One ongoing question reviewers have been digging into for the past few weeks is the expected behavior of AMD 7nm Ryzen CPUs at high boost clock versus the actual, measured behavior. AMD promised to update the user community today, September 10, as to the expected behavior of its CPUs and what changes would be incorporated in upcoming UEFI revisions.
To briefly recap: Reports in late July showed that some AMD CPUs were only reaching top boost clock frequency on a single CPU core. Last week, overclocker Der8aurer reported the results of a user survey showing that only some AMD 7nm Ryzen CPUs were hitting their full boost clocks (the exact percentage varies by CPU model). Late last week, Paul Alcorn of Tom’s Hardware published an extensive test of how different AMD AGESA versions and UEFI releases from motherboard impacted motherboard clocking. AGESA is the AMD Generic Encapsulated Software Architecture — the procedure library used to initialize the CPU and various components. Motherboard vendors use the AGESA as a template for creating UEFI versions.
What THG found was that different UEFI versions and AGESA releases have shown subtly different clocking results. Later releases have hit slightly lower boost clocks compared with the earlier versions that were used for reviews. At the same time, however, these later versions have also frequently held their boost clocks for longer before down-throttling the CPU.
There’s also evidence that the throttle temperatures have been subtly adjusted, from 80C initially down to 75 before creeping back upwards to 77. These changes would not necessarily impact performance — the CPU is boosting a bit lower, but also boosting longer — but it wasn’t clear what, exactly, AMD was trying to accomplish. During its IFA presentation last week, Intel argued that these subtle variations were evidence that AMD was trying to deal with a potentially significant reliability issue with its processors. THG was unwilling to sign on to that explanation without additional information.
While all of this was unfolding, AMD notified us that it would make an announcement on September 10 concerning a new AGESA update.
The text that follows is directly from AMD and concerns the improvements that will be baked into updated UEFIs from various motherboard manufacturers. I normally don’t quote from a blog post this extensively, but I think it’s important to present the exact text of what AMD is saying.
ur analysis indicates that the processor boost algorithm was affected by an issue that could cause target frequencies to be lower than expected. This has been resolved. We’ve also been exploring other opportunities to optimize performance, which can further enhance the frequency. These changes are now being implemented in flashable BIOSes from our motherboard partners. Across the stack of 3rd Gen Ryzen Processors, our internal testing shows that these changes can add approximately 25-50MHz to the current boost frequencies under various workloads.
AMD Reference Motherboard (AGESA 1003ABBA beta BIOS)
AMD Wraith Prism and Noctua NH-D15S coolers
Windows 10 May 2019 Update
22°C ambient test lab
Streacom BC1 Open Benchtable
AMD Chipset Driver 1.8.19.xxx
AMD Ryzen Balanced power plan
BIOS defaults (except memory OC)
These improvements will be available in flashable BIOSes starting in about two to three weeks’ time, depending on the testing and implementation schedule of your motherboard manufacturer.
Going forward, it’s important to understand how our boost technology operates. Our processors perform intelligent real-time analysis of the CPU temperature, motherboard voltage regulator current (amps), socket power (watts), loaded cores, and workload intensity to maximize performance from millisecond to millisecond. Ensuring your system has adequate thermal paste; reliable system cooling; the latest motherboard BIOS; reliable BIOS settings/configuration; the latest AMD chipset driver; and the latest operating system can enhance your experience.
Following the installation of the latest BIOS update, a consumer running a bursty, single threaded application on a PC with the latest software updates and adequate voltage and thermal headroom should see the maximum boost frequency of their processor. PCMark 10 is a good proxy for a user to test the maximum boost frequency of the processor in their system. It is expected that if users run a workload like Cinebench, which runs for an extended period of time, the operating frequencies may be less than the maximum throughout the run.
In addition, we do want to address recent questions about reliability. We perform extensive engineering analysis to develop reliability models and to model the lifetime of our processors before entering mass production. While AGESA 1003AB contained changes to improve system stability and performance for users, changes were not made for product longevity reasons. We do not expect that the improvements that have been made in boost frequency for AGESA 1003ABBA will have any impact on the lifetime of your Ryzen processor. (Emphasis added).
Separately from this, AMD also gave information on firmware changes implemented in AGESA 1003ABBA that are intended to reduce the CPU’s operating voltage by filtering out voltage/frequency boost requests from lightweight applications. The 1003ABBA AGESA now contains an activity filter designed to disregard “intermittent OS and application background noise.” This should lower the CPU’s voltage down to 1.2v as opposed to the higher peaks that have been reported.
Finally, AMD will release a new monitoring SDK that will allow anyone to build a monitoring tool for measuring various facets of Ryzen CPU performance. There will be more than 30 API calls exposed in the new application, including:
Current operating temperature: Reports the average temperature of the CPU cores over a short sample period. By design, this metric filters transient spikes that can skew temperature reporting.
Peak Core(s) Voltage (PCV): Reports the Voltage Identification (VID) requested by the CPU package of the motherboard voltage regulators. This voltage is set to service the needs of the cores under active load but isn’t necessarily the final voltage experienced by all of the CPU cores.
Average Core Voltage (ACV): Reports the average voltages experienced by all processor cores over a short sample period, factoring in active power management, sleep states, VDROOP, and idle time.
EDC (A), TDC (A), PPT (W): The current and power limits for your motherboard VRMs and processor socket.
Peak Speed: The maximum frequency of the fastest core during the sample period.
Effective Frequency: The frequency of the processor cores after factoring in time spent in sleep states (e.g. cc6 core sleep or pc6 package sleep). Example: One processor core is running at 4GHz while awake, but in cc6 core sleep for 50% of the sample period. The effective frequency of this core would be 2GHz. This value can give you a feel for how often the cores are using aggressive power management capabilities that aren’t immediately obvious (e.g. clock or voltage changes).
Various voltages and clocks, including: SoC voltage, DRAM voltage, fabric clock, memory clock, etc.
Ryzen Master has already been updated to give average core voltage values. AMD expects motherboard manufacturers to begin releasing new UEFIs with the 1003ABBA AGESA version incorporated within two weeks. As we wrote last week and despite rumors from Asus employee Shamino, AMD is not portraying these adjustments to clocking behavior as being related to reliability in any way.
As for AMD’s statements about the improved clocks, I want to wait and see how these changes impact behavior on our own test CPUs before drawing any conclusions. I will say that I don’t expect to see overall performance change much — 25-50MHz is only a 0.5 to 1 percent improvement on a 4.2GHz CPU, and we may not even be able to detect a performance shift in a standard benchmark from such a clock change. But we can monitor clock speeds directly and will report back on the impact of these changes.
TomsHardware tested a leaked beta version of a fixed BIOS, and it does improve performance measurably, though I'm not sure it'd be noticeable, granted they were using a processor which was only boosting 25mhz lower than advertised. Is possible the bug AMD is talking about is related to the CPPC, since the difference between enabled and disabled with boosting behavior is quite a lot.
I ran a userbenchmark on my last official UEFI 2703, and a modded UEFI that contained the updated SMU code 46.49.00.
The results are shown below and then tabulated.
The new SMU microcode has a bigger impact the few cores are in use. I saw a 4% or more increase in score from 4 cores to 1. 8 cores was below 2%, while 12 cores was within margin of error.
So while the new boosting code does impact performance, it didn't have any impact on the all-core boost for me.
This pretty much had the effect that rhallock noted in the blog post over in the gaming section of this forum.
What is really incredible is that this small performance gain moved me from the 74th percentile in Userbenchmark to the 99th percentile. It is impressive how similar the performance of all test 3900Xs really are.