Hacker News Reader

Comment by jowdones | Hacker News Reader

Core count is becoming more and more bullshit. If I get performance X using 1 core and Y using all cores, then true core count is Y / X, regardless of what advertised CPU count the manufacturer may dream.

Geekbench scores: -----------------

Intel Core i7-8086K: 6288 single core, 29927 multicore, asvertised cores: 6, real cores: 4.75, bullshit factor (advertised/real): 125%. Decent.

AMD Ryzen Threadripper 3970X: 5838 single core, 68457 multicore, advertised cores: 32, real cores: 11.7, bullshit factor: 270%. Ridiculous.

View on HN

I wouldn't say bullshit. Diminishing returns should be expected, even in embarrassingly parallel workloads, because of Amdahl's law as well as physical limitations regarding power delivery and cooling when utilizing all the cores.

That said, I like your "real cores" metric. Looking at $/"real core" would be very informative. But even then, there's still folks for whom the time saved would be worth the price premium.

Also, I'm not sure about Geekbench's MC score when run on TR. They've had problems in the past.

If I run your bullshit factor analysis with the i7-8086K vs the Ryzen 9 3950X using the single-core and 64-core workload tests from User Benchmark (no 3rd gen Threadrippers there yet) both get a bullshit factor of about 155%. Even the Threadripper 2990WX only has a bs factor of about 192%.

This just means Geekbench isn't good at scaling to multiple cores. It could be the workload or the implementation. If you have the right workload and a good implementation, modern processors do scale reasonably well (all numbers with Threadripper 3970X):

Cinebench R15: 206 single, 7302 multi, 35.4x scaling

POV-Ray: 9:02 single, 0:19 multi, 28.5x scaling

The scaling can be greater than the number of physical cores due to SMT. On the other hand, turbo clocks increases the single-core speed by 10% or more on most CPUs.

Or they demonstrate that the CPU is throttling cores due to excessive heat. Compare the 24-core 3960x to the 32-core 3970x, and also look at the power consumed when running cores at 100% utilization. It would appear that the 3970x starts throttling the cores once ~21-22 cores are in full utilization, leading the 3970x to be only marginally faster than the 3960x, even though it has 50% more cores. EDIT: 3970x has a GeekBench score of 70655 (single-threaded: 5684) while 3960x has a score of 66676 (single-threaded: 5703; according to https://www.anandtech.com/show/15044/the-amd-ryzen-threadrip...). Meanwhile, power consumption per core starts dropping off significantly at 21 cores for both (https://www.anandtech.com/show/15044/the-amd-ryzen-threadrip...)

> Meanwhile, power consumption per core starts dropping off significantly at 21 cores for both

Which you would expect with so many cores. Power consumption per core is probably going to be even lower for the 3990X, especially considering it has the same TDP.

So your "true core count" for the entire Summit supercomputer will be one if you run Dijkstra's SSSP on it?

What is the bullshit factor for your "true core count"?

You have an extremely warped understanding of what is actually going on and how things work. Non linear utilization of more cores is a matter of software and thermals.

Reply on news.ycombinator.com