AMD Posts New "AMD-PSTATE" CPUFreq Driver Leveraging CPPC For Better Perf-Per-Watt

[Michael]

For those wondering about CPU support... The benchmarks I am running at the moment are working out fine for 5950X / 5900HX / EPYC 7543... So far only with one 5900X box did I find the driver not binding to any device.

[numacross]

What about chiplets puts Zen CPU's at a disadvantage? As far as I can tell, chiplets allow them to more aggressively bin cores, add a bunch of cache, and increase core count in a cheap enough way that they can sell 64-core/128-thread chips for a third of the price of Intel's 28-core/56-thread chips.

The fact that Zen chiplet designs are split into processing and IO dies. They are physically separate and thus require a relatively high-power, fast and low latency interconnect over a, from the point of IC, huge distance. Monolithic designs thus have lower overall power usage, as indicated by Zen APU and laptop chips.

1) The same idle CPU power consumption can be observed while RAM is running at default 3200MHz.
2) The only thing I've done to RAM is changing its speed and tightening some timings. Its voltage remains at 1.36V. No other electrical parameters have been changed in BIOS - everything is running by default.

By the JEDEC standard that's overvoltage. Stock 3200MHz is 1.2V. While considered "normal" for the consumer market to take more, it is above the standard.

3) The CPU is running 100% stock. No overclocking of anything including SoC/Uncore. In fact only RAM is "overclocked" (I put it in quotes because I don't think raising its speed amounts to OC'ing per se, more so considering the IF speed is at 1800MHz which is the figure AMD calls "standard". 1900-2000MHz amount to overcloking).

I'm going to have to ask for the source of that. Since AMD considers anything over 3200MHz overclocking then IF over 1600MHz is overclocked as well, and together with RAM overvoltage might, theoretically, lead to excessive SoC load, especially if the BIOS defaults to SoC OC mode due to stability concerns (which disables power saving features).

So, that's it. I'm not terribly concerned with it but it's still kinda sad to see considering Core i9 11900K idles at around 6W while being built on a much inferior node (14nm). As far as I understand it's AMD's IO complex in desktop Ryzen 3000/5000 CPUs which doesn't properly power down. If you check the screenshot again, CPU cores themselves draw basically nothing.

Yeah, I even checked my laptop and it's powering down IF there correctly:


Granted that's a monolithic chip, not a chiplet design.

[avem]

Yeah, I even checked my laptop and it's powering down IF there correctly:

Nice to see, mobile Ryzens can downclock IF dynamically. On desktop it's always running at full speed.

Speaking of IF. For Ryzen 5000 CPUs 1800MHz IF clock is considering by many as "normal", 1900MHz and above is OC: https://www.pcgamer.com/amd-zen-3-in...abric-lottery/ https://www.tomshardware.com/feature...5000-ram-guide

Though you're correct in that 3200MHz RAM/1600MHz FCLK are truly standard by AMD's own definition but then again my Ryzen 7 3700X idled at a much lower wattage. I didn't change anything after I replaced it with 5800X.

[Myownfriend]

The fact that Zen chiplet designs are split into processing and IO dies. They are physically separate and thus require a relatively high-power, fast and low latency interconnect over a, from the point of IC, huge distance. Monolithic designs thus have lower overall power usage, as indicated by Zen APU and laptop chips.

Yes but that's still just one aspect of their architecture. Since chiplets allow for higher cache sizes, core counts, and more aggressive binning, cores can be more efficient and less traffic needs to travel off-chiplet and off-chip. Chiplets also allow lower performance parts of the design to be manufactured on more power or space efficient fabrication processes and allow for stacking. Both of these things are allowing AMD to make those 64MB SRAM chiplets which seem to be twice as dense as the SRAM on the core chiplet while needing to travel far less distance than a monolithic, planar chip due to stacking.

I understand that you mean that a monolithic chip would be more efficient if all else were equal but that's not the case with Zen desktop and mobile. Compare the highest-end Zen 3 laptop chip, the 5980HX, to it's closest chiplet-based desktop equivalent, the 5800X. The 5980HX has half as much L3 cache, doesn't support PCIE 4.0, uses a slightly different memory controller that supports LPDDR4 but supports less memory overall and no ECC, and has lower base clocks. The OEM-only 5800 is a variant of the 5800X that brings it's base clocks down to roughly the same as the 5800 and it's TDP is within 11 watts of the 5980HX's max TDP.

[numacross]

Nice to see, mobile Ryzens can downclock IF dynamically. On desktop it's always running at full speed.

Yes, but it is the BIOS' fault! I managed to make my 5800X downclock the IF and reduce the idle power by 10W. It took many reboots and even one recovery BIOS re-flash from USB



The downsides are quite heavy, I had to manually disable every "Auto" setting of my Gigabyte BIOS. It also required disabling SoC OC mode in two different places (one in Tweaker menu and one in the standard AGESA AMD tree). My memory sticks are unable to boot at any timings in 3200MHz at 1.2V, so I had to reduce the frequency in order to keep within JEDEC standard... even so, the timings are awful.
I haven't tested it, but I assume that reducing PCIe to 3.0 would drive the power usage even more.

[numacross]

Yes but that's still just one aspect of their architecture. Since chiplets allow for higher cache sizes, core counts, and more aggressive binning, cores can be more efficient and less traffic needs to travel off-chiplet and off-chip. Chiplets also allow lower performance parts of the design to be manufactured on more power or space efficient fabrication processes and allow for stacking. Both of these things are allowing AMD to make those 64MB SRAM chiplets which seem to be twice as dense as the SRAM on the core chiplet while needing to travel far less distance than a monolithic, planar chip due to stacking.

I understand that you mean that a monolithic chip would be more efficient if all else were equal but that's not the case with Zen desktop and mobile. Compare the highest-end Zen 3 laptop chip, the 5980HX, to it's closest chiplet-based desktop equivalent, the 5800X. The 5980HX has half as much L3 cache, doesn't support PCIE 4.0, uses a slightly different memory controller that supports LPDDR4 but supports less memory overall and no ECC, and has lower base clocks. The OEM-only 5800 is a variant of the 5800X that brings it's base clocks down to roughly the same as the 5800 and it's TDP is within 11 watts of the 5980HX's max TDP.

Hold on, you're touching many points at once. I provided an explanation why desktop parts have higher idle power usage - the Infinity Fabric is an "external" interconnect, and that still holds. No matter how well the CPU chiplets are made they still have to communicate with the IO die, and the distance is long in terms of IC design. That will take a lot of power.

As avem wrote Intel's equivalent monolithic desktop chip can provide roughly equivalent features and it manages to achieve way lower idle power usage than Zen desktop CPUs. I then theorized that in Zen desktop case it's probably a matter of BIOS settings meant to disable power saving features on the IO die in favor of increasing stability. It seems that I was, at least partially, right.

Instead of 5980HX you should be comparing desktop APU idle power usage. How do they do at idle? I wasn't able to find a HWiNFO idle screenshot, but by this benchmark it seems that the idle usage has improved by 5-8W.

[avem]

Yes, but it is the BIOS' fault! I managed to make my 5800X downclock the IF and reduce the idle power by 10W. It took many reboots and even one recovery BIOS re-flash from USB

The downsides are quite heavy, I had to manually disable every "Auto" setting of my Gigabyte BIOS. It also required disabling SoC OC mode in two different places (one in Tweaker menu and one in the standard AGESA AMD tree). My memory sticks are unable to boot at any timings in 3200MHz at 1.2V, so I had to reduce the frequency in order to keep within JEDEC standard... even so, the timings are awful.
I haven't tested it, but I assume that reducing PCIe to 3.0 would drive the power usage even more.

That looks like too much of a sacrifice but thanks for the effort anyway. The range of IF clocks looks pretty narrow though from 857 to 1067Hz which is a far cry from what mobile Ryzens can accomplish. Looks like AMD didn't create desktop Ryzen CPUs with good idle power consumption in mind or maybe it didn't allow to meet the performance targets. Ryzen CPUs love fast RAM and running it at 2166 MHz is really not a good idea.

[numacross]

That looks like too much of a sacrifice but thanks for the effort anyway. The range of IF clocks looks pretty narrow though from 857 to 1067Hz which is a far cry from what mobile Ryzens can accomplish. Looks like AMD didn't create desktop Ryzen CPUs with good idle power consumption in mind or maybe it didn't allow to meet the performance targets. Ryzen CPUs love fast RAM and running it at 2166 MHz is really not a good idea.

With a bit more tuning it would probably be possible to optimize, but I won't do that since my normal setup is performance-first. My GPU idles at 25W so the Ryzen is in good company

[leogx9r]

avem Desktop definitely downclocks IF at idle, assuming C-states are enabled. This is from AGESA 1.2.0.2 (with about 10 tabs opened in my browser atm) and you can see IF drops to ~450 MHz from the default 1800 MHz (3600 MT/s memory).



Due to the IOD, there's basically always at least 5 W being used for desktop systems (naturally that's based on your SoC voltage so if you can get that < 0.9V, you're gonna have slightly better idle power draw), lowest I've ever managed to see was around 11 W in a pure TTY environment.

[numacross]

avem Desktop definitely downclocks IF at idle, assuming C-states are enabled. This is from AGESA 1.2.0.2 (with about 10 tabs opened in my browser atm) and you can see IF drops to ~450 MHz from the default 1800 MHz (3600 MT/s memory).

Unfortunately that's BIOS dependent. My Gigabyte B550 board defaults to SoC OC mode, which disables power management and keeps IF at max speed, regardless of C-states, since those are working fine.