AMD's Ryzen 3000 Boost Fix Works, But Often Hits Wrong Cores

We've discovered quite a few encouraging traits with AMD's new boost-fixing BIOS and the Ryzen 9 3900X processor that extend beyond 'just' the expected heightened frequencies, but we also found a weakness in the chips' inability to hit the rated maximum boost speeds on all cores (the chips come with a mix of faster and slower cores). That often results in our Ryzen 9 3900X boosting on inactive cores, thus providing no performance improvement in some workloads.

AMD's Ryzen 3000 series debuted to wide acclaim with up to 15% higher instruction per cycle (IPC) throughput borne of the Zen 2 microarchitecture paired with the 7nm process, leading to a new level of value for the mainstream desktop. It also had the most cores we've seen on a mainstream platform, with more to come in the form of the 16-core 32-thread Ryzen 9 3950X later this month.

Unfortunately AMD has been criticized due to some chips' inability to hit the rated boost frequencies, which eventually led the company to issue the new firmware we're testing today. AMD says the new firmware improves clock rates by 25 to 50MHz, along with other fine-grained improvements. After several preliminary tests with beta firmware, we've found that AMD's updates correct the small clock rate deficit that we've seen with a few of our single-die Ryzen 3000 models, like the Ryzen 7 3700X.

While the new firmware delivers smaller boosts than expected, especially in light of the up to 300 MHz shortfall reported with the Ryzen 9 3900X in a recent survey, AMD does say that it corrects a firmware bug that reduces performance in some scenarios. With our testing of the single-die models complete, we're now focusing on the improvements for the multi-die Ryzen 9 3900X, which seems to be most impacted by the frequency deficit. We'll also test some adjustments AMD made to idle power consumption and see how the new update impacts AMD's auto-overclocking Precision Boost Overdrive (PBO) feature, but most importantly, examine how the Windows scheduler is still causing AMD problems.

Ryzen 3000 Boost Clock Fix: The Long View

For those of you interested in the testing particulars, you'll find them at the bottom of the page. For the rest of you, here's the quick rundown of the production-quality Gigabyte BIOSes we're testing today.

The F7A BIOS is the most interesting, as it contains the fixes for the boost clock issues, while the F5P BIOS is Gigabyte's last pre-patch BIOS. Finally, the N11 BIOS, which was provided to reviewers at launch, serves as our baseline.

We begin by recording the frequencies of each core during a series of commonly-used tests that should expose the peak frequencies. The first three tests include LAME, POV-RAY, and Cinebench in single-core test mode. These programs only execute on one core of the processor, which typically allow the chip to reach its peak boost frequency within its power, current, and thermal envelope.  

AMD's announcement of its BIOS fix noted that Cinebench, a program that enthusiasts often use to measure boost frequencies, isn't the best program to measure peak frequency due to the extended nature of the workload, instead suggesting that tests with intermittent "bursty" workloads, like PCMark 10, are better suited for testing. We added PCMark 10, Geekbench, and VRMark to factor in those burstier workloads.

Image 1 of 3

With 12 cores hammering away, the per-core frequency recordings create almost unintelligible charts, so the album above only includes the maximum and minimum frequencies recorded during each 1-second measurement interval. That means these measurements could come from any one core, but it makes the charts easier to digest. We've also plotted chip temperature on the right axis (the dark red line), and we'll break these charts up into zoomed versions for detailed observations during key sections of the test.

The obvious top-line conclusion is that our Ryzen 9 3900X now boosts up to 4.65 GHz, which is 50 MHz above the rated 4.6 GHz. That's impressive, as our peak measurements with previous BIOSes, including the 'reviewer' version, typically topped out at 4.575 GHz, meaning we've picked up ~75 MHz of extra boost speed from a simple BIOS update.

As we'll soon see, that is often a hollow victory.

We also noticed that these boosts typically come when the processor is under 53C, but that's easy to misinterpret. We'll take a closer look at that behavior, and what it really means, shortly.

Zooming In: PCMark 10, GeekBench 4, VRMark Test Results

We're zooming in on the latter portion of the benchmarks first because the PCMark 10 results are especially interesting. The approximate location of each benchmark in our chart is detailed in the table below. We also include the performance results in the album. 

Image 1 of 12

PCMark 10 spans from ~2,200 to ~4,200 seconds (bottom axis). The new firmware has much more boost activity during the benchmark than we see with older BIOSes, both in the number of boosts and the peak frequencies. We notice that the clock spikes largely correlate with lower temperatures, which is easy to wrongly interpret as boost activity only occurring during idle periods of the test. However, PCMark 10 is not a single-threaded test, instead using multiple tests that include multi-core workloads that cause the temperature to rise and the chip to shift into lower frequencies (just like Intel's chips).

That means many of these lower temperature readings aren't actually during idle periods. Instead they are just single- or lightly-threaded workloads. In either case, the chip only gives us a single temperature reading that doesn't include separate measurements for each of the two 7nm compute dies, making the metric unsuitable for drawing conclusions (localized heat is the main factor that impacts boosting behavior).

It does send us down the right path of investigation though, which ends up fingering Ryzen's inability to boost on all cores as the culprit in our odd performance results for this test.

Remember, as we discovered after the launch, AMD's Ryzen 3000 processors come with a mix of both faster and slower cores, with the slower cores unable to reach the rated boost specification. To help mitigate the issue, AMD uses the new Windows 10 scheduler to target workloads at the fastest cores first.

For our Ryzen 9 3900X sample, it appears that only cores 1, 2, 3, and 7 can reach 4.6 GHz or beyond. Contrary to what the temperature readings suggest, many of the boosts do happen during active portions of the test, but only when the workload is running on a single core, which naturally contributes to the lower temperature. Unfortunately, after analyzing the test output in detail, we see that, in many of these cases, the processor simply boosts on the wrong core.

For instance, a single core may be at 100% utilization, but that core may not be able to reach the peak boost frequency. That doesn't stop the processor from boosting an inactive core, though, which is exactly what happens. Because the core isn't doing any work, the boost is wasted and doesn't equate to faster performance.

This implies the scheduler isn't directing these bursty workloads to the fastest cores at all times, and because not all cores can reach the single-core boost speed, the boost often doesn't have an impact on performance. There are a few rare instances during the PCMark 10 benchmark where the workload lands in a core that can hit the 4.6 GHz boost, but the boosts tend to trigger on inactive cores more often than not. We also see this same behavior in the GeekBench tests.

In either case, after further scrutiny, this behavior also exists on previous versions of the Gigabyte and MSI BIOSes with the 3900X, so it isn't likely that AMD has nefariously tuned the BIOS to provide a perceivable, but hollow boost. This highlights the disadvantage of relying on the still-imprecise Windows scheduler to direct workloads at the fastest cores. This uneven boosting results in variance in our suite of PCMark 10 test results at the end of the album. The new BIOS is faster in aggregate, but it is really down to luck if the workload lands on the correct cores during the test. As a result, many of these figures land within the standard 3% of expected variance listed by PCMark.

You can clearly see the GeekBench test run three times, with three increases in temperature where the test runs in multi-core model, near the end of the chart. The new BIOS is clearly faster in single-threaded GeekBench at stock settings, but we see some variance with the overclocked (PBO) configuration. We also see a notable decrease in multi-threaded Geekbench with the new BIOS. We also run VRMark at the tail end of the sequence, and the chip boosts much more frequently than it did with older BIOSes. However, because the chip often boosts on inactive cores, the results are inconclusive and largely fall within the expected variance.

We'll see a few examples below where the scheduler is much better at pinning the workload in the 3900X's fastest cores.

Zooming In: LAME, Pov-Ray and Cinebench Test Results

Image 1 of 6

Our LAME encoder workload, which is inherently single-threaded, runs through its paces five times during the first 100 seconds of the test. This test finds the Ryzen 9 3900X boosting less frequently with the new firmware, but achieving a higher boost speed. The test results show very little variation between the three BIOS revisions, with the fixed BIOS falling within the expected variance compared to the original reviewer BIOS, but still being ever-so-slightly faster than the previous F5P revision. We also include test results with the auto-overclocked PBO (Precision Boost Overdrive) configuration, and here we can see the new BIOS take a tangible lead over the older versions.

The POV-Ray portion of the plot finds the 3900X with the new BIOS again boosting less frequently, and all three BIOSes peak at roughly 4.35 GHz during the benchmark. This isn't unexpected, as POV-Ray uses densely-packed AVX instructions that result in lower overall clock rates, just like we see with Intel's processors during AVX workloads. In either case, the performance variation between the three BIOSes is negligible at stock settings, though we see a performance improvement with PBO once again.

The Cinebench segment of the test has a few interesting caveats. Again, we see the processor boosting to its peak clocks less frequently with the new BIOS, but these boosts are largely confined to two cores as the workload migrates back and forth. In contrast, the older BIOS revisions boosted more frequently on all cores, even ones that weren't active.

This more efficient boosting behavior results in a less-impressive chart, but the peaks are higher during the test, reaching 4.65 GHz, than with the older BIOSes. Much of that is enabled merely because the other cores aren't unnecessarily boosting as frequently, which allows the chip to conserve its thermal budget for more effective boosting. We also notice that the Windows 10 scheduler is much more effective at sending the Cinebench workload to cores with the highest boost potential, which results in much higher single-threaded Cinebench scores, too.

Five Jillion Data Points

Image 1 of 3

Flipping through these charts, we can see the eyesore that comes from plotting the frequency of every core on a single chart. But as a general observation, we can see that fewer cores are boosting with the new BIOS during the first ~2250 seconds of the test run, leading to a 'spikier' plot.

That's actually desirable, as these single-core loads don't require multiple cores to boost because the workload is resident on only core. This improvement comes from AMD's new 'activity filter' that prevents lightly-active or inactive cores from boosting when they aren't in use. This prevents the cores from boosting due to intermittent OS and application 'background noise' that's likely metered either by a utilization or power consumption monitor for each core. Unfortunately, this doesn't seem to entirely filter out the intermittent maximum boosts on totally inactive cores.

In either case, the filter reduces the temperature of the chip during lightly-threaded workloads, which we can see in the temperature plot. That means the chip consumes less power and generates less heat, too, which opens up further opportunities for boost activity.

Ryzen 3900X Idle Power Consumption, Voltage, and Clocks

Image 1 of 10

Many Ryzen 3000 owners complained of high voltages at idle, which AMD says it also addressed with its activity filter. When the activity filter is active on a core, AMD says it will drop to ~1.2V but can still boost to the full 1.5V (the rated safe voltage) on demand.

We begin by measuring package power, which is the full power consumption of the chip at the AM4 socket, and see that the new BIOS reduces overall power consumption by roughly 3 watts compared to the older versions. A quick glance at the clock rates during idle shows us that the chip remains at a steadier clock with the new BIOS during our five-minute measurement window.

Finally, we can see that the new 1.0.0.3ABBA firmware doesn't hit the peak of 1.5V at idle like the older firmwares, instead peaking at only ~1.4V. Those peaks are also less frequent due to the filter preventing unnecessary boosts.

Ryzen 3900X Loaded Power Consumption and Voltage

Image 1 of 8

The activity filter also helps with power consumption under load, as it minimizes unnecessary boost activity. Here we can see the new BIOS (in red) has much lower package power consumption throughout the heavily threaded portions of the long series of tests we used for measuring boost clocks. The improvements are much more pronounced during the POV-Ray and Cinebench tests, but extend throughout all workloads.

Flipping through to the voltage charts, we can see that the chip often ran at ~1.5V during the test with older firmwares, but the new BIOS reduces that upper threshold to 1.4V. These solid improvements in power consumption are a hidden benefit of the update that will result in less heat being dumped into your case, and more thermal headroom for more frequent boosts.

Thoughts

As with all Ryzen 3000 processors, we can't tell you how the new BIOS will impact your specific chips' boost frequency. Much of that boils down to the quality of your individual chip and the BIOS/motherboard combination. Overall, AMD's fix has corrected the clock deficits we recorded with our Ryzen 9 3900X sample, but the increased peak frequencies often don't equate to noticeable differences in performance.

That's due to the reality that AMD's chips come with a mix of faster and slower cores. Because your chip can't boost to its top frequency on every core, any wayward workload that lands in slower cores will suffer. More often than not, we observed the chip boosting on an inactive core during some workloads, which does nothing for performance. This is a persistent trend with the 3900X with both current and older BIOS revisions on the Gigabyte and MSI motherboards.

Like we've already seen in the mobile space, this new binning strategy is innovative and promises to wring the utmost performance out of every piece of silicon that comes out of the fab. But as we've seen today, AMD's implementation is still in the midst of teething pains. The first big hurdle is to get the Windows scheduler to utilize the faster cores more efficiently, and that may already be in the works.

The next Windows 10 insider build has another scheduler alteration that rotates workloads more efficiently across favored cores (ie, faster cores) to improve performance and reliability, but whether that just represents a broader, or more refined, implementation of the changes already made for Ryzen processors remains to be seen. We hope it helps to keeps workloads targeted at the faster cores, as the current scheduler seems to be playing a bit of roulette with the 3900X's cores.

Many of the alterations likely hide behind the black box of the system management unit (SMU). This unit alters the parameters of the chip in ways that can't be seen by the end user, making it difficult to track the changes being made. Our most recent testing suggests that AMD has altered the thermal thresholds to correct the boosting behavior, but we can't test that with the Ryzen 9 3900X due to its single temperature measurement for both compute dies.

In either case, the current 46.49.0 SMU revision goes a long way to improving performance, but we're also told by our sources that a new 46.49.2 revision is already filtering out to motherboard vendors. That means we should see more improvements on the BIOS front quickly. That makes sense given the pending release of the 16-core 32-thread Ryzen 9 3950X later this month.

AMD has also dismissed the reliability concerns that we recently investigated. The company says unequivocally that the heightened boost speeds will not impact processor longevity.

Overall, AMD's boost fix should meet the demands of most users that want to see their chips hit the numbers printed on the box, but you shouldn't expect miracles: Those increased boost clocks don't equate to much extra performance. We hope the company continues to work on optimizing the silicon and working with Microsoft to fine-tune the scheduler. Given AMD's history with the first-gen Ryzen processors, which were born into a world with zero optimization for the new Zen architecture but quickly served up massive post-launch gains, it's possible that we can see even more performance gains through optimization in the future.   

We have to keep things in perspective, though. We're analyzing the finer details of the architecture to see how AMD can improve these chips in the future, but our recommendations remain unchanged. We stand by the recommendations we've made in both our reviews and our Best CPU articles: The Ryzen 3000 series processors bring a new class of performance, and value, to the mainstream desktop. 

Test Setup

Currently only Gigabyte and Asus have launched final versions of the new 1.0.0.3ABBA firmware that corrects the boost behavior, while ASRock and MSI firmwares remain in beta form. We're testing three of Giagbyte's BIOSes, each with a different form of the AGESA (AMD Encapsulated Software Architecture), which is the AMD code that it provides to motherboard vendors to create motherboard firmwares. Each BIOS version also has its own unique SMU (System Management Unit) code, which is a small unit inside the processor that can alter several settings inside the chip, such as parameters that alter the boost algorithm's behavior.

The F7A BIOS is the most interesting, as it contains the fixes for the boost clock issues, while the F5P BIOS is Gigabyte's last pre-patch BIOS. Finally, the N11 BIOS, which was provided to reviewers at launch, serves as our baseline.

We topped the Ryzen 9 3900X with a beefy Corsair H115i cooler and fed it with a 1500W SilverStone power supply to assure the chip has plenty of cooling and power to achieve the maximum boost clocks. We're following the general best practices we detail here, but be aware that these results will vary based on the quality of your chip and the type of motherboard you use. We're also using the latest AMD chipset drivers that have the latest version of the Collaborative Power and Performance Control 2 (CPPC2) software that manipulates Ryzen 3000's power states from within the operating system, reducing power state transition latency from 30ms to 1ms. This software is critical to expose peak boost frequencies in bursty workloads.