Tom's Hardware Verdict
A solid attempt at a Raspberry Pi alternative, it just falls short by not unleashing the full power of the N100. The form factor and the RP2040 GPIO along with the x86 CPU make this a great deal but there are caveats to take into account.
Pros
- +
Comparable Price to the Pi 5
- +
Powerful x86 CPU
- +
Pi-like form factor
- +
Onboard NVMe SSD slot
- +
RP2040 GPIO
Cons
- -
Full CPU power not used
- -
Needs more power
- -
Just too big for Raspberry Pi cases
Why you can trust Tom's Hardware
The Raspberry Pi form factor is seen as the ideal and it has been aped by many different boards since it was revised back in 2014. Radxa, traditionally makers of Arm based single board computers (SBC) have released their first Intel N100 based board, the Radxa X4. The X4 shares a similar form factor to the Raspberry Pi 5, but it has an Intel x86 CPU at its heart.
On paper, the N100’s quad-core 3.4 GHz CPU is more powerful than the Raspberry Pi 5’s quad-core Arm Cortex-A76 64-bit CPU running at 2.4 GHz. We’ve already seen proof of that in the Lattepanda Mu, which also uses the N100 but in a Compute Module 4 form factor. But how will the change of form factor impact the performance of the N100?
There is a GPIO on the Radxa X4, but the Intel N100 has nothing to do with it. Instead, there is a Raspberry Pi RP2040 Arm based SoC. The same as found inside the Raspberry Pi Pico and Pico W. It may not be the newer RP2350 as found in the Raspberry Pi Pico 2, but it has a full 40 pin GPIO that we can use, with no USB cables to connect the X4 to the RP2040.
Could the Radxa X4 take the throne as the best single board computer? Or is its dominion still absolute? Let's find out.
Radxa X4 Technical Specifications
Header Cell - Column 0 | Radxa X4 | LattePanda Mu |
---|---|---|
Processor | Intel N100 Quad-Core, Four-Thread up to 3.4 GHz | Intel N100 Quad-Core, Four-Thread up to 3.4 GHz |
GPU | Intel UHD Graphics up to 750 MHz | Intel UHD Graphics up to 750 MHz |
RAM | 4/8/12GB LPDDR5 4800 MHz (8GB in review unit) | 8GB LPDDR5 4800 MHz |
Storage | 1 x M.2 M Key PCIe 3.0 4-lane 2230 NVMe SSD (not included) Optional onboard eMMC | 64GB eMMC |
Connectivity | 2.5 GB Ethernet Wi-Fi 5 and 6 Bluetooth 5 and 5.2 | Via carrier board |
GPIO | 40 pin GPIO via Raspberry Pi RP2040 2 x SPI 2 x I2C 2 x UART 16 x PWM 8 x Programmable IO (PIO) | 4 x UART 4 x I2C 64 GPIOs via carrier board |
USB | 1 x USB 2 Type A 3 x USB 3 Type A | Via carrier board Up to 8 x USB 2 4 x USB 3.2 |
Display | 2 x Micro HDMI up to 4Kp60 | Via carrier board 1 x eDP 1.4 3 x HDMI 2 / DisplayPort 1.4 |
Expansion Slots | None | Varies depending on carrier board |
Co-Processor | Raspberry Pi RP2040 Arm Cortex M0+ Dual Core at 133 MHz | N/A |
Power | USB Type C PD 12V at 2.5A | Varies depending on carrier board |
Dimensions | 85 x 56mm | 60 x 69.6mm |
Price | $60 for 4GB $80 for 8GB Add $9 for 32GB eMMC $10 for 64GB eMMC | $139 for module $190 with Primer Carrier and Active Cooler. $240 with Full Evaluation Carrier |
Design of the Radxa X4
The Radxa X4 is very much in the Raspberry Pi form factor design camp. It closely mimics the Raspberry Pi 4 but we are certain that it won’t fit inside the best Raspberry Pi cases as along the GPIO edge the board protrudes out by a few extra millimeters. If you want a case, then the $15 Radxa Heatsink is a smart purchase. Why? Because it provides cooling and protection for the board. The Intel N100 needs cooling! You cannot run this board for long without some form of cooling, more on that later. We feel that the $15 case should be included, simply because if you don’t have it, then the N100 will throttle, badly!
To attach / remove the Radxa X4 to the heatsink case requires 10 screws! You need to remove the “legs” of the case in order to gain access to the four screw holes necessary to hold the X4 to the heatsink.
Instead of a PCIe connector, like on the Raspberry Pi 5, Radxa X4 has an M.2 M key slot for a 2230 NVMe SSD. This is a clever way to add fast and cheap storage to an SBC without the need for extra components. The drive will rest over the Wi-Fi / Bluetooth chip, but the antennas mean that wireless performance is not impacted.
Wi-Fi and Bluetooth antennas stick out like the antenna of a bug, and the real-time clock battery provides a bright yellow “tail” which makes the overall aesthetic seem untidy. We’d prefer an integrated antenna and RTC battery, but those would need extra space on the PCB.
There are no CSI or DSI ports for Raspberry Pi cameras or displays. This is pretty much an Intel desktop PC in an SBC form factor.
Radxa X4 Performance Tests
We previously mentioned the Radxa Heatsink for X4, a $15 heatsink and active cooling case that keeps the N100 tamed. Unusually the Radxa X4 attaches upside down to the case, making GPIO access awkward, but bearable. The included cooling pads work ok, but others in the SBC community have noted that they can crumble and break apart. Hopefully this is just down to early review unit issues. If yours is a bit crumbly, use some good thermal paste instead.
Attaching the heatsink case to the X4 is simple, once you take it apart. The only tricky part is the fan connection. The cable is way too short and it requires some dexterity to correctly install, without damaging the cable or its connector.
The fan is always on, and always at 100%. There is no PWM speed control. It is all or nothing, and you really need it on, to move the warm air from the heatsink.
How hot does it get? At idle, the CPU sits at around 34 degrees Celsius, comfortable and 3°C cooler than the Lattepanda Mu, 5.5°C cooler than the Raspberry Pi 5. Impressive. Under stress, the CPU hits 62°C, 22°C cooler than the Lattepanda Mu’s 84°C. But the Raspberry Pi 5 runs cooler at 59.3°C
Overall, cooling performance is great, but that is solely down to the CPU fan being at full power all of the time. Which leads nicely to power consumption. The Raspberry Pi 5 is the clear winner here. Its Arm CPU sips power while the Intel N100 drinks deeply from the well.
At idle, the Radxa X4 uses 4.84 Watts of power, the Lattepanda Mu 4.9 Watts, and the Raspberry Pi 5 2.6 Watts. During the stress test, the power consumption jumps to 16.9 Watts for the Radxa X4, then settles to 10.89 Watts (all CPU cores at 2.1 GHz). The Lattepanda Mu hit 15.8 Watts, and the Raspberry Pi 5 is way down at 6.8 Watts.
What about boot times? This is largely down to what NVMe drive you are using. We haven’t got an eMMC module to test, so you’ll need to pick up the best SSDs for the job. Our test SSD came from a recent Steam Deck upgrade. The Phison based 256GB NVMe SSD is more than enough for the Radxa X4.
The Radxa X4 booted faster than the LattePanda Mu with Windows 11 (18.3 seconds versus 31.7 seconds) but for Ubuntu, The LattePanda Mu came in at 22.7 seconds, versus the 25.6 seconds on the Radxa X4. Note that the LattePanda Mu was running Ubuntu 22.04, and the Radxa X4 24.04. The three second difference between the two can be chalked up to human error and new features in the more recent OS.
Geekbench 6 is where we see the computational power of the N100 shine, but not as brightly as the LattePanda Mu. The Radxa X4’s form factor means that there isn’t enough PCB to fully handle the N100’s 3.4 GHz clock. Later on we tested gaming on the Radxa X4, and noted that the CPU never went above 3 GHz for any usable length of time.
The Radxa X4’s single and multi scores for both Windows and Linux were lower than the LattePanda Mu’s equivalent scores. Enough to show that the CPU was being throttled by the inadequate cooling provided by the official cooler. The scores are better than the Raspberry Pi 5, which we are using as a comparison. But the Raspberry Pi 5 idles at 2.5W, half the idle power consumption of the N100.
Using the Radxa X4
Essentially the Radxa X4 is an x86 PC in an SBC form factor. That means we can install any operating system that we wish. Our review unit came with no onboard storage so we dug through our bag full of NVMe SSDs and found a Phison based drive that used to be inside our Steam Deck. We installed Ubuntu 24.04 to the drive and tinkered.
Ubuntu 24.04 ran smoothly, even playing back a 1080p60 YouTube video with only a few dropped frames. The Gnome desktop of Ubuntu was responsive, and never felt sluggish. That said, if you want to eke out the most performance, install a lighter window manager. LXDE or XFCE (Lubuntu and Xubuntu respectively) will give you a little more performance.
Remember, the Radxa X4 is based on an x86 CPU and that means we can install pretty much any Linux distro. DietPi, the lightweight Linux distro would be an ideal choice. It can be used to make Linux appliances via its easy to use menu system. An N100 powered home server created by DietPi would be a great balance of power efficiency and performance.
Windows 11 on the Radxa X4 is a little slower than Linux. For basic browsing and general work tasks, it will get the job done. For maker projects, machine learning, robotics etc, the N100’s power and the GPIO provide a great platform for advanced projects.
One thing that permeates across Linux and Windows is Radxa’s documentation. It is “all over the place” making it harder to get the basics done. For example the Linux install was a joy. It just worked, even the RP2040 GPIO just worked with Thonny. On Windows 11 I had to download many driver packages for Wi-Fi, Ethernet, Bluetooth etc. I spent 90 minutes installing Windows 11 and the many drivers before I could do anything. Linux, 20 minutes tops.
Radxa X4 versus the Raspberry Pi 5
The N100 has a top speed of 3.4 GHz, but during our time with the board we never managed to hit that high. At best when running a y-cruncher stress test we saw 3 GHz for around five seconds, the CPU then dropped down to 2.1 GHz for all cores. So on paper, the N100 is more powerful than the Arm CPUs of the Raspberry Pi 5, and the LattePanda Mu proved that, but the Radxa X4 just hits the brakes before it reaches top speed. We went into the BIOS but could not find a reliable means to up the power to the CPU, noting that Intel states that the N100 is a 6W chip, so there isn’t much more power that we can provide. Better cooling may enable higher performance, but the stock cooler does a decent job, we just can’t hit the highest speed.
The Radxa X4 is priced similarly to the Raspberry Pi 5, so it all boils down to your power budget and CPU architecture preference. Both have 4 and 8GB RAM models, with the N100 offering a possible 16GB RAM in the future. The Radxa X4 page does list a 12GB model for sale, which may offer a little better performance over the 8GB model, especially when using the GPU.
Radxa X4 RP2040 GPIO
Yes there is a full 40-pin GPIO on the Radxa X4, but if you are hoping to use the best Raspberry Pi HATs, then I am sorry to say that you can’t just place them on. Firstly, because the X4 is upside down and inside the heatsink case, HAT access is limited. You could use a breakout board, but then the next issue is pinout. The 40 pin GPIO appears to be the same as the Raspberry Pi, but we can bet that it doesn’t behave in the same way. Check your pinout before connecting any HATs! Lastly, software. Yes we have an RP2040 which can run MicroPython, CircuitPython, C and any other Pico compatible language, but that doesn’t mean that HATs will work.
Using the RP2040 with the GPIO requires us to flash our chosen firmware to the RP2040, just like how we would for a Pico connected via USB. The BOOTSEL button is located just under the D of the Radxa logo on the board. Drop the UF2 file onto the RPI_RP2 drive and then use your favorite code editor to write code to the device. In the Thonny workflow, we connect to a serial device via a COM port, CircuitPython uses a USB flash drive approach. It's easy and in a few moments we had blinking lights and NeoPixels running directly from the GPIO.
Can the Radxa X4 game?
Yes it can, but unlike the LattePanda Mu’s N100 performance, the Radxa X4 leaves a lot to be desired. The problem that we face is that the N100 CPU is not receiving adequate cooling to hit the 3.4 GHz top speed. The CPU will throttle down to 1.5 GHz with Stray at 720P 50% resolution scaling and low textures. We only managed a high of 22 FPS even after all that tweaking. Boomer shooter Warhammer 40,000: Boltgun is always my favorite test, but even this game at 720P performed poorly. Any weapons with particle effects will drop the FPS down into the teens.
Hades, our favorite rogue-like dungeon crawler (it's awesome on Steam Deck) fared better than expected. It had the same issues at 1080P as the LattePanda Mu, but 720p saw improved results. Call of Duty 4: Modern Warfare at 1024 x 768 was an interesting case. On the LattePanda Mu, it played perfectly, and while it played well on the Radxa X4, there was a significant performance difference in the FPS.
The bottom line for gaming on the Radxa X4. Use it for Steam Link, cloud gaming from the comfort of your couch. Older games will run well enough, but your mileage may vary.
Who is the Radxa X4 for?
If you want a little more horsepower than the Raspberry Pi 5, but don’t want to sacrifice the form factor or pay any more than a Raspberry Pi 5. If you need a cheap x86 PC to control makerspace machinery like laser cutters, CNC tools or to act as cheap Windows machines, then the Radxa X4 will do the job. Robotics and machine learning projects will benefit from the extra CPU horsepower, if you can unlock its full potential.
The onboard RP2040 GPIO is a great help and elevates the Radxa X4 over other Intel SBCs.
Bottom Line
I like the Radxa X4. It may have the same heart as the LattePanda Mu, but it is a totally different beast. The Raspberry Pi 5 form factor and the RP2040 GPIO elevates the Radxa X4 as a real Pi 5 alternative for makers. It is priced the same, looks similar, and can be used in maker projects.
That said, the N100 is restricted by the form factor, so if you want to get the most from the N100, you’ll need the LattePanda Mu. But that comes in at a much steeper price.
It boils down to what do you want to do with an SBC? If you need, or want the power of the N100 in your project, get the Radxa X4. You’ll have to factor in its power requirements, especially for mobile projects. If you don’t need the extra power, stick with the Raspberry Pi 5; it is a potent platform that sips power compared to the N100 of the Radxa X4.
Les Pounder is an associate editor at Tom's Hardware. He is a creative technologist and for seven years has created projects to educate and inspire minds both young and old. He has worked with the Raspberry Pi Foundation to write and deliver their teacher training program "Picademy".
30-year-old Pentium FDIV bug tracked down in the silicon — Ken Shirriff takes the microscope to Intel's first-ever recall
Cyberpunk 2077 update 2.2 claims to improve Arrow Lake performance by up to 33%, theoretically matching the Ryzen 7 7800X3D
Empyrean Technology gives control to CEC after U.S. blacklisting — China’s top developer of chip design systems hands reins to state-owned firm
-
chrsphr I grabbed one of these to add as a proxmox node. Got the advantage of being an x86 system so I can extended my existing proxmox system, but small enough to be powered by PoE.Reply
It's also got pretty good video hardware encoding/decoding, for the size, I think.
Either way, a fun little toy and an interesting change from ARM boards -
ryann Most PC's player got tips for using Intel CPU: BIOS can set the PL1 to improve CPU performance, and c-state enable will make cpu running over 3Ghz. just make sure CPU cooling has margin.Reply
It can be regarded as an additional performance, of course it also requires more power. -
bit_user
Disagree. Its main benefit is that it packs lots of features in the smallest space. However, it's really not ideal for use cases like ultra-mini desktops, where it would be better to have a more NUC-like arrangement of connectors on the front and back edges, only. For this sort of thing, the Pi's size also poses challenges for cooling and having things like M.2 slots, which the Pi 5 could only do by adding an extra carrier board. The external antenna and RTC battery further make the case that this is an inappropriate form factor.The article said:The Raspberry Pi form factor is seen as the ideal and it has been aped by many different boards since it was revised back in 2014.
Furthermore, there are standard form factors, like nano-ITX and min-STX that would provide us with a broader array of cases, if more Pi competitors would actually use them.
If they put an active cooler on it, like the Pi 5 can use, then it would be fine.The article said:how will the change of form factor impact the performance of the N100?
No, it's a laptop SoC that can be found in many Chromebook-class machines.The article said:This is pretty much an Intel desktop PC in an SBC form factor.
In my experience, thermal pads have been a disaster with the more powerful N97. The biggest cooling upgrade I made was to remove the spacers and replace the thermal pad with good quality heatsink TIM.The article said:The included cooling pads work ok
That's really unfortunate, as is their decision to go with a non-standard fan header. The SoC is capable of idling at a couple Watts but can boost at up to 25 W.The article said:The fan is always on, and always at 100%. There is no PWM speed control.
I think the "stress test" used to measure that temperature of 62 C is not very stressful. As I've said in other articles, written by Les Pounder, you can't just fire up stress-ng without paying attention to what stress test it's running, as some of them hardly stress the CPU at all. Try --cpu-method=fft.
Because, last I checked, browsers are still using software decoding (would be nice if you actually said anything about how you tested it).The article said:Ubuntu 24.04 ran smoothly, even playing back a 1080p60 YouTube video with only a few dropped frames.
If you use something that supports Intel's hardware decoder, like VLC, then I think it should even do 4k/60 AV1 playback.
That's because you don't understand what you're trying to measure. I'm having a bit of trouble finding the multi-core frequency limits for the N100, but the limit of 3.4 GHz only applies to a single core. Basically, as soon as another core is doing anything, it can drop to the 2-core frequency limit. The 4-core limit is much lower, still.The article said:The N100 has a top speed of 3.4 GHz, but during our time with the board we never managed to hit that high. At best when running a y-cruncher stress test we saw 3 GHz for around five seconds, the CPU then dropped down to 2.1 GHz for all cores.
No matter how high your power limits or how good your cooling, a N100 will never run 2 or 4 cores at 3.4 GHz. This is baked right into the firmware and you can't change or override it.
What you want is this:The article said:We went into the BIOS but could not find a reliable means to up the power to the CPU, noting that Intel states that the N100 is a 6W chip, so there isn’t much more power that we can provide.
https://github.com/horshack-dpreview/setPL
It works fine on my N97 (Ubuntu 24.04). You just need to be sure you have the prerequisites installed. The settings will revert to the manufacturer's defaults, after a reboot. When you run the script, it prints the old values for PL1 & PL2, so you can see what they were set to.
These use cases generally fall in the category of where you'd wan ]in-band ECC. The ODROID-H4 series provide that as a BIOS option (more on that, later).The article said:If you need a cheap x86 PC to control makerspace machinery like laser cutters, CNC tools or to act as cheap Windows machines, then the Radxa X4 will do the job. Robotics and machine learning projects will benefit from the extra CPU horsepower, if you can unlock its full potential.
My take:This board is an absolute bargain! That said, if you really want an Alder Lake-N that's done right, and you don't mind a bit larger (almost NUC) form factor, then the ODROID-H4 series is the one you want. It's still a good value for money (better than Latte Panda Mu), but supports a full 80 mm M.2 drive (with more lanes than the Latte Panda), supports up to 48 GB DDR5, and lets you choose between the more powerful N97 and N305 SoCs.
The only two things lacking from the ODROID-H4 are built-in wifi and GPIO. Both can be added via USB, however. Another big plus is that it has a mini-ITX adapter kit, enabling you to put it in any mini-ITX case (but not Thin mini-ITX!). The plus and ultra versions also have SATA x4. -
bit_user
I think it's not the power, but the cooling that's inadequate.Devoteicon said:ConsNeeds more power
That said, it'd be interesting to know how the manufacturer configured the PL1/PL2 values. I assume PL1 is the standard 6W, but the real question is what they did with PL2. The stock value for that SoC is 25 W, which shows just how hungry those little Gracemont cores + the 24 EU iGPU can get!
The good news is that you should be able to hack a decent heatsink onto it and use the script mentioned in my above post to make it perform as good as any N100 out there! Just be sure to use either a copper heatsink or at least a copper shim between the die and the heatsink, if it's not copper. The tiny Gracemont cores have major hotspotting issues, especially for single-core and dual-core boosting scenarios. -
bit_user
As far as I'm aware, disabling C-state will not raise the upper limits. I'm pretty sure C-state is just about letting the CPU clock down, during periods of low-load. If you disable C-state, then it will limit your ability to boost, because that will increase the average power consumption and eat into the headroom normally used by the turbo algorithm.ryann said:c-state enable will make cpu running over 3Ghz. just make sure CPU cooling has margin.
As far as I'm aware, there's simply to way to exceed the baked-in core vs. frequency limits of the SoC. That counts as overclocking and these SoCs don't allow it.
The performance discrepancy vs. the Latte Panda Mu comes down to one of two things (or both):ryann said:It can be regarded as an additional performance, of course it also requires more power.
lower default PL2
thermal throttling
Although they showed thermal data for the "stress test", they didn't show it for Geek bench, which is where we saw the big performance discrepancy. That's a shame, especially because Geek Bench was almost certainly more stressful than what they had tried to do for a "stress test".