Coffeelake--revenge of the pentium 4

Hi I just read this
http://www.tomshardware.com/picturestory/764-best-of-computex-2017.html#s4
It says the 15 W coffeelake will be >30% faster with double the cores & an increase of 500 Mhz in turbo boost.
I confirmed it here
https://www.notebookcheck.net/Intel...15W-chips-to-be-up-to-30-faster.224716.0.html
And confirmed it here
https://www.macrumors.com/2017/05/30/intel-coffee-lake-30-pct-performance-boost/
And here
https://www.pcper.com/news/Processors/Computex-2017-Intel-8th-Gen-Core-Processors-30-Faster-7th-Gen
And here
https://arstechnica.com/gadgets/201...formance-boost-for-8th-generation-processors/.


let's compute
2×cores = 100% ideal performance increase.
500 Mhz = 14.28% performance increase.
Taking real world restrictions (but this was a benchmark) there should be a minimal 100% increase.
Q.e.d. 3 has 30% or less IPC/core vs kaby lake.

 

From your first link:

What HP the platform offers isn't what will necessarily be used by the O/S and the programs you run.

While 30% doesn't seem as flashy as your ideal 114.28% increase or more... - it is more truthful than what other manufacturers are throwing around as the performance gain from their latest processors...

 

Yes but still the performance perclock per core seems to have dropped like a rock. Like they substituted 1 A72 with 4 A53. Or a core processor with 4 atoms. Might even make sense. A still more powerful & faster core would have consumed more power and die area per core, so they might have ended up with a 50 W TDP, for coffeelake U

 

That 30% will not be across the entire product stack, for all KL vs CL. It's one test for 15W parts using a "user experience" workload. A high power unconstrained CPU doing 100% utilisation synthetic load is completely different

Manufacturers do also routinely cherrypick the best scenario for their demos.

 

I do not understand. Why would intel cherrypick a result which shows that canon lake is 30% faster when it could have shown that it is 130%+ faster? Both the sku's are running the same scenario at the same power.

 

Performance per clock has dropped, how?

I'll repeat what I've already stated: what an O/S and programs deliver to a user (i.e. 'productivity') is a far cry from what the theoretical compute power a platform has.

Intel choses to report on the performance difference a user will likely experience.

Other companies chose to boast about theoretical improvements that may or may not bring about any real, real world improvements in todays compute landscape for workstation type workloads (i.e. single user).

Always has been like this, always will be... well, until the hardware can start programming itself in the future for always optimum performance.

 

1st sysmark 2014 is not a synthetic test.
https://bapco.com/products/sysmark-2014/
SYSmark® 2014 ver 1.5 is an application-based benchmark that reflects usage patterns of business users in the areas of office productivity, media creation, and data/financial analysis. SYSmark 2014 ver 1.5 features the latest and most popular applications from each of their respective fields.


Let me put it this way
SYSmark® 2014 ver 1.5 is an application-based benchmark that reflects usage patterns of business users in the areas of office productivity, media creation, and data/financial analysis. SYSmark 2014 ver 1.5 features the latest and most popular applications from each of their respective fields.

The compared CPU is i7-7500U which is 2.7 GHz Base and 3.5 GHz Turbo as mentioned by intel. Traditionally, common applications are supposed to parallel single threaded performance.
Now let the score be X for the performance of a single core
Then for a dual core processor the performance would ideally be 2X.
Adding in the effect of turbo that would be 29.6% (considering a single core)
That would make the net performance of i7-7500U roughly 2.3X.
Then with 4 cores, the performance would be 4x.
Intel says this has a turbo of 4 GHz. Assuming they are talking about single core turbo (it could be dual core turbo) the performance comes to 4X+0.3X+0.15X = 4.45 X. Assuming a poorly threaded workload that leaves 1 core at 0% leaves 3.45X.
3.45X divided by 2.3X gives cannon lake a 50% lead. Which is greater than 30(+/-7)%.
Of course, you could posit that intel deliberately choose the benchmark and system config to show it's latest chip in a bad light.

Please note 4x A53 could outperform 1× A57. But in that case the IPC will still be lower in the A53 chip.
Which is all I am saying.

 

Nobody said Sysmark was a synthetic test (quite the opposite, in fact; re-read what has already been posted).

Your math for how powerful a multicore platform is for workstation type usage (i.e. single user) was outdated (and proven wrong) a decade ago by... real life workloads.

Thanks for your colorful reply, but I really don't need to see fonts bigger than default to get the idea you're trying to convey.

I'll try once again to point out the obvious that you're ignoring.

Intel is reporting real world performance gains - using a current O/S and current workloads - (they are not reporting the theoretical gains their platform jump achieves with a synthetic (server-like) workload like Cinebench would show).

The actual gains seen by a consumer do not depend just on the processor. The O/S, programs, workflows and workloads and other platform choices also make a big contribution too. This doesn't change the processors IPC or any other narrow parameter of choice you may focus on.

But it does change how truthfully Intel is communicating with it's customers vs. other manufacturers.


An example is you have 'x' much horsepower to have an 8 second car in the quarter mile. 2 times 'x' is needed to shave off tenths of seconds in the real world; but that doesn't negate the fact that that slightly faster car is actually twice as powerful. i.e. 2x 'x' doesn't equal 4 seconds real world results.

Btw; your poll doesn't reflect any real world applicability. None of those choices make sense.

Right now, for almost all consumers that need performance, an i7 QC (4C/8T) of the latest generation gives the best overall performance (not too many cores, not too few threads, not too slow clocks - balance). When it's paired with the fastest and highest capacity memory possible (minimum 16GB, 32GB or more recommended for any current processor/platform available).