AMD Details "SQUIP" Side Channel Vulnerability For Zen's Execution Unit Scheduler

[piotrj3]

What's the benefits of having those units in numbers bigger than 1 instead of just splitting them into separate cores? I assume you have different numbers for different types of units so that probably wouldn't really work out though.

Thank you!



Oh wow, is 8 the optimal number for most cases? Can you tweak it, whether in BIOS or on the fly, to a smaller number if it works out better?

Thanks!

Because if you split them to seperate cores, your single threaded performance will be horrible. You want to have high single thread performance (by using multiple units for certain operations so if it is possible to execute few at once you can), but you also want to have good multicore performance and increase utilization of those units when they can't be used by 1 core.

This is whole paradox - if your workload can be perfectly parallel you would make as simple core as possible and make as many cores as possible. If workload is strongly bound by single thread you want as powerful single thread as possible - crank that GHz, make tons of execution/decoder units for 1 core etc.

This is a thing that if you want power saving processor, you want low frequency many simple cores with agressive downclocking.

Best gaming CPUs are exactly opposite of that.

I hope that Intel's aproach long term will good answear for that question, as Intel 12600k/12700k can compete with 5800X/5600X and if you pick non-K version efficiency is even better. Source:

https://www.techpowerup.com/review/intel-core-i5-12600k-alder-lake-12th-gen/20.html

E cores are more simple cores that take less die space and P-cores are those for single-threaded bound tasks.

[PublicNuisance]

But how much performance did you lose in going back to a FX-9590? Did you "Before & After" benchmark your changes?

I haven't done any fresh ones but when I first bought the Ryzen 5 2600X years ago I saw most games give me at least a 20-30% jump over the FX-9590 with some even doubling. To be fair though at the time I didn't even know about stuff like the PSP and had I known I would never had bought the Ryzen. I would have instead spent my money on either a Libreboot or Coreboot solution. I'm trying to sell what parts I can now to move forwards in that direction and to also save up to get a Talos II. Gaming is my main hobby but I have to draw a line in the sand somewhere. From now on any hardware I buy has to obey some ground rules. No Intel ME or PSP that is not at least neutered in a FOSS fashion; No closed source BIOS; and compatible with the Libre kernel preferred.

[archkde]

Ah. I think suspend being affected by SMT is firmware-related. I don't think it's a bug, but I'm curious on what exactly happens with SMT on/off in the firmware.

I have an ASUS PRIME-X470 PRO motherboard, and for the SMT option the BIOS specifically mentions S3/Sleep not functioning when cores are removed. In AMD CBS there's a similar note, and that implies whatever is happening is intentional and done by AMD.

The BIOS SMT option is not relevant here, Linux disables SMT by putting the extra logical CPUs into a deep idle state (C10) and never scheduling a task on or assigning an interrupt to them.

But from what I've found out by now, and unlike I had assumed, it's actually the kernel's job to restore the hardware state on resume from suspend. Firmware puts the CPU into the boot state. So this is actually a Linux kernel bug.

[geearf]

Because if you split them to seperate cores, your single threaded performance will be horrible. You want to have high single thread performance (by using multiple units for certain operations so if it is possible to execute few at once you can), but you also want to have good multicore performance and increase utilization of those units when they can't be used by 1 core.

This is whole paradox - if your workload can be perfectly parallel you would make as simple core as possible and make as many cores as possible. If workload is strongly bound by single thread you want as powerful single thread as possible - crank that GHz, make tons of execution/decoder units for 1 core etc.

This is a thing that if you want power saving processor, you want low frequency many simple cores with agressive downclocking.

Best gaming CPUs are exactly opposite of that.

I hope that Intel's aproach long term will good answear for that question, as Intel 12600k/12700k can compete with 5800X/5600X and if you pick non-K version efficiency is even better. Source:

https://www.techpowerup.com/review/intel-core-i5-12600k-alder-lake-12th-gen/20.html

E cores are more simple cores that take less die space and P-cores are those for single-threaded bound tasks.

Thank you for this explanation!

[Alex/AT]

"constant-time algorithms" are now "best practices", oh really?
Data-independent random-time algorithms feel much more secure. You can't infer anything from a random time. The only thing we need is some 'RANDDELAY x,y' CPU command that uses software-independent entropy source to delay for some small amount of CPU cycles between x and y.