For the last four years the cheap end of the mini PC market has been Alder Lake-N: the N95, the N100, the N305, all of them stacks of Gracemont efficiency cores with no performance core anywhere in the design. They were fine. They were also uniformly frustrating in the same way, because four small cores handle background work well and handle *you* poorly. Anything interactive, a cold app launch, a heavy page, a single-threaded build step, ran at the speed of one small core and no faster.
The Core 3 304 is the first budget Intel part I have tested that breaks that pattern, and the way it breaks it is worth the review on its own. It is Wildcat Lake, Intel's new low-power SoC on the 18A process, and it is not a pile of E-cores. It is one Cougar Cove performance core running up to 4.3 GHz with four Darkmont efficiency cores behind it. Five cores, five threads, an odd-looking configuration that turns out to be exactly the right shape for what these machines are actually asked to do.
System Configuration
- Processor: Intel Core 3 304 (Wildcat Lake), 5 cores / 5 threads: 1x Cougar Cove P-core at 1.5 GHz base and 4.3 GHz turbo, 4x Darkmont LP E-cores at 1.4 GHz base and 3.3 GHz turbo. 4 MB L2, 6 MB L3. Intel 18A process. 15 W processor base power, 35 W maximum turbo power
- Memory: 12 GB LPDDR5, soldered, no SO-DIMM slots. Nameplate 5600 MT/s. Windows reports 11.7 GiB usable
- Graphics: Intel Graphics, 1x Xe3 core (16 EU), up to 2.3 GHz, 9 TOPS. PCI ID
8086:FD81 - NPU: Intel NPU 5, 15 TOPS INT8
- Storage: 512 GB NVMe SSD (CF620, no-name controller and NAND)
- Networking: Realtek RTL8125 2.5 GbE, Realtek RTL8852BE Wi-Fi 6 (2.4/5 GHz, not 6E)
- BIOS: American Megatrends
P3WL21001, dated 2026-06-26 - Operating System: Windows 11 Pro, build 10.0.26200
- Power Consumption: 5.6 W idle, 37.9 W peak at the wall (ESP32 meter)
The Processor in Context
Wildcat Lake launched in May 2026 as the replacement for the Alder Lake-N and Twin Lake families that have carried Intel's budget mini PC business since 2023. Two things changed, and only one of them is the process node.
The interesting change is the topology. Alder Lake-N was E-cores only: the N100 is four Gracemont cores, the N305 is eight. There was no P-core in the design at any SKU. Wildcat Lake introduces a hybrid layout at this price tier, pairing a single Cougar Cove P-core with four Darkmont LP E-cores. That is why this chip enumerates as five cores and five threads, a count that looks like a typo until you understand the design. There is no hyperthreading on the P-core, so cores and threads match.
The second change is the process. 18A is Intel's leading node, and seeing it under a budget SKU this early is unusual. It shows up in the power numbers later in this review, which are the best I have measured in this review set.
The specification that matters most for interpreting every number below is the power envelope: 15 W base, 35 W maximum turbo. This is a genuinely low-power part.
How much faster is the P-core, really
I can answer this precisely, because I accidentally measured both. An early Geekbench run on this machine was scheduled entirely onto the E-cores (the methodology note at the end explains how). Same binary, same machine, same session, same thermal state, one run on Darkmont only and one with the P-core available:
| Geekbench 6 | Darkmont E-cores only | With the Cougar Cove P-core | Difference |
|---|---|---|---|
| Single-Core | 1526 | 2396 | +57.0% |
| Multi-Core | 4618 | 5755 | +24.6% |
A 57% single-thread gap between the two core types on one die is the whole argument for this design in one number. An N100 owner is living on the left column permanently. That is what a P-core buys, and it is why this chip feels different from the Alder Lake-N machines it replaces even though it has fewer cores than an N305.
Port Layout and Connectivity
Front:
- 2x USB-A 3.2 Gen 2 (10 Gbps)
- 1x USB-C with DisplayPort Alt Mode
- 3.5mm combo audio jack
Rear:
- 4x USB-A 3.2 Gen 1 (5 Gbps)
- 1x HDMI
- 1x DisplayPort
- 1x 2.5 GbE RJ-45
- DC input
Seven USB ports plus dual digital display outputs and 2.5 GbE is a generous layout at this tier, and it is physically identical to the port arrangement on the Kamrui Hyper H1 I reviewed previously. That is almost certainly the same chassis and the same board family carried across to a new SoC.
A sidebar on how those ports are wired
The split above is worth explaining, because it tells you which ports to use for fast external storage.
Intel's specification for the Core 3 304 provides 2x USB 3.2, 8x USB 2.0, and up to 2x Thunderbolt 4. Windows enumeration on this unit shows a Genesys Logic GL3523 hub (VID_05E3&PID_0626), a four-port USB 3.1 Gen 1 part, meaning 5 Gbps shared across its downstream ports.
Those two facts explain the layout exactly. The SoC's two native SuperSpeed lanes go to the two front USB-A ports at 10 Gbps, and the four rear ports hang off the GL3523 at 5 Gbps. So the front pair is where an NVMe enclosure belongs, and the rear four are for peripherals that live plugged in. Note also that the four rear ports share the hub's single upstream link, so running several fast drives off the back simultaneously will contend in a way the front ports do not.
Unlike the previous-generation Hyper H1 in this same chassis, which offered a 10 Gbps USB-C, this unit's USB-C is not fed by one of those two native lanes; it carries DisplayPort Alt Mode on the display lanes. Treat it as a display output first and a data port second until its data rate is measured.
Windows cannot report the maximum speed of an empty port, so if you want to verify any of this, plug a USB 3.2 Gen 2 drive into each port and watch the sequential rate: above roughly 470 MB/s indicates a 10 Gbps link, around 440 MB/s indicates 5 Gbps, and around 40 MB/s means it fell back to USB 2.0.
Expandability and Internal Storage Options
The RAM is soldered LPDDR5 and there are no SO-DIMM slots. 12 GB is permanent.
That number deserves an honest treatment rather than a reflexive complaint. 12 GB runs Windows 11 comfortably for the workloads this machine suits: browser, office applications, media, light development. It is more than the 8 GB that the N100 boxes shipped with for years, and it is enough that you will hit the CPU ceiling before the memory ceiling in most realistic use. Under Linux, where the desktop baseline is several gigabytes lower, it is genuinely roomy.
What you cannot do is change your mind later. If you want to run several VMs, keep a large local language model resident, or work with big datasets, this is a hard ceiling and you should buy something else. The trade is the usual one: soldering LPDDR5 close to the package is part of what buys the idle power figures below.
The 512 GB NVMe is in an M.2 2280 slot and is replaceable. The CF620 is not a publicly documented SSD SKU and carries no advertised TBW rating, so treat it as commodity storage. As the storage section shows, it is the weakest measured component in the system, and it is also the one component you can trivially upgrade.
Performance Analysis
Every number below was measured on the unit with a scripted benchmark harness, with wall power sampled at 1 Hz by an ESP32 meter throughout. The comparison unit is the Kamrui Hyper H1 (Ryzen 7 7735HS, 8 cores / 16 threads at 28 W), because it is the same chassis at roughly twice the price class.
CPU Performance
| Benchmark | Kamrui P2 (Core 3 304) | Hyper H1 (7735HS) | Delta |
|---|---|---|---|
| Geekbench 6 Single | 2396 | 1966 | +21.9% |
| Geekbench 6 Multi | 5755 | 8755 | -34.3% |
| y-cruncher Pi-1B | 102.4 s | 58.8 s | 1.74x longer |
| y-cruncher multi-core efficiency | 89.1% | 96.8% | -7.7 pts |
| 7-Zip compress | 5,672 MIPS | 5,843 MIPS | -2.9% |
| 7-Zip decompress | 9,025 MIPS | 9,293 MIPS | -2.9% |
| Blender Classroom CPU | 18.15 spm | 37.31 spm | -51.4% |
The single-core score is the headline. At 2396 the Core 3 304 beats the Ryzen 5 7430U by 22.7%, the Ryzen 7 7735HS by 21.9%, and the i9-11900H by 7.7%. It lands 7.0% behind the i5-14450HX and 9.4% behind the i9-13900HK, a chip in a $749 machine. From a 15 W part.
That is the practical argument for this design. Single-thread performance is what determines whether a machine feels quick, and this one is within 10% of parts that cost three times as much and draw four times the power.
The multi-core score is the price of admission. 5755 is last in the review set, 11% below even the six-core 7430U and 34.3% below the 7735HS. Five cores are five cores, and no amount of architectural cleverness changes that.
Two results complicate the simple story in useful ways. y-cruncher's Pi-1B takes 1.74x as long as on the 7735HS, which is roughly what the core counts predict, and its multi-core efficiency of 89.1% is only 7.7 points behind the eight-core Ryzen. The hybrid topology scales better than I expected. And 7-Zip lands within 2.9% of the 7735HS on both compression and decompression, which is a genuinely surprising result: an eight-core, sixteen-thread chip at 28 W has essentially no advantage here over five cores at 15 W. 7-Zip's dictionary work is memory-latency-bound and branch-heavy rather than purely throughput-bound, and this platform's memory subsystem is strong (see below).
Blender is the counterexample and the honest one: sustained, embarrassingly parallel floating-point rendering is exactly where core count is destiny, and the Core 3 304 gives up 51.4%.
Graphics Performance
| Benchmark | Kamrui P2 (Xe3, 16 EU) | Hyper H1 (Radeon 680M, 12 CU) | Delta |
|---|---|---|---|
| Geekbench 6 OpenCL | 6,347 | 26,009 | -75.6% |
| Geekbench 6 Vulkan | 7,428 | not measured | - |
| Blender Classroom GPU | 24.39 spm (oneAPI) | 59.54 spm (HIP) | -59.0% |
| Blender GPU vs own CPU | 1.34x faster | 1.60x faster | - |
| Heaven Basic | 40.1 FPS / 1010 | 100.5 FPS / 2533 | -60.1% |
Against the Intel iGPU it most directly replaces, this is a clear generational win: 6347 in OpenCL beats the i5-14450HX's UHD Graphics (4027) by 57.6%. Against everything else it loses, trailing the 11900H's UHD Xe by 13.5%, the Vega 7 by 25.0%, and the Radeon 680M by 75.6%.
Heaven tells the same story from a rendering rather than compute angle. At 40.1 FPS and a score of 1010 it beats the i5-14450HX's UHD Graphics by 49.6%, edges past the 7430U's single-channel Vega 7 by 7.5%, and trails the i9-11900H's UHD Xe by 15.0%, the i9-13900HK's Iris Xe by 41.0% and the Radeon 680M by 60.1%. That ordering matches the OpenCL ranking closely (Heaven has it 15.0% behind the 11900H where OpenCL says 13.5%), which is a useful sign that neither result is an artefact.
Those numbers were taken at Heaven's Basic settings: Direct3D9, 1280x720 windowed, 2x anti-aliasing, quality Medium, tessellation disabled. Worth stating explicitly, because Heaven's score is meaningless without them and the free Basic Edition locks the options so every run in this review set shares the configuration.
Two further details. First, Vulkan at 7428 is 17.0% above the OpenCL score on identical hardware. Intel's Vulkan compute path is better optimised than their OpenCL runtime, and since most iGPU comparisons quote OpenCL only, Intel parts are routinely undersold by that convention. Second, rendering Blender's Classroom scene on this iGPU is 1.34x faster than on its own CPU, so the oneAPI path is a working compute device rather than a checkbox, even though the absolute numbers are small.
The 9 TOPS from the iGPU and 15 TOPS from the NPU are not figures I have validated, but an NPU at this price point is new. The CPU also carries AVX-VNNI, which is the instruction set that matters for small-model inference on the cores themselves.
None of this makes it a gaming machine. It is 16 EUs.
Media Engine (transcoding)
For a machine whose best role is a home server, the fixed-function media engine matters more than the shader count, and this is where the Xe3 block earns its keep. All figures are frames per second on Jellyfin's own test clips, every one of them 60 fps, with the 4K pair being 10-bit HDR10.
| Workload | fps | vs realtime (60 fps source) |
|---|---|---|
| 1080p H.264 encode (QSV) | 401 | 6.7x |
| 1080p AV1 encode (QSV) | 395 | 6.6x |
| 1080p HEVC encode (QSV) | 379 | 6.3x |
| 1080p x264 encode (CPU, veryfast) | 113 | 1.9x |
| 4K AV1 HDR10 decode (QSV) | 484 | 8.1x |
| 4K HEVC HDR10 decode (QSV) | 446 | 7.4x |
| 4K HDR10 to 1080p SDR, tone-mapped (QSV) | 215 | 3.6x |
Three things stand out.
AV1 encode exists at all. This is the first machine in the review set with hardware AV1 encode, and it runs at 395 fps, within 2% of its H.264 speed. AV1 delivers meaningfully better quality per bit than H.264, so a machine that can encode it in fixed-function hardware at 6.6x realtime is genuinely useful for archiving a library or re-encoding camera footage.
The hardware is 3.5x the CPU, and that understates it. x264 at the veryfast preset manages 113 fps, so the media engine is 3.5x quicker in raw throughput while producing a better-compressed format. The power difference is the real story, below.
The tone-mapping number is the one to judge it on. Taking a 4K HDR10 HEVC stream, decoding it, tone-mapping to SDR and re-encoding to 1080p H.264 is the workload that makes Plex and Jellyfin servers stutter, and this machine sustains 215 fps, about 3.6x realtime. That is roughly three simultaneous 4K HDR transcodes, or many more of the 24 fps film content most libraries actually hold.
And it does it on almost no power. Measured at the wall over a sustained 64 second run: 10.6 W average, 12.1 W peak, against a 5.6 W idle floor. Call it six watts to transcode 4K HDR in realtime three times over. For comparison, the CPU-bound phases of this review pushed the same machine to 37.9 W. This is what a fixed-function media block is for, and it is the single strongest argument for this machine as an always-on server.
One caveat on the AV1 decode figure. 484 fps is measured decoding to null, which skips the presentation path, so treat it as the engine's throughput rather than a playback number.
Memory Performance
This is where the Wildcat Lake platform quietly does something the N100 generation could not.
| Benchmark | Kamrui P2 (LPDDR5) | Hyper H1 (LPDDR5-5500, 64-bit) | Delta |
|---|---|---|---|
| MLC peak bandwidth | 40,698 MB/s | 35,741 MB/s | +13.9% |
| MLC idle latency | 128.8 ns | 124.1 ns | +3.8% (worse) |
| STREAM Copy | 38,552 MB/s | - | - |
| STREAM Scale | 25,968 MB/s | - | - |
| STREAM Add | 29,555 MB/s | - | - |
| STREAM Triad | 29,475 MB/s | 25,005 MB/s | +17.9% |
The comparison that matters for a buyer coming from this segment is against what these machines used to ship with: a single SO-DIMM of DDR4-3200, 25.6 GB/s theoretical and realistically 18 to 20 GB/s measured. Against that baseline, 40.7 GB/s of MLC bandwidth and 29.5 GB/s of STREAM Triad is roughly double, and it is the single biggest platform improvement after the P-core. It also helps explain the 7-Zip result above.
Be precise about the ceiling of that claim. Against dual-channel DDR4-3200, which the DDR4 mini PCs elsewhere in this review set run at roughly 40 GB/s measured, this is parity rather than a win. The LPDDR5 nameplate is not magic. What it beats is the single-channel configuration budget vendors actually shipped, and it beats that decisively.
Two caveats. Latency at 128.8 ns is mediocre, slightly worse than the Hyper H1's 124.1 ns and well behind a good DDR4 platform's 80 to 95 ns. LPDDR5 buys bandwidth and power, not latency. And I cannot confirm the bus width without a teardown: at a nameplate 5600 MT/s, 40.7 GB/s implies either a 64-bit bus running at an implausibly high 91% of theoretical or a wider bus at ordinary efficiency. Geekbench reports memory size but not speed or channel count on Windows, so this needs HWiNFO to settle.
For anyone reproducing these numbers: STREAM Copy (38,552 MB/s) sits far above Scale (25,968 MB/s) because the compiler turns the Copy kernel into a memcpy with non-temporal stores. Triad is the figure to compare across systems.
Storage Performance
| Benchmark (diskspd) | Kamrui P2 (CF620 512 GB) | Hyper H1 (G932E1Q 1 TB) | Delta |
|---|---|---|---|
| SEQ Read | 1,691 MB/s | 3,365 MB/s | -49.7% |
| SEQ Write | 1,599 MB/s | 2,254 MB/s | -29.1% |
| RND 4K Q32 Read | 380.9 MB/s (97,509 IOPS) | 362.8 MB/s (92,882 IOPS) | +5.0% |
| RND 4K Q32 Write | 294.4 MB/s (75,364 IOPS) | 264.8 MB/s (67,785 IOPS) | +11.2% |
| RND 4K Q1 Read | 45.9 MB/s (11,737 IOPS) | 68.2 MB/s (17,457 IOPS) | -32.8% |
| RND 4K Q1 Write | 120.8 MB/s (30,927 IOPS) | 122.5 MB/s (31,366 IOPS) | -1.4% |
Measured with Microsoft diskspd rather than CrystalDiskMark, so read these as sustained rates from a well-tuned benchmark rather than the CDM figures reviews usually quote.
Sequential read at 1,691 MB/s is roughly half the Hyper H1's drive. Some of that is the drive being a cheaper part, and some is the platform: the Core 3 304 provides six PCIe Gen4 lanes in total, so a budget board has less to spend on the M.2 slot than a Rembrandt design does.
The high-queue-depth random numbers are actually *better* than the more expensive unit's, by 5.0% on read and 11.2% on write. The Q1 read figure of 45.9 MB/s is the one that touches perceived snappiness and it trails by 32.8%. This is a competent commodity drive in a machine where it is the easiest component to replace.
Network Performance
2.5 GbE via the Realtek RTL8125, plus Realtek RTL8852BE Wi-Fi 6.
I did not run a wired throughput test, and I would rather say so than publish a number I did not take. The RTL8125 is thoroughly mature silicon and there is no realistic scenario in which it fails to saturate its link. The expected figure is roughly 2,389 Mbps of TCP goodput, which is 2.5 Gbps line rate after Ethernet, IP and TCP framing overhead (2500 x 1420 / 1486).
What I did measure, inadvertently, is the Wi-Fi. The unit's Ethernet port was unplugged during the benchmark run, so iperf3 measured the RTL8852BE instead: 440 Mbps down and 511 Mbps up against a LAN server on a 1.2 Gbps negotiated link. That is unremarkable Wi-Fi 6 performance. The card is a Wi-Fi 6 part with no 6 GHz band, and it is an M.2 2230 module you can replace.
Power Consumption
This is the best result in the review.
At 5.6 W idle this is the most efficient machine I have measured, edging out the 7430U's 6.0 W and using 31.6% less than the Hyper H1's 8.2 W. Running 24/7 at $0.15/kWh that is $7.40 per year in electricity, against $10.81 for the Hyper H1 and $18.66 for the i9-13900HK.
Peak draw was 37.9 W at the wall during Blender CPU, with y-cruncher a hair behind at 37.8 W and sustaining a 31.6 W average. Those are whole-system numbers including power supply losses, so the SoC itself was drawing meaningfully less. After conversion losses the machine is pulling roughly 32 W DC at its hardest, which puts the processor in the neighbourhood of its 35 W turbo ceiling only in bursts and much nearer the 15 W base power in sustained work.
I was not able to read CPU package power or die temperature directly. LibreHardwareMonitor exposed only CPU load, GPU clock and SSD temperature on this machine, most likely because Wildcat Lake is new enough that its sensor definitions are not yet implemented. The wall meter is the honest number and it is the one that shows on your electricity bill.
Noise
The unit stayed quiet for the entire benchmark run, including sustained y-cruncher and Blender CPU phases at 29 to 38 W. There is no fan ramp to speak of, no pitch change under load, and no throttling cycle. On a desk next to a monitor it is effectively inaudible at idle and only a soft baseline hum at full load.
This is the direct consequence of the power envelope. A machine that never dissipates more than about 32 W sustained does not need to move much air, and Kamrui reused a chassis and cooler designed for a 28 W eight-core part. The cooling solution is comfortably oversized for this SoC, and the acoustic result is the best in the review set.
Use Case Recommendations
Home Server. Excellent, and the best argument for this machine. 5.6 W idle is the lowest I have measured, 2.5 GbE keeps NAS traffic off the bottleneck, and the acoustics let it live in a closet or behind a TV. The media engine seals it: 4K HDR10 tone-mapped to 1080p at 3.6x realtime for about six watts above idle, plus hardware AV1 encode, which no other machine in this review set has. As a Plex or Jellyfin box this punches far above its CPU. The 12 GB RAM ceiling is the constraint: containers and light services fit comfortably, several memory-hungry VMs do not.
Productivity Workstation. Genuinely good, and this is where the P-core earns its place. A single-core score within 10% of an i9-13900HK means browsing, office work and general desktop interaction feel quick in a way no N100 machine manages. 12 GB is adequate rather than generous.
Development Environment. Adequate with caveats. Interactive work (editing, language servers, incremental rebuilds) benefits from the strong single-thread result. Full parallel compiles do not: five cores at 34% below a 7735HS is the reality. For scripting, web development, or driving a remote build host, it is fine.
Light Content Creation. Limited. The Xe3 iGPU renders Blender 1.34x faster than the CPU, which is a working compute path, but the absolute numbers are small. Photo editing works. Video editing beyond 1080p does not.
Gaming. No. 16 execution units is enough for emulation of older consoles and for indie titles, and that is the honest boundary.
Linux. Probably the best fit of all. 12 GB goes further under a Linux desktop than under Windows 11, the power floor should be as good or better, and none of the weak points of this machine matter for a headless server role.
Value Proposition
At $360.99 the Kamrui P2 sits $20.99 (6.2%) above the Kron K1 at $340 and $129 (26.3%) below the Hyper H1 at $489.99 in the same chassis. That places it at the bottom of this review set on price, next to the K1, and the per-dollar numbers split cleanly along the same line as everything else in this review.
| Per dollar | Kamrui P2 (Core 3 304, $360.99) | Kron K1 (7430U, $340) | Hyper H1 (7735HS, $489.99) |
|---|---|---|---|
| Geekbench 6 Single, pts/$ | 6.64 | 5.74 | 4.01 |
| Geekbench 6 Multi, pts/$ | 15.94 | 19.03 | 17.87 |
| Idle power | 5.6 W | 6.0 W | 8.2 W |
Single-thread performance per dollar is the best in the set: 15.5% ahead of the Kron K1 and 65.4% ahead of the Hyper H1. Multi-core per dollar is the worst: 16.2% behind the K1 and 10.8% behind the Hyper H1. If you are buying for responsiveness, this is the cheapest way into a fast core I have tested. If you are buying for throughput, it is not.
Against the Kron K1 ($340). The closest call. The K1 has six Zen 3 cores and twelve threads, so it wins multi-core by 12.4% and costs $21 less. The P2 wins single-thread by 22.7%, idles slightly lower, has a hardware media engine with AV1 encode, and doubles a single-channel DDR4 memory configuration on bandwidth. On graphics it is a split: it edges the K1's single-channel Vega 7 in Heaven by 7.5% but trails the same iGPU on dual-channel memory by 35.7%. For a desktop or a media server, the extra $21 buys the better machine. For a batch-compute box, the K1 is still the better buy.
Against the Hyper H1 ($489.99). The 7735HS is the stronger computer in every throughput metric: 52% more multi-core, two and a half times the Heaven frame rate, four times the OpenCL score. What it does not have is a faster core; the Core 3 304 beats it by 21.9% in single-thread. Electricity does not close the gap: the P2's 2.6 W lower idle saves about $3.40 a year at $0.15/kWh, which would take roughly 38 years to repay the $129 difference. Buy the Hyper H1 if you need the cores or the iGPU. Otherwise the cheaper machine is the more sensible one.
Against the Mac Mini M4 ($499). $138 (38.2%) more buys substantially more CPU and GPU in every measure, plus 16 GB of memory to this machine's 12 GB. The Mac's base configuration ships with soldered 256 GB storage and gigabit Ethernet, against a replaceable 512 GB NVMe and 2.5 GbE here, and it runs macOS rather than Windows or a stock Linux distribution. For a desktop, the Mac is worth the difference. For a headless home server that wants an x86 Linux host, a swappable drive and a faster wired port, the P2 is the more practical box at well under the price.
Against the Kamrui Hyper (~$500) and AceMagic M1 (~$749). Both are a class up: more cores, more power draw and, at 20 W and 14.2 W idle, two and a half to three and a half times the power floor. They are not alternatives for the same buyer.
Upgradeable vs Permanent
- Soldered and permanent: RAM (12 GB LPDDR5, forever), Ethernet NIC (Realtek RTL8125 2.5 GbE), CPU, iGPU, NPU
- Replaceable: NVMe SSD (M.2 2280), Wi-Fi module (M.2 2230, requires opening the chassis)
The 12 GB memory ceiling is the decision point for this machine, more than the price will be. Everything else that is weak is either replaceable (the CF620 SSD, the Wi-Fi 6 card) or irrelevant to the roles it suits. The RAM is neither.
The 2.5 GbE choice deserves credit rather than criticism. The RTL8125 is soldered and permanent, and it is the right part: budget vendors shipping gigabit in 2026 are cutting a corner that costs the buyer something, and Kamrui did not.
The comparison that matters
The real competition is not the other units in this review set, which are all more powerful and more expensive. It is the enormous installed base of N100 and N305 mini PCs this chip replaces. Against those:
- A single-thread result 57% above this machine's own E-cores, which is roughly the class of core an N100 owner has
- Roughly double the memory bandwidth, if they shipped single-channel DDR4-3200, which most did
- A lower idle power floor than any of them
- An iGPU that is a supported compute device rather than a display adapter
- Fewer total cores than an N305, and a lower multi-core ceiling
For anything interactive, this is the better machine by a wide margin. For a parallel batch queue, the N305 still has an argument.
Software Experience
Windows 11 Out-of-Box Experience
Windows 11 Pro build 26200 was preinstalled and setup was uneventful, with none of the multi-hour update pathology the i5-14450HX unit exhibited. The SMBIOS manufacturer and model fields are both literally Default String, so Windows reports the system model as "Default String" in About and any inventory tooling does the same. Cosmetic, but sloppy, and it is the second Kamrui unit to ship this way.
Linux Compatibility (not tested)
Not tested in the review window. The Realtek RTL8125 and RTL8852BE are both well supported on current mainline kernels. Xe3 graphics support is the open question given how new the silicon is.
A note on methodology
One finding from this review is worth passing on, because it would silently corrupt anyone else's numbers on a hybrid chip.
My benchmark harness drives the machine over SSH. On Windows, an SSH command runs in Session 0 as service-context work, and Windows Thread Director treats that as background priority, which prefers E-cores. On a homogeneous CPU that costs nothing. On this 1P+4E part it meant benchmarks ran without ever touching the performance core. Geekbench's own hardware enumeration gave it away: the affected runs reported 32 KB L1 data cache across five cores (Darkmont's geometry) and could not see the 6 MB L3 at all, while a correctly scheduled run reported 48 KB L1d on the Cougar Cove core plus the full L3.
The scale of the error, same machine, same binaries:
| Measurement | Scheduled onto E-cores | Correctly scheduled | Understated by |
|---|---|---|---|
| Geekbench 6 Single | 1526 | 2396 | 36% |
| Geekbench 6 Multi | 4618 | 5755 | 20% |
| y-cruncher Pi-1B | 145.5 s | 102.4 s | 30% slower |
| 7-Zip decompress | 5,967 MIPS | 9,025 MIPS | 34% |
| Blender Classroom CPU | 13.73 spm | 18.15 spm | 24% |
| Blender CPU peak power | 36.7 W | 37.9 W | - |
Affinity is not the mechanism and pinning does not fix it: the affinity mask already covers every core. It is a scheduling preference, and the fix is to set the process priority class explicitly on the benchmark process and every one of its descendants. Memory bandwidth, storage and GPU results were unaffected, which is a useful sanity check on the diagnosis: only CPU-throughput-bound work moved.
If you benchmark a hybrid machine remotely, check which cores your work actually landed on before trusting the numbers.
Conclusion
The Core 3 304 is a more interesting chip than its position in Intel's stack suggests, and the reason is structural. Four years of budget mini PCs were built on E-core-only silicon that handled background work well and interactive work poorly. Putting one real performance core into that design fixes precisely the thing that made those machines annoying, and it does so without giving up the power envelope that made them attractive. A Geekbench single-core score of 2396 puts this 15 W machine within 10% of an i9-13900HK, while idling at 5.6 W, the lowest figure in this review set.
The costs are real. Multi-core throughput is last place, 34% behind the 7735HS and below even a six-core 7430U, so anything that scales across many threads belongs on a different machine. Blender gives up 51%. The memory subsystem doubles what single-channel DDR4-3200 budget boxes delivered but only reaches parity with dual-channel DDR4, and 128.8 ns latency is mediocre. The SoC has only two SuperSpeed USB lanes, both spent on the front USB-A pair, which leaves the USB-C's data rate unaccounted for. And 12 GB of soldered LPDDR5 is a ceiling you will never raise.
What Kamrui built around the chip is coherent. The cooler is oversized for a 15 W part, which is why the machine is silent. The 2.5 GbE is the right port. The SSD is the weak component and it is the one you can replace. The RAM is the constraint you cannot fix, and 12 GB is the number to think hardest about before buying.
Final recommendation: at $360.99, buy this if you want a silent, genuinely low-power always-on machine for a home server, a media box, or a responsive everyday desktop, and you are confident 12 GB will always be enough. Buy an N305 box instead if your workload is a parallel batch queue. Buy the Hyper H1 or a larger machine if you need multi-core throughput or a capable iGPU. If you plan to run Linux on it, this is a better machine than its Windows numbers suggest.
