ASRock B650E Taichi motherboard – breaker of several records

Ľubomír Samák

1 year ago

Simulations

The AMD B650E chipset is a compromise solution to some extent, but the ASRock Taichi motherboard that is based on it makes an ultimate impression. And it’s not just a “feel”, it really is that… The VRM of the CPU didn’t fit in our thermal image with the standard procedure. There are a few quirks and things that you might find it worth tweaking, but those are usually related to other things, like the more modest chipset features.

Combining an expensive, high-end board with an AMD B650E chipset may seem odd at first, but when you realize that the connectivity of the higher-end X670E chipset (i.e. two B650E chips) is beyond what you can grasp, it starts to make sense. Often you may not even hit the limits of the B650E chipset’s capabilities.


Parameters		ASRock B650E Taichi

Socket		AMD AM5
Chipset		AMD B650E
Format		E-ATX (305 × 267 mm)
CPU power delivery		27-phase
Supported memory (and max. frequency)		DDR5 (6600 MHz)
Slots PCIe ×16 (+ PCIe ×1)		2× (+ 0×)
Centre of socket to first PCIe ×16 slot		85 mm
Centre of socket to first DIMM slot		56 mm
Storage connectors		4× SATA III, 3× M.2 (1× PCIe 5.0 ×4: 30–110 mm + 2× M.2 PCIe 4.0 ×4: 80 mm)
PWM connectors for fans or AIO pump		8×
Internal USB ports		1× 3.2 gen. 2×2 type C, 2× 3.2 gen. 1 type A, 4× 2.0 type A
Other internal connectors		1× TPM, 3× ARGB LED (5 V), 1× RGB LED (12 V)
POST display		yes
Buttons		Start, Reset, Flash BIOS, Clear CMOS
External USB ports		1× 4 (Thunderbolt 4) type C, 3× 3.2 gen. 2 type A, 8× 3.2 gen. 1 type A
Video outputs		1× HDMI 2.1
Network		1× RJ-45 (2,5 GbE) – Intel Killer E3000, WiFi 6E (802.11 a/b/g/n/ac/ax), Bluetooth 5.3
Audio		Realtek ALC4082 (5.1) with DAC ESS Sabre9218
Other external connectors		–
Manufacturer's suggested retail price		475 EUR

/* Here you can add custom CSS for the current table */ /* Lean more about CSS: https://en.wikipedia.org/wiki/Cascading_Style_Sheets */ /* To prevent the use of styles to other tables use "#supsystic-table-2462" as a base selector for example: #supsystic-table-2462 { ... } #supsystic-table-2462 tbody { ... } #supsystic-table-2462 tbody tr { ... } */

ASRock B650E Taichi

In ASRock’s motherboard hierarchy for the AMD AM5 platform, the B650E Taichi is a cheaper derivative of the X670E Taichi. We will touch on the basic things that make the difference in the text.

The width of this board is as much as 267 mm, which at 305 mm of height means that it is an E-ATX format. This is used where the PCB area of smaller ATX boards isn’t enough, although some places are perhaps a little more barren than you might expect. There are only four SATA ports (if you need more, the X670E Taichi has eight) and only three M.2 slots for SSDs (there are four on the X670E Taichi). These are limitations imposed at the chipset level, which naturally has a finite number of PCIe lanes, and this is lower than the X670E variant. On the B650E Taichi, you can also forget about connecting a second pair of USB 3.2 gen. 1 ports on the front panel of the case.

In terms of PCI Express 5.0 interface support for the graphics card, the B650E Taichi has only one such slot. On the X670E Taichi, PCIe 5.0 is supported by two slots. The latter has both of them connected to the CPU. On the B650E Taichi board, the second PCIe ×16 slot is only connected to the chipset, thus the slower PCIe 4.0 lanes. This has one advantage, however, in that the lanes are not redistributed (×8/×8) when using both slots and the full number of lanes is always available.

As befits a high-end board, a backplate is also included with the B650E Taichi. This not only protects the PCB during mounting (there are cases where you tap the SMD into the spacer and the fun is over…), but also contributes to the cooling of the VRM (Vcore and SOC). It is in contact with the PCB through a thick thermal pad.

Two things about the power delivery attract particular attention. One is the enormous number of phases, 27, and the other you have to take a closer look at. On the decorative cover between the heatsink and the external I/O connectors, there is a small grille next to the “Taichi” lettering. Behind it hides a small, 40-millimeter axial fan. If you start looking for its parameters according to its designation (T124010SH), you’ll see from the datasheet that it has double ball bearings (i.e. decent, durable ones) and reaches a speed of 7000 rpm. However, you will not encounter it in practice.

Fans on motherboards are a bogeyman, but mostly probably because users portray unrealistic scenarios of “how noisy it’s going to be”, it’s not. The fan nestled between the fins of a heatsink is to avert situations that can cause possible performance drops due to overheating of voltage regulators during aggressive overclocking. However, even a situation with the Ryzen 9 7950X processor, which runs up against PPT limits, is very far from such a scenario. Even then, the fan does not switch on and the operation of the VRM cooler is passive.

We could not get the fan to run even in conditions outside the wind tunnel, without system cooling, when the passive cooler decreases in effectiveness. Of course, the fan does work and you can modify its behavior in the BIOS (for example, as part of making active cooling reduce temperatures and thus prolong the life of the motherboard), but you definitely don’t need to be afraid of it.

Active cooling is an added value here. It does take up some space that could have been used for a larger heatsink, but it’s already big enough as it is (at 484g without the fan) and quite nicely articulated, so it has a above-standard fin surface area as well. Those few square centimetres would probably be pretty useless already, while removing the fan could mean that the cooler won’t reach such a high cooling performance to keep VRM temperatures below the critical threshold when overclocking hard.

The power delivery, as already mentioned, is robust (it’s quite a crowd of components), but the used voltage regulators Renesas RAA210040 are not among the very top of the range, or rather they are built for a significantly lower current load, so they heat up more than, for example, the Infineons TDA21490 on the MSI MEG X670E Ace board.

The ASRock B650E Taichi also takes advantage of the extra width of the PCB to unconventionally mount one of the three available M.2 SSD slots. The PCIe 5.0-enabled slot is positioned between the 24-pin ATX and DIMM slots (alongside them). In addition to having room to support long 110mm SSDs in these spots, there is also a large heatsink that exceeds the normal 80mm SSDs, but the mounting is customized to accommodate them. It is an aluminium monolith, but it is bigger (weighing 76 grams), but also nicely finned, so it is comparable in effectiveness with the Axagon CLR-M2XL. The SSD coolers are effective overall, including the second one that is shared for the second and third M.2 slots.

One of the highlights of the B650E Taichi is the presence of the Intel JHL8340 Thunderbolt 4 controller with USB4 support. But beware, although you will read in various places about backward compatibility of USB4 with older USB standards, this is not completely true here. USB 3.2 gen. 2×2 is not supported. If you connect a device with support for this interface (typically an external SSD) to the external USB-C connector, it will only use 10 Gbps (instead of 20 Gbps). This may be off-putting to some given that many motherboards externally support USB 3.2 gen. 2, in this case it’s only through the internal connector.

When we mentioned how much less the B650E Taichi has compared to the X670E Taichi, it’s certainly fair to turn it around a bit and point out where it has the upper hand – in almost triple the number (8) of external USB 3.2 gen. 1 connectors. This is thanks to the ASMedia ASM1074 external controllers. There are more external USB ports on the B650E Taichi overall (up to 12 versus 10 on the X670E Taichi), but you have to do without a second USB4 port on the rear panel, and three (instead of five on the X670E Taichi) 10 Gb USB 3.2 gen. 2 ports must suffice.

Intel Killer E3000? This network adapter uses a Realtek RTL8125 chip, but above that is a special Rivet Networks (now Intel) software layer used for packet prioritization, for example, which can come in handy for higher “performance” in games as well. In the context of Ethernet, it’s perhaps a shame that ASRock has dropped the 10-gigabit Aquantia adapters that it has pushed on its boards in this class in the past. However, one can understand that ASRock judged the 2.5 Gb connectivity supported by the chipset to be sufficient.

ASRock also counts on builds without a graphics card, i.e. with the use of an iGPU in the processor, an HDMI connector is brought out for it, but 8K@60 Hz (HDCP 2.3) is also supported via the USB-C connector. What is not expected, on the other hand, is the connection of an audio system with multiple separate satellites. Although ASrock boasts ESS Sabre9218 DACs and WIMA capacitors from audiophile tech, and the codec is Realtek ALC4082, there are only two 3.5mm jacks (line out and microphone). However, multi-channel systems (up to 5.1) can be connected via the S/PDIF optical output.

And what would a “gaming” board be without RGB LED elements? One light guide is on the chipset cooler cover, in its lower half (apparently, so as not to be overshadowed by at least the smaller graphics cards), and the other is also in the lower right quadrant, but from the back. It’s directed at the case’s sheet metal to give a soft light diffusion effect.

What is looks like in BIOS

The initial screen of the user interface won’t surprise you with anything, the layout of the individual elements is traditional regardless of whether you are in advanced or simple mode, where there is a basic component diagnostics and mainly, a button to read the EXPO memory profile.

AMD EXPO can then also be activated in advanced mode in the “OC Tweaker” tab, where there is also an option to adjust the speed of the memory controller. Uncore is at 3000 MHz by default, double that of many other AM5 boards, including the MSI MEG X670E Ace.

Despite Uncore’s high bandwidth, the ASRock B650E Taichi memory subsystem is slower from the perspective of the Aida64 benchmark, and the relatively “weaker” real-world test results correspond to that as well. The reason why this behavior occurs must be sought in the DRAM profile with detailed timing options presets. The basic ones (CL, RDC, RP and RAS) are always the same, but the more detailed ones already differ.

When the factory DRAM mode (AMD Agesa default) is changed to “Aggressive”, the detailed timings and latencies are reduced to a lower number of cycles along with increasing the overall speed of the memory subsystem. The intervals between periodic refreshes of memory contents are also shorter (because DRAM is volatile). This in itself does not help performance, but it may be a means to achieve stability by more aggressively adjusting memory clock speeds and timings.

Overall, the default settings around memories are more conservative, but can be tweaked for better performance with manual intervention. ASRock has grasped this in the interest of the highest compatibility possible to prevent the board from not POSTing after activating the EXPO memory profile.

CPU power management can be adjusted from within the AMD PBO (Precision Boost Overdrive) environment by manually entering both the PPT limit (for a TDP of 105, a value of 142,000 must be entered, which is in mW) and the current limits. It is easier to get to a similar result in the SMU (Advanced AMD CBS\SMU Common Options) settings by enabling “Eco Mode” with 105 W.

The options of the Fan-tastic management interface are a bit poorer compared to competing boards. You can indeed set an arbitrary speed curve for each of the eight headers, but it can only be derived by the temperature from two sources (CPU and MB), which is a bit of a shame considering there are more internal sensors, even on the CPU VRM.

If you don’t like the “drag and drop” technique, the PWM duty cycle for a certain temperature can also be entered in the H/W Monitor by typing in the exact numbers. This tab also exclusively allows you to customize the VRM fan, which Fan-tastic does not detect. Its profile is preset to “Silent”, in which you can’t easily force the cooler to active mode.

However, if you want to make the cooling more effective regardless of the VRM temperature, but at the same time so that you still don’t hear the fan, the “Standard” mode is suitable. In this mode, the speed is always only symbolic during normal operation and the fan only “whispers” a little. In Performance mode it is noisier and at Full speed (about 7000 rpm) it is already rumbling, but that might not bother typical overclocking enthusiasts so much and the most important thing is that the effectiveness of the VRM cooler is not the limit when overclocking.

Gaming tests…

The vast majority of tests are based on the methodology for processors and graphics cards. The choice of games is narrower with motherboards, but for this purpose there is no need for more of them. The processor we use is always the powerful AMD Ryzen 9 79500X or on Intel platforms It’s the Core i9-13900K. These processors highlight both the strengths and weaknesses of any motherboard well. In the past we have tested with two processors, including a cheaper, more low-power model, but we don’t do that anymore. The hypothesis that more expensive motherboards might give an “advantage” to cheaper processors in performance has not been confirmed, so it’s rather pointless.

We’ve selected five titles from games we’re testing in two resolutions. There are significantly fewer games than in the CPU or graphics card tests, but these are just enough for the motherboard tests. Few people consider performance in a particular game when choosing a motherboard. But an indicative overview of which motherboard shapes gaming performance in what way (compared to another motherboard) is necessary. To avoid significant discrepancies over time, we’ve reached for relatively older titles that no longer receive significant updates.

These are Borderlands 3, F1 2020, Metro Exodus, Shadow of the Tomb Raider and Total War Saga: Troy. With newer games, there might be some performance changes over time (updates) and especially in high resolutions with high details. This is one of the test setups (2160p and Ultra, or the highest visual detail but without ray-tracing graphics) that focuses on comparing performance, for which the bottleneck is the graphics card. In other words, it will be clear from these tests which motherboard can affect the performance of which graphics card to what extent for any reasons. In contrast, a setup with Full HD resolution and with graphical details reduced to “High” will also reflect the CPU’s contribution to the final gaming performance.

We use OCAT to record fps, or the times of individual frames, which are then used to calculate fps, and FLAT to analyze the CSV. The developer and author of articles (and videos) for the GPUreport.cz website is behind both.
For the highest accuracy, all runs are repeated three times and average values of average and minimum fps are displayed in the graphs. These multiple repetitions also apply to non-game tests.

… Computing tests, SSD tests, USB ports and network tests

We test application performance in a very similar way to the processor tests. Almost all tests are included, from the easier ones (such as those in a web environment) to those that push the CPU or graphics card to the limit. These are typically tests such as 3D rendering, video encoding (x264, x265, SVT-AV1) or other performance-intensive computing tasks. As with processors or graphics cards, we have a wide range of applications – users editing video (Adobe Premiere Pro, DaVinci Resolve Studio), graphic effects creators (Adobe Premiere Pro), graphic designers or photographers (Adobe Photoshop and Lightroom, Affinity Photo, AI applications Topaz Labs, …) will find their own in the results, and there are also tests of (de)encryption, (de)compression, numerical calculations, simulations and, of course, tests of memory.

SSD performance tests are also important for motherboards. Therefore we test the maximum sequential read and write speeds on an empty Samsung 980 Pro SSD (1 TB) in the well distributed CrystalDiskMark, in all slots. We approach the USB port tests in the same way. We use a WD Black P50 external SSD to test them. It supports fast USB 3.2 gen. 2×2, so it won’t be a bottleneck for even the fastest USB controllers. We report only one result for each USB standard. This is calculated from the average of all available ports.

We won’t deprive you of network bandwidth tests either. We move large files in both directions within a local network between the motherboard network adapters and the Sonnet Solo10G 10-gigabit PCIe card. This from the aforementioned Samsung 980 Pro SSD to the Patriot Hellfire (480 GB), which is still fast enough to not slow down even the 10 Gb adapter.

The results of all performance tests are averaged over three repeated measurements for high accuracy.

CPU settings…

We primarily test processors without power limits, the way most motherboards have it in factory settings. For tests that have an overlap with power, temperature and CPU clock speed measurements, we also observe the behavior of boards with a power limit according to Intel’s recommendations, where we set PL1 to the TDP level (125 W) while respecting the Tau timeout (56 s). The upper limit of the power supply (PL2/PTT) is set in the BIOS according to the official values. For Core i9-13900K it is 253 W, for Core i9-12900K it is 241 W. On AMD platforms with the Ryzen 7950X test processor, the reduced power supply mode represents a TDP setting of 105W with a PPT of 142W. Such a load also corresponds to unconstrained power supply of the Ryzen 7 7700X and Ryzen 5 7600X processors. Aggressive overclocking technologies such as PBO2 (AMD) or MCE (Asus) and similar are not covered in standard motherboard tests.

… and application updates

Tests should also take into account that over time, individual updates may skew performance comparisons. Some applications we use in portable versions that do not update or can be kept on a stable version, but for some this is not the case. Typically games get updated over time, which is natural, and keeping them on old versions out of reality would also be questionable.

In short, just count on the fact that the accuracy of the results you are comparing with each other decreases a bit as time goes on. To make this analysis easier, we’ve listed when each board was tested. You can find this out in the dialog box, where you can find information about the date of testing. This dialog is displayed in the interactive graphs, next to any result bar. Just hover over it.

Methodology: How we measure power draw

In contrast to the Z690/B660 tests, we’ll simplify it a bit and measure only the CPU power draw on the EPS cables. This means that (also for the sake of best possible clarity) we omit the 24-pin measurements. We have already analysed it thoroughly and the power draw on it doesn’t change much across boards. Of the ten boards tested with an Alder Lake processor (Core i9-12900K), the power draw at 12 volts of the 24-pin connector ranges from 37.3–40.4 W (gaming load, graphics card power supply via PCI Express ×16 slot), at 5V (memory, ARGB LEDs and some external controllers) then between 13.9–22.3 W and finally at the weakest, 3.3-volt branch, the power draw of our test setup tends to be 2.2–3.6 W.

On top of the CPU power draw, which also takes into account the efficiency of the power delivery, this adds up to some 53–66 W under gaming/graphics load and only 15–25 W outside of it, with the graphics card idle. We already know all this from older tests, and it will be no different on the new boards, and as the number of measurements increases, reducing measurements that worsen orientation is beneficial. But from the text above, you know how much to add for the total power draw of the motherboard components to the CPU’s majority power draw.

The situation will be a bit different on AMD platforms, for those we will deal with what is the power draw on which branch of the 24-pin, but already in a separate article that will better highlight this topic. In a large comprehensive motherboard test, these measurements disappear, they do not attract enough attention.

We measure the power draw of the CPU (and its VRM) on the power supply cables, with calibrated Prova 15 current clamps and a calibrated Keysight U1231A multimeter. The clamps measure the electric current, the multimeter measures the electric voltage. In the union of these two electrical quantities, we finally obtain the exact power draw. We measure this in different loads on the CPU. The maximum multithreaded load is represented by Cinebench R23.

Lower, gaming load by Shadow of the Tomb Raider (1080p@high), single-threaded load by audio encoding (reference encoder 1.3.2, FLAC with bitrate 200 kbps) and idle power draw is measured on the Windows 10 desktop when only basic operating system processes and launchers of some test applications are running in the background.

Methodology: Temperature and clock speed measurements

By far the most critical part in terms of temperatures on the motherboard is the power delivery (VRM) for the CPU. This is where we return to the Fluke Ti125 thermal imager, which produces temperature maps that can be used to locate the average temperature, as well as the hottest point. We record both these values (average and maximum temperature on the Vcore) in graphs, and we will later evaluate the efficiency of the VRM heatsinks based on the maximum one. However, we lack a suitable thermometer for that yet. Of course, the thermovision is implemented without a heatsink, and a thermocouple needs to be installed on the hottest MOSFET to detect the reduction of temperature with a heatsink. This will be added soon.

Thermal imaging always relates to operating with the more powerful of the pair of test processors. With it, the differences and possible limitations or impending risks (for example, even from thermal throttling) become more apparent. In order to have a good view of the VRM, we use an Alphacool Eisbaer 360 liquid cooler with the fans fixed at full power (12 V) instead of a tower cooler (from the CPU tests). The temperature tests also include CPU temperatures for completeness, and we also test the efficiency of the supplied SSD heatsinks as part of the motherboard tests. These are already included with virtually all better motherboards, and so the question naturally arises whether to use them or replace them with other, more finned ones. We will test these heatsinks on a Samsung 980 Pro SSD during ten minutes of intense load in CrystalDiskMark. Finally, the temperature of the chipset’s southbridge and the cooling efficiency in this direction is noteworthy as well.

All tests are conducted in a wind tunnel, so full system cooling is provided. This consists of three Noctua NF-S12A PWMs@5V (~550 rpm) . Two of which are intake, one is exhaust. But the three fast AIO fans also function as exhaust fans, so there is underpressure in the case.

The temperature at the entrance to the tunnel is properly controlled and ranges between 21-21.3 °C. Maintaining a constant temperature at all times during testing is important not only for the accuracy of the temperature measurements, but also because a higher or lower ambient temperature also affects the eventual behaviour of the processors’ boost. And we also properly monitor and compare the clock speeds, whether under all-core load or even single-threaded tasks. We use the HWiNFO application to record the clock speeds and temperatures of the cores (sampling is set to two seconds).

Maintaining a constant temperature at the intake is necessary not only for a proper comparison of processor temperatures, but especially for objective performance comparisons. The clock speed development, and specially the single core boost, is precisely based on the temperature. Typically in summer, at higher temperatures than is normal in living quarters in winter, processors can be slower.

Temperatures are always read as maximum (both from the VRM thermovision and average, but still from the local maximum values at the end of Cinebench R23). For Intel processors, for each test we read the maximum temperature of the cores, usually all of them. These maxima are then averaged and the result represents the final value in the graph. From the single-threaded workload outputs, we extract only the recorded values from the active cores (there are usually two of these, and they alternate between each other during the test). For AMD processors it is a bit different. They don’t have temperature sensors for each core. In order to make the procedure methodically as similar as possible to the one we apply on Intel processors, we define the average temperature of all cores by the highest value reported by the CPU Tdie (average) sensor. However, for single-core workloads we already use the CPU sensor (Tctl/Tdie), which usually reports a slightly higher value that better corresponds to hotspots of one or two cores. However, these values as well as the values from all internal sensors should be taken with a grain of salt, the accuracy of sensors across CPUs varies.

Clock speed evaluation is more accurate, each core has its own sensor even on AMD processors. However, unlike the temperatures, we write the average values of the clock speeds during the tests in the graphs. We monitor the temperatures and clock speed of the CPU cores in the same tests in which we also measure power draw. Thus, sequentially from the lowest desktop idle load in Windows 10, through audio encoding (single-threaded load), gaming load in Shadow of the Tomb Raider to Cinebench R23.

The higher price of motherboards is one of the main topics around desktop PCs. Especially in the context of the AMD platform. However, if you decide to build your rig on DDR5 memory with the vision of a longer life expectancy, the B650 Tomahawk WiFi, in comparison to similarly equipped motherboards with Intel chipsets, is rather higher on the “price” scales and with a more attractive price/value ratio.

Test setup

Alphacool Eisbaer Aurora 360 liquid cooler w/ a metal backplate

G.Skill Trident Z5 Neo memory (2×16 GB, 6000 MHz/CL30). Motherboards with DDR4 memory support are tested with Patriot Blackout (4×8 GB, 3600 MHz/CL18) and Z690/B660 motherboards with DDR5 memory support were tested with Kingston Fury Beast (2×16 GB, 5200 MHz/CL40)

MSI RTX 3080 Gaming X Trio graphics card

Patriot Viper VP4100 (1 TB) and Patriot Viper VPN100 (2 TB) SSDs

Note: Graphics drivers used at the time of testing: Nvidia GeForce 466.77 and OS Windows 10 build 19045.

3DMark

We use 3DMark Professional for our tests and from the tests, Night Raid (DirectX 12), Fire Strike (DirectX 11) and Time Spy (DirectX 12). In the graphs you will find the CPU sub-scores, the combined scores, as well as the graphics scores. From this you can see to what extent a given CPU is limiting the graphics card.

Borderlands 3

Test environment: resolution 1920 × 1080 px; graphics settings preset High; API DirectX 12; extra settings Anti-Aliasing: None; test scene: built-in benchmark.

Test environment: resolution 3840 × 2160 px; graphics settings preset Ultra; API DirectX 12; no extra settings; test scene: built-in benchmark.

F1 2020

Test environment: resolution 1920 × 1080 px; graphics settings preset High; API DirectX 12; extra settings Anti-Aliasing: off, Skidmarks Blending: off; test scene: built-in benchmark (Australia, Clear/Dry, Cycle).

Test environment: resolution 3840 × 2160 px; graphics settings preset Ultra High; API DirectX 12; extra settings Anti-Aliasing: TAA, Skidmarks Blending: off; test scene: built-in benchmark (Australia, Clear/Dry, Cycle).

Metro Exodus

Test environment: resolution 1920 × 1080 px; graphics settings preset High; API DirectX 12; no extra settings; test scene: built-in benchmark.

Test environment: resolution 3840 × 2160 px; graphics settings preset Extreme; API DirectX 12; no extra settings; test scene: built-in benchmark.

Shadow of the Tomb Raider

Test environment: resolution 1920 × 1080 px; graphics settings preset High; API DirectX 12; extra settings Anti-Aliasing: off; test scene: built-in benchmark.

Test environment: resolution 3840 × 2160 px; graphics settings preset Highest; API DirectX 12; extra settings Anti-Aliasing: TAA; test scene: built-in benchmark.

Total War Saga: Troy

Test environment: resolution 1920 × 1080 px; graphics settings preset High; API DirectX 11; no extra settings; test scene: built-in benchmark.

Test environment: resolution 3840 × 2160 px; graphics settings preset Ultra; API DirectX 11; no extra settings; test scene: built-in benchmark.

PCMark

Geekbench

Speedometer (2.0) and Octane (2.0)

Test environment: To ensure that results are not affected by web browser updates over time, we use a portable version of Google Chrome (91.0.472.101), a 64-bit build. Hardware GPU acceleration is enabled as well, as it is by default for every user.

Note: The values in the graphs represent the average of the scores obtained in the subtasks, which are grouped according to their nature into seven categories (Core language features, Memory and GC, Strings and arrays, Virtual machine and GC, Loading and Parsing, Bit and Math operations, and Compiler and GC latency).

Cinebench R20

Cinebench R23

Blender@Cycles

Test environment: We use well distributed projects BMW (510 tiles) and Classroom (2040 tiles) and the renderer Cycles. Render settings are set to None, with which all the work falls on the CPU.

LuxRender (SPECworkstation 3.1)

Adobe Premiere Pro (PugetBench)

Test environment: PugetBench tests set. We keep the version of the application (Adobe Premiere Pro) at 15.2.

DaVinci Resolve Studio (PugetBench)

Test environment: set of PugetBench tests, test type: standard. App version of DaVinci Resolve Studio is 17.2.1 (build 12).

Graphics effects: Adobe After Effects

Test environment: set of PugetBench tests. App version of Adobe After Effects is 18.2.1.

Test environment: For video conversion we’re using a 4K video LG Demo Snowboard with a 43,9 Mb/s bitrate. AVC (x264) and HEVC (x265) profiles are set for high quality and encoder profiles are “slow”. HandBrake version is 1.3.3 (2020061300).

x264 and x265 benchmarks

Naposledy sme sa zaoberali základnou doskou, ktorá, ktorá je aj vďaka nižšej cene vhodná najmä na použitie s lacnejšími procesormi. Teraz tu máme o zhruba 50 eur drahšiu Gigabyte B660 Aorus Master DDR4. Príplatok tu má jasné opodstatnenie a odzkadľuje sa na lepších vlastnostiach. Napájacia kaskáda je výrazne efektívnejšia, chladiče sú účinnejšie a výbava je celkovo bohatšia, vrátane svetielok.

Audio encoding

Test environment: Audio encoding is done using command line encoders, we measure the time it takes for the conversion to finish. The same 42-minute long 16-bit WAV file (stereo) with 44.1 kHz is always used (Love Over Gold by Dire Straits album rip in a single audio file).

Encoder settings are selected to achieve maximum or near maximum compression. The bitrate is relatively high, with the exception of lossless FLAC of about 200 kb/s.

Note: These tests measure single-thread performance.

FLAC: reference encoder 1.3.2, 64-bit build. Launch options: flac.exe -s -8 -m -e -p -f

MP3: encoder lame3.100.1, 64-bit build (Intel 19 Compiler) from RareWares. Launch options: lame.exe -S -V 0 -q 0

AAC: uses Apple QuickTime libraries, invoked through the application from the command line, QAAC 2.72, 64-bit build, Intel 19 Compiler (does not require installation of the whole Apple package). Launch options: qaac64.exe -V 100 -s -q 2

Opus: reference encoder 1.3.1, Launch options: opusenc.exe –comp 10 –quiet –vbr –bitrate 192

Adobe Photoshop (PugetBench)

Test environment: set of PugetBench tests. App version of Adobe Photoshop is 22.4.2.

Affinity Photo (benchmark)

Test environment: built-in benchmark.

Topaz DeNoise AI, Gigapixel AI and Sharpen AI. These single-purpose applications are used for restoration of low-quality photos. Whether it is high noise (caused by higher ISO), raster level (typically after cropping) or when something needs extra focus. The AI performance is always used.

Test settings for Topaz Labs applications. DeNoise AI, Gigapixel AI and Sharpen AI, left to right. Each application has one of the three windows

Test environment: As part of batch editing, 42 photos with a lower resolution of 1920 × 1280 px are processed, with the settings from the images above. DeNoise AI is in version 3.1.2, Gigapixel in 5.5.2 and Sharpen AI in 3.1.2.

The processor is used for acceleration (and high RAM allocation), but you can also switch to the GPU

WinRAR 6.01

7-Zip 19.00

TrueCrypt 7.1a

Aida64 (AES, SHA3)

Aida64, FPU tests

FSI (SPECworkstation 3.1)

Kirchhoff migration (SPECworkstation 3.1)

Python36 (SPECworkstation 3.1)

SRMP (SPECworkstation 3.1)

Octave (SPECworkstation 3.1)

FFTW (SPECworkstation 3.1)

Convolution (SPECworkstation 3.1)

CalculiX (SPECworkstation 3.1)

RodiniaLifeSci (SPECworkstation 3.1)

WPCcfd (SPECworkstation 3.1)

Poisson (SPECworkstation 3.1)

LAMMPS (SPECworkstation 3.1)

NAMD (SPECworkstation 3.1)

Memory tests…

… and cache (L1, L2, L3)

M.2 (SSD) slots speed

USB ports speed

Ethernet speed

In the second test setup we use a Sonnet Solo10G network card to measure the LAN adapter transfer speeds

Analysis of power draw without power limits

Analysis of power draw with power limits

Achieved CPU clock speed w/o power limits…

… and with power limits

Disclaimer: The temperatures of the Core i9-12900K with the Core i9-13900K are incomparable. With the Intel Raptor Lake processor (Core i9-13900K) we use a metal backplate, while with Alder Lake (Core i9-12900K) the Alphacool Eisbaer Aurora 360 cooler has a plastic backplate. The latter has lower pressure and the heat transfer intensity is worse, as our tests show.

… and with power limits

VRM temperature w/o power limits…

… and with power limits

SSD temperature

Chipset temperature (south bridge)

Conclusion

The ASRock B650E Taichi motherboard is based on a rather unique design. Competing Asus, Gigabyte, and MSI don’t really have a similar board with the B650E chipset. The E-ATX format isn’t used to its full potential here and some things about it may seem a bit bizarre (for example, only four SATA ports), but others have room to shine in the best sense of the word. For example, the really elite SSD cooler above the first M.2 slot. It is the most effective design we have tested on motherboards so far (and there have been a few of those…), yet in its atypical position it doesn’t interfere with anything.

But most people don’t buy a motherboard because of the SSD cooler, the features of the CPU power delivery probably matter more. The latter is monstrous in this case, although the voltage regulators are “weaker”, but their relatively higher temperature compared to other high-end motherboards is compensated by an effective cooler. And it is such (effective) even without the activity of the fan. It only occurs without interfering with its settings in critical situations of very aggressive overclocking, i.e. in conditions where the fan increases the TDP of the cooler. This creates better conditions for tuners to wring the maximum out of the processor. Without a fan, purely with passive cooling, you might not get to this point because you will hit a limit in VRM cooling rather than a limit in CPU capabilities.

CPU power supply efficiency in high load is high, at the level of the MSI MEG X670E Ace (It’s worse in idle, where we’ve measured the highest power draw so far. Not only within the AMD AM5 platform, but overall).The ASRock B650E Taichi’s voltage regulators may be less efficient and heat up more, but the B650E Taichi took the top spot in terms of lowest power draw thanks to its more efficient power supply voltage management. This gives less margin for undervolting, but in the default settings the lowest-power operation is achieved under very high CPU load (with the AVX instructions engaged). The differences are small, and it’s true that they’re also due to slightly lower computing performance (on the B650E Taichi). Why? (answer below the image with the editorial award “Approved“)

In addition to the lower average CPU core clock speed in the order of tens of megahertz, this is also due to the lower memory speed associated with the less aggressive DRAM mode. This means that some settings prioritize the best possible compatibility over maximum performance. This could hopefully limit the cases where, especially with faster modules, the board does not POST after activating a memory profile. In case it does POST, there is of course room for additional tuning where the memory subsystem bandwidth can be increased and thus you get better performance in practice. For some applications more, for some less, it depends on how dependent they are on memory speeds.

Overview of tests of speeds of M.2 slots, USB ports and Internet connection shows that results are achieved as expected (neither significantly above nor significantly below the average of other boards). However, in sequential read and write tests, the Intel Killer E3000 network adapter does not show its main advantage over the Realtek RTL8125 without the Rivet software add-on. Thanks to it, prioritization of selected packets can achieve, for example, lower ping in games and the like.

The features around the audio adapter are also above standard, and the portion of external USB ports (led by the 40 Gb USB4 interface) is also exceptionally generous. But there are some things that don’t belong in this class of boards. Whether it’s the USB 3.2 gen. 2×2 only in internal form (i.e. no representation between external connectors), “only” three M.2 slots for SSDs, and someone may also miss PCI Express 5.0 support in the second PCIe ×16 slot, as the more expensive Taichi model (with X670E chipset) has it. But at least on the B650E Taichi there is no redistribution of PCIe lanes, as they are only routed from the processor to the first slot (on the X670E Taichi to both), the second PCIe ×16 slot is connected to the south bridge (B650E chip).

We can’t help but say that the B650E Taichi is an expensive board and there will be plenty of users who won’t be able to justify the price with its features. But then there will also be those for whom it will be the closest thing to what they are looking for. It’s as always about different preferences for different things.

English translation and edit by Jozef Dudáš


ASRock B650E Taichi
+ Extremely powerful 27-phase power delivery (VRM)...
+ ... can handle an overclocked Ryzen 9 7950X with margin
+ Very efficient power management under higher load
+ Effective, very heavily finned VRM cooler with active capability
+ High-end SSD cooler, we haven't come across a more effective one on a motherboard yet
+ Detailed fan management options
+ As many as twelve fast USB connectors on the rear I/O panel...
+ ... including USB4 port
- ... but without backwards compatibility with USB 3.2 gen. 2×2
- Higher idle power draw
- Absence of several things that are common in this price category...
- ... only one internal 19-pin USB, only four SATA ports or only three M.2 slots for SSDs
Suggested retail price: 475 EUR

/* Here you can add custom CSS for the current table */ /* Lean more about CSS: https://en.wikipedia.org/wiki/Cascading_Style_Sheets */ /* To prevent the use of styles to other tables use "#supsystic-table-2463" as a base selector for example: #supsystic-table-2463 { ... } #supsystic-table-2463 tbody { ... } #supsystic-table-2463 tbody tr { ... } */

Some of the tested boards are also available in the Datacomp e-shop

Special thanks to Blackmagic Design (for licenses for DeNoise AI, Gigapixel AI and Sharpen AI) and Topaz Labs (for licenses for DeNoise AI, Gigapixel AI and Sharpen AI)

Continue: Memory and cache tests