Running The AMD "ABBA" Ryzen 3000 Boost Fix Under Linux With 140 Tests

[tweak42]

Great news. With this "fix", and the spectre mitigations Michael mentioned, I'm curious if there will be further optimizations down the line to that could bump performance even more.

Regarding Ryzen owners complaining, I think it's less missing performance but more on AMD false advertising in comparison to Intel. When Intel says cpu clocks at x.xx Mhz it can reach that speed. Users should not stand for AMD marketing to lie about cpu clockspeeds, regardless of the actual software performance.

[GrayShade]

Great news. With this "fix", and the spectre mitigations Michael mentioned, I'm curious if there will be further optimizations down the line to that could bump performance even more.

Regarding Ryzen owners complaining, I think it's less missing performance but more on AMD false advertising in comparison to Intel. When Intel says cpu clocks at x.xx Mhz it can reach that speed. Users should not stand for AMD marketing to lie about cpu clockspeeds, regardless of the actual software performance.

AMD run their processors closer to the silicon limit than Intel, so there's less headroom for boost frequencies. There's also no guarantee that you'll be hitting that turbo frequency. There's a good analysis on this here: https://www.anandtech.com/show/14873...uency-metrics-.

[ms178]

Great news. With this "fix", and the spectre mitigations Michael mentioned, I'm curious if there will be further optimizations down the line to that could bump performance even more.

There is still work for AMD to do to squeeze the last bit of performance out of their CPUs, such as this task here: https://sourceware.org/bugzilla/show_bug.cgi?id=24979

In short, the current GLIBC doesn't use a fast path on AMD CPUs even though at least the recent architectures support all the neccessary instructions.

[shmerl]

. When Intel says cpu clocks at x.xx Mhz it can reach that speed. Users should not stand for AMD marketing to lie about cpu clockspeeds, regardless of the actual software performance.

Users should first learn, that Intel boost and AMD boost mean totally different things. Basically, to reach maximum advertised clocks, you also need that CPPC governor support.

See: https://www.anandtech.com/show/14873...uency-metrics-

[edwaleni]

Interesting results on the Windows side shows that the scheduler is assigning tasks to cores are either not at boost or can't be boosted some of the time.

As an example, the BIOS update can boost Core 6, but Windows places the thread at Core 9 which cannot be boosted.

This means certain applications cannot get access to the core with the highest boost at the very time it needs it.

This is causing inconsistencies in certain test results.

The testers were blaming the Windows Scheduler, though I can't see how a scheduler would know if a core can't be physically boosted to its peak frequency unless it does some sort of training ahead of time.

[AndyChow]

Regarding Ryzen owners complaining, I think it's less missing performance but more on AMD false advertising in comparison to Intel. When Intel says cpu clocks at x.xx Mhz it can reach that speed. Users should not stand for AMD marketing to lie about cpu clockspeeds, regardless of the actual software performance.

I hate marketing lies. In this case also. 25-100 Mhz difference on a boost clock that tests the very fabric limit of hardware doesn't seem to me that big. Intel lies about it's TDP, marketing wise, by quite a lot.

The worst AMD lie to me was their Fury liquid-cooled GPU. Before launch, they said it would be an over-clockers "dream". Turns out you couldn't over-clock it, not one bit.

I agree that forward, all tech companies should clearly advertise some average expectation and understanding of silicon lottery. As we go forward, I bet more and more silicon will have their top performance be determined by variable limit of the physics of our production capacities. I can imagine a world where you buy a chip, and it has 256 cores. But really it's 240-280 cores, depending on how many solar flares happened during the chip lithography. Yeah, chip production might reach that level in the next 20 years. Or maybe it will be 65536 cores +/- 100.

The marketing needs to be more accurate. But +/- 100 Mhz on a 16 core 4.6 Ghz silicon really isn't that outrageous.

[numacross]

The testers were blaming the Windows Scheduler, though I can't see how a scheduler would know if a core can't be physically boosted to its peak frequency unless it does some sort of training ahead of time.

The AMD chipset drivers are providing this information from the ACPI CPPC2 tables and Windows 10 1903 has a scheduler that's using this information to be aware of which cores are the best. It's also keeping threads within the CCX to minimize InfinityFabric latencies and prefers non-SMT cores first just like on Intel.

You can see the core binning allocation in AMD's Ryzen Master software.

[numacross]

Real world isn't a game. The issue from AMD's perspective is that it could result in a lawsuit which AMD might not win and would have to compensate some of the customers who bought Ryzen 3000 CPUs and were never able to reach the declared boost frequency in a single-threaded scenario. Note that the number of users who never succeeded to reach the declared maximum boost is very large: about 80% of Ryzen 3000 users. If a CPU is marketed as having a boost frequency and is never able to reach that frequency, it should be considered false advertising.

See also: https://www.anandtech.com/show/14804/amd-settlement

I wouldn't be so hasty in believing the "80%" number. It's based on an unofficial survey by Der8auer for which some are questioning the methodology. There was no verification mechanism so the results could've been poisoned. What is more it might be a case of a silent majority - when everything is working fine there would be no reason to investigate the boost clock issue so the chances of sending a "good" report is lower than sending a "bad" report.

There is also the issue of motherboard manufacturers providing very weird defaults. For example one of the high-end ASRock X570 motherboards is defaulting to not enabling Cool'n'quiet when set to "Auto" which results in ACPI CPPC tables not being populated and the CPU not boosting properly.

I've built a few Ryzen systems and haven't had problems with boost clocks on any of them, but I took the time to understand how Zen 2 works and what's required for boosting.

Don't get me wrong, I'm not an AMD zealot and I know they've done some questionable things in the past even with regards to Zen 2 - the infamous PBO video is a good example. The reality is that Zen 2 is so finely tuned out of the box that PBO/AutoOC will usually do nothing but decrease performance. This might a BIOS/AGESA teething issue, but even manual overclocking results in sacrificing single threaded performance for multi-threaded increases.

All in all the issue is a complex one with many factors influencing the end result. The fact that AMD supports a very large range of motherboards for Ryzen 3000 doesn't help.

As we can see from this article the CPU did indeed hit and exceed the max boost clock with the new AGESA version. The performance has also increased which suggests this was not just a fast and dirty fix by AMD meant to avoid a lawsuit by faking max frequency.

[eva2000]

Thanks Michael just the type of testing I was looking at in terms of cpu frequency etc. Under Linux/CentOS/Ubuntu what is the best tool(s) to use to monitor cpu frequency (outside of PTS suite) for AMD Ryzen for idle and load right now ? I ask as I am about to do some comparative testing for Intel Xeon E-2276G vs AMD Ryzen 3600 vs AMD Ryzen 3800X and interested in tracking cpu frequency more accurately The Ryzens are using ASRockRack X470D4U with latest P3.20 bios

example on CentOS 7 for

E-2276G I get current, min max frequency of

cat /sys/devices/system/cpu/cpu0/cpufreq/{cpuinfo_cur_freq,cpuinfo_min_freq,cpuinfo_max_freq} | awk '{printf "%.2f\n", $1/1000}' | xargs
802.49 800.00 4900.00

Ryzen 3600

cat /sys/devices/system/cpu/cpu0/cpufreq/{cpuinfo_cur_freq,cpuinfo_min_freq,cpuinfo_max_freq} | awk '{printf "%.2f\n", $1/1000}' | xargs
2200.00 2200.00 3600.00

Ryzen 3800X

cat /sys/devices/system/cpu/cpu0/cpufreq/{cpuinfo_cur_freq,cpuinfo_min_freq,cpuinfo_max_freq} | awk '{printf "%.2f\n", $1/1000}' | xargs
2200.00 2200.00 3900.00

Ryzen 3800X

cpupower frequency-info
analyzing CPU 0:
  driver: acpi-cpufreq
  CPUs which run at the same hardware frequency: 0
  CPUs which need to have their frequency coordinated by software: 0
  maximum transition latency:  Cannot determine or is not supported.
  hardware limits: 2.20 GHz - 3.90 GHz
  available frequency steps:  3.90 GHz, 2.80 GHz, 2.20 GHz
  available cpufreq governors: conservative userspace powersave ondemand performance
  current policy: frequency should be within 2.20 GHz and 3.90 GHz.
                  The governor "conservative" may decide which speed to use
                  within this range.
  current CPU frequency: 2.20 GHz (asserted by call to hardware)
  boost state support:
    Supported: yes
    Active: yes
    Boost States: 0
    Total States: 3
    Pstate-P0:  3900MHz
    Pstate-P1:  2800MHz
    Pstate-P2:  2200MHz

[starshipeleven]

Well, for many this was the first time in years that they even considered AMD and immediately got burned. It was only like 50% cheaper and 200% more powerful. Pretty disappointing.

Also somewhere around 30 and 60% less hot