Hacker News Reader

Comment by LeicaLatte | Hacker News Reader

"The problem here is the same problem we’ve seen with big core parts and Windows 10, however: these large processors can only take so much before having to move threads around, to keep both the frequency high and the energy density low. All it takes is for a minor internal OS blip and the single-thread power begins to dissipate. Windows 10 famously kicks in a few unwanted instruction streams when you are not looking, and as a result the average single thread performance seen on the 8700K might be equal or lower than that of the previous generation. It becomes an infuriating problem to debug as a reviewer."

I didn't know this was a big issue with multicore processors, very interesting. Does this mean that if you OC a single core to the limit to get the best single threaded performance, you will see drops in performance when your other cores start getting utilized?

View on HN

I can't quite tell from the current replies if the mechanism of what is going on is clear. Apologies if this is redundant.

What is happening is a thermal management issue. "turbo" lets you boost briefly to speeds that will cause the chip to overheat if run long enough. Their idea to manage this was to play hot potato with which core gets baked. As soon as it reaches critical temperatures, they turbo up another core and the OS migrates threads.

So one part of the chip gets hot while the other cools off, and then they continue back and forth.Personally I'm surprised this doesn't crack silicon or pop bonds.

With regards to overclocking, I guess I didn't know you could overclock a single core? Is this pinned or is it adjusting what the "turbo" frequency is?

In anycase, if there is a differential in speeds among cores, the hot potato thing will probably happen. And if you overclock it will probably happen more.

The only way to combat this would be to either mess with the windows cpu driver and put your cpu at risk, or make the speed homogenous among all cores.

> mess with the windows cpu driver and put your cpu at risk

AFAIK, the thermal throttling is hardware controlled. I doubt you can override the protections in software. Driver tweaking may help

> "turbo" lets you boost briefly to speeds that will cause the chip to overheat if run long enough. Their idea to manage this was to play hot potato with which core gets baked. As soon as it reaches critical temperatures, they turbo up another core and the OS migrates threads.

This is not an accurate description of how turbo works. You don't have to migrate away from a core and switch to another.

Turbo simply recognizes that if you have a multi-core processor, and many of the cores stay asleep and don't contribute to power and thermal demands, then the remaining cores can take up the slack. The more cores stay asleep, the faster the remainder can go.

Does this mean that if you OC a single core to the limit to get the best single threaded performance, you will see drops in performance when your other cores start getting utilized?

This is how all CPUs with some form of "Boost" work, yes. People who overclock generally lock the multiplier to avoid it. So you will tend to have this problem when you do not overclock.

To further complicate things, some of the larger many cored parts have a favored core which will turbo higher than the others, but only if the OS schedules work on it.

Yes.

At least for their high-end server parts, Intel has docs on how many turbo bins you get. AVX can also slow things down all on its own (it reduces the available turbo). A co-worker of mine is (hoping to) speak at KVM forum about this as relates virtualization.

Ah, found some docs for the Core i5: https://www.intel.com/content/www/us/en/support/processors/0...

There's a nicer version of these somewhere with graphs, but I'm on my phone and it's early :)

It is called thread migration (or, more abstractly, process migration) and high memory clock speeds favor environments where this is expected to occur frequently.

As such, there is significant performance degradation when high-powered CPUs are paired with low-powered RAM. Incidentally, most OEMs/ODMs are susceptible to this as they frequently use high-powered CPUs as advertisement anchors but cut costs with low-powered RAM.

It is the source of unpleasant lag spikes for end-users and, as mentioned elsewhere, they are difficult to work around.

When you say low-powered ram, are you talking about clock speeds?

Reply on news.ycombinator.com