SkatterBencher #13: AMD Ryzen 9 3900XT Overclocked to 4550 MHz
We overclock the AMD Ryzen 9 3900XT processor up to 4550 MHz with the ASUS ROG Crosshair VIII Impact motherboard and EK water cooling.
Table of Contents
Introduction
The Ryzen 9 3900XT is the top part in the new Matisse XT product line.
Matisse XT is the refresh of the original Matisse architecture launched last year. The new processors have the same architecture and 7nm process node as its predecessors. AMD claims it has made significant progress on yields and that’s why faster products can come to market. The Matisse XT features high clock speeds, enhanced overclocking support, and more tightly tuned boost algorithms.
The Ryzen 9 3900 XT offers 12 cores and 24 threads with a base frequency of 3.8GHz and boost frequency of 4.7GHz. The TDP is 105 watt and the MSRP just below $500. The CPUs should be available from July 7.
In this video we will cover the basic overclocking steps to get your CPU all the way to 4.55 GHz using custom loop water cooling.
We will dig into three overclocking scenarios.
First, we’ll overclock the CPU to its maximum stable Prime 95 settings. Second, we’ll push the CPU further to its all-core stable maximum frequency. Lastly, we look into individual CCX overclocking.
Before we get going, we need to talk a little bit about the overclocking constraints and our platform.
AMD Ryzen 9 3900XT: Platform Overview
Along with the AMD Ryzen 9 3900X processor, in this guide we will be using the ASUS ROG Crosshair VIII Impact motherboard, an ROG Strix RTX 2080 Ti, a pair of G.SKILL Trident Z Royal DDR4-3200 memory sticks, and of course EK water cooling. All this is mounted on top of our favorite Open Benchtable.
The cost of the components should be around $3,200
- CPU: $499
- Cooling: $200 + $400
- Motherboard: $430
- Benchtable: $200
- Memory: $160
- VGA: $1300
AMD Ryzen 9 3900XT: Overclocking Constraints
Before we get started with the overclocking, let’s cover some of the constraints we will face.
A Ryzen 3000 CPU consists of a couple of parts. Each CPU has multiple chiplets. A chiplet is a die with specific functions such as CPU cores, IO hub, Memory controller, and so on. All the chiplets communicate each other via the fabric interconnect.
A Core Chiplet Die, or CCD, is one of the chips on the AMD CPU. A CCD consists of two CCXs paired together. CCX is short for Core Complex. The Core Complex consists of 4 individual cores each with their L1 and L2 cache. They also share a larger L3 cache.
The Ryzen 9 3900XT has 2 CCDs with each 2 CCXs. Each CCX has 1 out of the 4 cores disabled. Adding up all the cores that are left, 3 times 4, makes 12 cores.
AMD’s default frequency offers aggressive clock frequency of individual cores based on the temperature and power consumption headroom, as well as the core quality.
While AMD offers aggressive frequency boost for single threaded applications, it does not offer to set the single threaded boost frequency when manually overclocking. This means that if you overclock manually you will lose the single thread frequency advantage.
Also, AMD also does not offer an AVX offset that can reduce the operating frequency when for example running Prime 95.
Lastly, by default the fabric, memory controller and memory frequency operate in “synchronous mode”. That means typically the CPU will run all frequencies in 1 to 1 ratio. Synchronous mode works up to 1800 MHz, after which the system switches to asynchronous mode. In asynchronous mode, the memory controller will operate at half the frequency of the system memory. The fabric clock will also run below system memory frequency. So, you will have a performance penalty. The penalty can be overcome by increasing the memory frequency to well over DDR4-4000 speeds.
With all this in mind, let’s jump into the benchmarks and overclocking.
Benchmarking Software
Here’s a list of the benchmarks and software utilities used in this guide, including download links:
- SuperPI 4M https://www.techpowerup.com/download/super-pi/
- Geekbench 5 https://www.geekbench.com/
- HWBOT X265 https://hwbot.org/benchmark/hwbot_x265_benchmark_-_4k/
- Cinebench R23 https://www.maxon.net/en/cinebench
- V-Ray 5 https://www.chaosgroup.com/vray/benchmark
- ROG RealBench V2.56 https://rog.asus.com/rog-pro/realbench-v2-leaderboard/
- Prime 95 Small FFT with AVX https://www.mersenne.org/download/
AMD Ryzen 9 3900XT: Stock Performance
Before we get started with pushing the performance of the AMD Ryzen 9 3900XT processor, let’s first take a look at the scoring at stock settings:
- SuperPI 4M: 44.485
- Geekbench 5 (single): 1,306 points
- Geekbench 5 (multi): 10,866 points
- HWBOT X264 4K: 20.513 fps
- Cinebench R20: 7,200 marks
- ROG Realbench: 210,790
- Final Fantasy XIV: 149.47 fps
Step 1: Manual CPU Overclocking
As a first step, we will overclock the CPU frequency to 4350 MHz with 1.365V and a level 2 loadline. We leave the other settings untouched.
Upon entering the BIOS, navigate to the Extreme Tweaker menu.
- Set Ai Overclock Tuner to XMP II
- Set AVX Instruction Core Ratio Negative Offset to 2
- Set CPU Core Ratio to Sync All Cores
- Set ALL-Core Ratio Limit to 52
- Disable Ring Down Bin
- Set Min. and Max CPU Cache Ratio to 45
- Set CPU Core/Cache Voltage to Adaptive Mode
- Set Additional Turbo Mode CPU Core Voltage to 1.45v
- Set CPU VCCIO Voltage to 1.24v
- Set CPU System Agent Voltage to 1.24v
We re-ran the benchmarks and checked the performance increase compared to default operation.
- SuperPI 4M: -1.91%
- Geekbench 5 (single): -2.53%
- Geekbench 5 (multi): 1.55%
- HWBOT X264 4K: 3.70%
- Cinebench R20: 6.49%
- ROG Realbench: : +3.56%
- Final Fantasy XIV: -0.29%
We can already see that by manually overclocking we are giving up the strong single thread boost frequency. In some benchmarks we lose performance compared to default, while in the multithreaded applications we gain performance
Let’s dial in the other frequencies.
Step 2: Manual CPU, Fabric, Memory overclocking
In addition to overclocking the CPU frequency to 4.35 GHz, we also overclock the fabric and memory controller to 1.8 GHz. We also manually increase the memory frequency to DDR4-3600 and set the memory timings.
This is also the highest Prime 95 small FFT with AVX stable configuration.
Upon entering the BIOS, navigate to the Extreme Tweaker menu.
- Set Ai Overclock Tuner to XMP II
- Set AVX Instruction Core Ratio Negative Offset to 2
- Set CPU Core Ratio to Sync All Cores
- Set ALL-Core Ratio Limit to 52
- Disable Ring Down Bin
- Set Min. and Max CPU Cache Ratio to 45
- Set CPU Core/Cache Voltage to Adaptive Mode
- Set Additional Turbo Mode CPU Core Voltage to 1.45v
- Set CPU VCCIO Voltage to 1.24v
- Set CPU System Agent Voltage to 1.24v
We re-ran the benchmarks and checked the performance increase compared to default operation.
- SuperPI 4M: -1.91%
- Geekbench 5 (single): -0.38%
- Geekbench 5 (multi): +6.25%
- HWBOT X264 4K: +10.40%
- Cinebench R20: +6.86%
- ROG Realbench: : +13.50%
- Final Fantasy XIV: +11.50%
We can clearly see the positive impact of increasing the fabric and memory frequency. Some single threaded benchmark applications are still in deficit compared to stock, but especially multi-threaded applications are benefiting nicely.
Running Prime 95 Small FFT with AVX at 4350 MHz, we’re seeing peak CPU temperature of 85C and an peak CPU package power of 220 watt.
Let’s look at post-Prime 95 overclocking capabilities
Step 3: Post Prime 95
If we ignore Prime 95, we can further increase the CPU frequency to 4.55 GHz while maintaining the same fabric and memory clock frequencies
Upon entering the BIOS, navigate to the Extreme Tweaker menu.
- Set Ai Overclock Tuner to XMP II
- Set AVX Instruction Core Ratio Negative Offset to 2
- Set CPU Core Ratio to Sync All Cores
- Set ALL-Core Ratio Limit to 52
- Disable Ring Down Bin
- Set Min. and Max CPU Cache Ratio to 45
- Set CPU Core/Cache Voltage to Adaptive Mode
- Set Additional Turbo Mode CPU Core Voltage to 1.45v
- Set CPU VCCIO Voltage to 1.24v
- Set CPU System Agent Voltage to 1.24v
Then save the changes and reboot.
We re-ran the benchmarks and checked the performance increase compared to default operation.
- SuperPI 4M: +2.71%
- Geekbench 5 (single): +3.91%
- Geekbench 5 (multi): +9.13%
- HWBOT X264 4K: +15.20%
- Cinebench R20: +11.99%
- ROG Realbench: : +14.34%
- Final Fantasy XIV: +11.98%
As expected, the performance continues to rise.
Interesting to note is that at 4.55 GHz even the single threaded workloads like SuperPI and Geekbench 5 are higher than stock. This may be confusing since the single threaded boost frequency is 4.7 GHz.
However, this can easily be explained. The 4.7 GHz is only a “best case” scenario which requires both great cooling but also great cores. Not all of the 12 cores in the 3900XT are of equal quality and we noticed that some would boost to only 4.5 GHz.
If you want to find out which core boosts to what frequency, I highly recommend you to use HWinfo’s “effective clock” measurement. You can find more information about this measurement from Elmorlabs blog article: https://www.elmorlabs.com/index.php/2019-11-15/a-look-at-the-amd-ryzen-9-3950x/
So, in short, the likely reason why we see higher performance at 4.55 GHz is (1) increased fabric and memory performance and (2) fixed near-turbo frequency for all cores.
Step 4: Per CCX overclocking
The last step in our overclocking adventure is to increase the frequency of the individual CCX. For this Ryzen 9 3900XT we found that CCX 2 on CCD 1 was able to run at half a ratio higher than the others. So, 4.6 GHz.
Upon entering the BIOS, navigate to the Extreme Tweaker menu.
- Set Ai Overclock Tuner to XMP II
- Set AVX Instruction Core Ratio Negative Offset to 2
- Set CPU Core Ratio to Sync All Cores
- Set ALL-Core Ratio Limit to 52
- Disable Ring Down Bin
- Set Min. and Max CPU Cache Ratio to 45
- Set CPU Core/Cache Voltage to Adaptive Mode
- Set Additional Turbo Mode CPU Core Voltage to 1.45v
- Set CPU VCCIO Voltage to 1.24v
- Set CPU System Agent Voltage to 1.24v
We re-ran the benchmarks and checked the performance increase compared to default operation.
- SuperPI 4M: +2.46%
- Geekbench 5 (single): +5.51%
- Geekbench 5 (multi): +9.81%
- HWBOT X264 4K: +15.22%
- Cinebench R20: +13.06%
- ROG Realbench: : +16.54%
- Final Fantasy XIV: +12.82%
Conclusion
Alright, let’s wrap this up.
When going through the numbers and discussing them with industry friends, it dawned on me that rather than having a variety of overclocking strategies, AMD users have a variety of overclocking trade-offs. But not in a bad way.
Frankly, the out of the box frequencies and resulting performance are excellent. The AMD engineers who were tasked with getting users the best possible performance at default settings did an amazing job. In fact, they did such a good job that manual overclocking can give you worse performance in certain scenarios. Specifically, single threaded light workloads.
When manually overclocking, you lose the benefits of automatic boost frequency. Also, you can not configure the boost frequencies by specific use case (i.e. by core usage or per core). This is the first overclocking trade-off: settle for lower single threaded performance with higher all core performance, or the other way around.
Another overclocking trade off is that there’s no way to configure the system for truly worst-case scenarios such as Prime 95 small FTT with AVX. On other platforms, you can use an AVX offset ratio to temporarily reduce the performance if such workloads come your way. But on AMD you can’t. That means you have to decide whether you’re willing to trade additional performance for a potentially less stable system in certain situations.
Side-note on this is that workloads like Cinebench R20, ROG Realbench, and Geekbench 5 actually use AVX instructions. So it’s not entirely correct to call for an AVX offset ratio as that would also lower the performance in scenarios where light AVX is actually stable.
In an ideal scenario, here’s how this Ryzen 9 3900XT would likely be able to overclock to:
- Single threaded workloads: 4.8 GHz (up from 4.7 GHz)
- All-core non-AVX workloads: 4.55 GHz (up from 4.5 GHz)
- All-core AVX high workloads: 4.35 GHz (up from 3.8 GHz)
As for now, if you stick to an absolute worst-case scenario you’ll have to settle for 450 MHz lower in single threaded workloads.
That’s all for today. If you have any questions or comments, feel free to drop them in the comment section below. If you liked the video, you know what to do.
See ya next time!
