what's a ballpark way to compare modern CPUs?

You can do that for specific systems and tasks only. No (fairly) accurate, easy general measure for CPUs anymore (if there was any, ever: I recall doing SPEC CPU95 once, and that was already pretty mixed test, with very mixed results on different computers).

 

[Drakkith]

You could try googling cpu benchmark ...

Seconded. Most of the sites I've used to compare CPU's have a single 'performance' number that is convenient to use to quickly gauge different CPU performance. Personally I also look at the single-thread rating as well since I'd rather have a slightly better single-thread performance vs slightly better multi-thread.

And if you REALLY want to know details, these sits often have a score for individual tests like number of operations per second for integer math, floating point math, string sorting, etc.

 

[kyphysics]

Is the new Intel Raptor Lake chip supposed to be better than Zen 4 of AMD? It comes out in a few days, right?

 

[kyphysics]

There are several benchmark tests that can be run on a system to give you an idea of the speed of processing. The tests are a mix of cpu intensive, and disk io intensive tasks.

Maybe a bit off-topic, but didn't Intel for years fake a lot of their benchmarks (and later get caught for it)?

Another question is how trustworthy are reviews with benchmarks. If it comes from Intel, should one trust it?

And if it comes from an "outside source," how can you know they're not being sponsored by Intel for that review?

 

[berkeman]

Maybe a bit off-topic, but didn't Intel for years fake a lot of their benchmarks (and later get caught for it)?

Source?

 

[kyphysics]

Source?

See from 15 min. 40 sec. onward.

 

[jedishrfu]

Looks like they played some clever marketing to steal attention away from competitors:

https://www.digitaltrends.com/computing/intel-28-core-cpu-fake-amd/

One time at my former company, we were doing an in-house tester demo. Upper-level mgmt came to see it. It was a real demo and would take several of minutes to fire up and compile the test program, load bit patterns into the hardware, and then run the test.

An upper-level manager after the demo tells us it's okay to speed through stuff (fake the demo) as its a demo.

I guess he wanted us to do a "Julia Child" demo where you start cooking and then tell the audience it will take 20 minutes to complete so you remove the pot from the stove and pull up another pot from under the table fully cooked and go on with the show.

 

[Rive]

Maybe a bit off-topic, but didn't Intel for years fake a lot of their benchmarks (and later get caught for it)?

Well, it's difficult to pick a describing term, but ... kind of.
Old, but persistent story.

Another question is how trustworthy are reviews with benchmarks. If it comes from Intel, should one trust it?

If you care, then you should just read the review of reviews

Honestly, it's better if you just stop caring. Unless you are some cutting edge performance enthusiast of some kind, you just won't really meet these any way. Better care for the possible future extensions and such than for these.
The best is try to find some user reviews and recommendations from users with similar computer usage as you.

 

[anorlunda]

Once upon a time computers came with flashing lights and spinning tape drives. They were sufficient for demos.

 

[jedishrfu]

I never understood why Hollywood went the route of spinning tape drives when the actual start/stop tape motion was so much cooler.

 

[Vanadium 50]

Ever see an optical paper tape reader? Whoosh!

 

[jedishrfu]

I used to think the pneumatic card readers used on large mainframe computers sounded like Star Trek's photon torpedoes.

 

[kyphysics]

Once upon a time computers came with flashing lights and spinning tape drives. They were sufficient for demos.

How small do you think they'll get in the future? ...Say in 25 years?

I'm already impressed we have Apple Watches.

 

[kyphysics]

^^^Maybe this is for another thread, but I'm also wondering if there is a natural limit that we currently know of for how small and powerful a computer can get? Like what is the smallest/most powerful computer that we can currently conceive of creating?

 

[jedishrfu]

There is an obvious limit issue here.

To make things smaller means to make thinner on-chip traces(wires) and smaller transistor components. In doing so, we limit the amount of power they can consume before they overheat which limits the processing speed.

We could change chip technology, but then we'd need to rethink everything from signaling issues to power to size to speed again. As an example, GaAs is faster than silicon technologies, but the signaling pulses are notoriously difficult to get right at high-clock speeds. (from my recollection of working on a GaAs project running the chips at 250 GHz)

https://en.wikipedia.org/wiki/Gallium_arsenide

Anyway, if we believe in the power of Moore's law, then we can expect a doubling of chip density every two years which would be 4096 times improvement ($2^12$) in 25 years. This can be viewed in a couple of ways but mostly as smaller chips with more function and less power but shorter traces.

 

[kyphysics]

Anyway, if we believe in the power of Moore's law, then we can expect a doubling of chip density every two years which would be 4096 times improvement ($2^{12}$) in 25 years. This can be viewed in a couple of ways but mostly as smaller chips with more function and less power but shorter traces.

Are you saying they'd also overheat faster?

 

[jedishrfu]

No, you have to balance the heat generation to insure that doesn't happen but in doing so you have to reduce the speed. That's why its so hard to determine how they will play out in the next 25 years. Traces(wires) that connect the transistors have to get thinner for higher density chips and can only carry so much current before they overload, pop and break like a wire fuse so you must balance the speed against the limits of chip size.

 

[anorlunda]

Don't forgret that not all processors need speed. Small and low power processors are very useful in sensors, and remote devices. It can be easier to use a flexible CPU than to design a custom circuit, even for the most trivial uses.

 

[Vanadium 50]

not all processors need speed

Your toaster doesn't toast any faster with a faster controller. Your elevator doesn't go faster. Clothes don't dry faster. Food doesn't bake faster.

I suspect a BCM2835 SOC is under a dollar in quantity. I also suspect the minimum quantity is in the hundreds of thousands.

 

[anorlunda]

I suspect a BCM2835 SOC is under a dollar in quantity. I also suspect the minimum quantity is in the hundreds of thousands.

Right idea, but the numbers are slightly different.

 

[Vanadium 50]

Hmmm...wonder why anyone would ever buy 140 of them? Cheaper to get a thousand

 

[jtbell]

Ever see an optical paper tape reader? Whoosh!

When I was an undergrad, I played with a PDP-5 that used a paper tape reader on the side of a Teletype terminal. Not quite as fast.

 

[jtbell]

I'm already impressed we have Apple Watches.

I'm still waiting for the Dick Tracy wrist TV.

Apple could make one basically by shrinking an iPhone down to Apple Watch size...

 

[Vanadium 50]

What is a picture "tube"? And an "aerial"?

 

[dlgoff]

When I was an undergrad, I played with a PDP-5 that used a paper tape reader on the side of a Teletype terminal. Not quite as fast.

I remember those. I played with the PDP-8.

 

[DrJohn]

How small do you think they'll get in the future? ...Say in 25 years?

I'm already impressed we have Apple Watches.

That tiny tiny screen just does not impress me one little bit. A tiny computer that can tell you the time, or let you read this forum two words at a time? Come on! They sell them to the terminally niave. IMHO

 

[Algr]

That tiny tiny screen just does not impress me one little bit. A tiny computer that can tell you the time, or let you read this forum two words at a time? Come on! They sell them to the terminally niave. IMHO

But what if it is so tiny that it fits in a contact lens? Or over your optic nerve? Then the screen will be as big as you want.

 

[DrJohn]

But what if it is so tiny that it fits in a contact lens? Or over your optic nerve? Then the screen will be as big as you want.

We will have to wait a bit for that to happen I suppose

 

[kyphysics]

That tiny tiny screen just does not impress me one little bit. A tiny computer that can tell you the time, or let you read this forum two words at a time? Come on! They sell them to the terminally niave. IMHO

My physical therapist says she uses her mainly to contact back and forth with her kids. Yeah, you can use a cell phone, so I'm not sure what she means by that.

Maybe it has features that make such an interaction easier? Mostly, I agree that the screen is too small for anything useful online.

 

[Algr]

I don't have one, but I suppose it could be good for notifications. Do you want to talk to this person who is calling/texting? Do you have an appointment? Is your pulse doing something strange?

 

[ETERNITY Starship]

Fritz Chess Benchmark is simple and easy use ... and has option to select single or multicore etc

http://www.jens-hartmann.at/Fritzmarks/

Fritzchessbench

*I write and run lots of benchmarks ... since around 2010 GHz speed increases have been incremental.

This means that other links in the computational chain have a far bigger impact than they used to.

Temperature ... a good CPU cooler and case fans can increase compute speeds by 5% or more

Cooler temps mean less thermal throttling General architecture has yeilded about a 1% or 2% increase with each iteration ... newer is better

True Cores are approx 5% to 10% faster than virtual cores (Hyper threading etc)

IPC ... instructions per clock is still a main factor to look for

Medium core counts CPU's are generally faster than high core counts at least for most typical applications ... Cinebench etc can utilise high core counts l... most software written only uses 4 cores or so ... high core count CPU's generally have lower clock speeds and lower / shorter Turbo Boost Speeds

GHZ is still king (mostly) ... But Cache (L1,L2,L3) and 'interconnects' are the new kings on the block

 

[Vanadium 50]

I'm not sure I agree. A chess benchmark might be useful for benchmarking chess programs, and it might be OK for workloads that involve searching through trees, but as a one-size-fits-all benchmark? Nope.

As far as the relationship between cores and threads, I can shown you workloads where additional threads adds nothing, where it adds a full core's worth of power, and anywhere in between. All depends on the workflow. I can even show you workflows where adding a software thread without the hardware to back it up helped a little bit. That sounds impossible, but the numbers are what they are.