Relative IPC and efficiency of some Intel CPUs in Cinebench R15

[mackerel]

Consider this pre-testing to a wider IPC test. I ran Cinebench R15 in various configurations on 3 CPUs representing 3 different generations.

• 12100F - Alder Lake, Intel 7 (formerly 10 Enhanced SuperFin)
• 11700k - Rocket Lake, 14nm
• 7980XE - Skylake-X, 14nm

Rocket Lake was the first desktop generation past Skylake architecture even though it was still made on 14nm process, and Alder Lake finally took desktop beyond 14nm nodes.

Why use Cinebench R15? Mainly because it is well known, and as older software it does not make use of AVX instructions so takes away an element of complication. It is known to not be much affected by memory performance which is good if I'm focusing on the CPU cores itself. I note the score, reported CPU power consumption, and where possible the CPU clock. In some scenarios with a power limit, the clock varied too much to get a good value. Skylake-X here should be near enough a substitute for regular Skylake and derivatives (Kaby Lake, Coffee Lake, Comet Lake).

Note this is "quick testing" and I didn't try to eliminate all variables and check of repeatability beyond doing a minimum of 2 runs each.

Relative performance (IPC best case):
Skylake-X: 52.2 points/core/GHz
Rocket Lake: 63.3 points/core/GHz, +21% relative to SKX
Alder Lake: 72.7 points/core/GHz, +39% relative to SKX, +15% relative to RKL

Relative efficiency:
This is where it gets really complicated. It depends on where on the efficiency curve you compare, and these are wildly differing CPU configurations. I used a metric of points/W for comparison.

7980XE unlimited (188W): 16.6 points/W
7980XE turbo off (117W): 20.9 points/W
11700k unlimited (183W): 12.7 points/W
11700k turbo off (88W): 20.4 points/W
11700k 45W limit: 22.5 points/W
11700k 25W limit: 23.1 points/W
12100F unlimited (56W): 20.3 points/W
12100F 45W limit: 23.9 points/W
12100F 25W limit: 30.2 points/W

I decided against testing the 7980XE at lower powers since it was reporting around 45W usage at idle!

We still have different core configurations. I tried adjusting the 11700k further, running it with only 4 cores enabled (4c8t), and also running it with all cores and HT off (8c8t).

11700k 4 cores 8 threads
125W limit: 10.1 points/W
45W limit: 18.3 points/W (Alder Lake 30% more efficient)
25W limit: 18.6 points/W (Alder Lake 62% more efficient)
As generally expected, this is much worse than running with 8 cores. While each core has more power available, it runs in a less efficient area. Alder Lake does seem to scale better at lower powers. Note although is is 4 core vs 4 core, by disabling 4 cores of Rocket Lake it may not scale exactly. Intel did not make a 4 core Rocket Lake.

11700k 8 cores 8 threads (HT off)
Unlimited (159W): 11.1 points/W
45W limit: 18.5 points/W
Cinebench R15 is on the higher end of HT scaling workloads. For a 45W power limit, it is doing 18% less work than with HT on. While it was not a part of my testing this time, previously I've seen typically around 30% more throughput at the same clock with HT, implying the difference here is due to the extra power consumed by HT. Arrow Lake is rumoured to not have HT so it will be interesting to see how that goes overall.


Overall we do see a bit more efficiency from the newer process, especially at lower powers. There is also a clear improvement in IPC between the generations. Of course, this is very limited in only looking at Cinebench R15, and I hope to expand this to wider workloads over time.

 

[MaddMutt]

WOW Thanks for this ^

 

[Evilsizer]

im going to say this is my personal preference. if we are going to compare, then it should be equal cores and equal clock speeds. I am not sure how we call it a increase when one cpu has higher cpu speed vs another even if the same number of cores/threads. its good info for sure!

 

[mackerel]

im going to say this is my personal preference. if we are going to compare, then it should be equal cores and equal clock speeds. I am not sure how we call it a increase when one cpu has higher cpu speed vs another even if the same number of cores/threads. its good info for sure!

This is a special case for Cinebench R15 only. It doesn't scale much at all with cache or ram, like a fraction of a % going from single channel at base speed to all the channels overclocked to infinity and beyond. Cinebench R15 you can closely predict the score of pretty much any CPU if you know the clock and cores/threads, assuming you previously worked out the appropriate scaling factors for the architecture used. That scaling factor is indicative of IPC for this specific software. Maybe it falls apart if you run it on a 256 core monster or similar, but that's way outside my reach. The power measurements I'm really not sure what to do with for now. Ideally I'd need a set of curves for each architecture and software. Unless I can automate that testing, it will be insane to manually run.

This would NOT apply to Cinebench 2024 for example. They have finally made it tough enough it shows ram limiting on faster CPUs making it much better at system level testing, if introducing a headache for testing since it takes forever to run with slow CPUs. When using software to test hardware, I also test the software to see how it responds. Take gaming testing for example, there is a balance between GPU and CPU. I basically have to consider similar for CPU testing, since it is a balance between cores and ram. Ok, cache impacts it too, but you get the idea. I need to reiterate at this point, this is not product or system testing, this is microarchitecture testing.

For other workloads, it isn't so much about keeping the cores and clock equal, as making sure that the cores and clocks are kept low enough not to be externally limited. As long as the software is known to scale well in the test conditions used, it doesn't have to be identical.