Honestly, I want to like the Geekom A9 Max, I really do. From a pure design perspective, the hardware, the engineering behind it, the connectivity bundled into this thing, it’s on another level. And it’s just a bloody pretty thing at that, at least as far as Mini PCs go. You’d think it’d be a shoo-in for our best SFF PC list, without question. But there’s a catch, because of course there is.
I may as well rip the band-aid off right now. It’s the memory. I can already see the PR team squirming as they read that sentence, and it genuinely does pain me to say it, I promise. Geekom has decided to ship the A9 Max AI with a single stick of 32 GB DDR5 @ 5600 MT/s. And that is a decision that has effectively starved AMD’s Radeon 890M of any modicum of the performance it can actually generate. I know this because I have extensively tested this hardware most recently, in Acemagic’s Mini PC Retro X5, which, like-for-like, practically has the same spec, bar a smaller SSD, and a slightly older CPU.
The performance impact that this one tiny decision has had on the overall architecture is nuts, and combined with the wildly high price tag being what it is, it’s by far the thing that’s decimated that review score. If Geekom turned around tomorrow and popped 2x 16 GB sticks in this config instead of this single-channel nonsense, I can happily tell you right now, the score would bounce up by a good 15%, because in almost every other way, the A9 Max is excellent.
A9 Max AI specs
CPU
AMD Ryzen AI 9 HX 470
Cores | Threads
12 | 24
Clock speed
up to 5.2GHz
GPU
AMD Radeon 890M
RAM
32GB (1x32GB) DDR5-5600
Storage
2 TB Kingston OM8TAP42048K1-A00 PCIe 4.0 (2x PCIe 4.0 M.2 slots available)
Rear I/O
2x 2.5Gb ethernet, 1x USB 2.0 Type-A, 1x USB 3.2 Gen 2 Type-A, 2x USB 4.0 Type-C, 2x HDMI 2.1, DC-in
Front I/O
4x USB 3.2 Gen 2 Type-A, SD Card Reader, 3.5 mm audio out
Power
120 W
Dimensions
135 x 132 x 46mm
Price
£1,600 | $1,799 (as configured)
Buy if…
✅ You need a compact AI-focused workhorse: The A9 Max has some serious engineering chops. Chuck an extra DRAM stick in here and an SSD, and you’re looking at one mean local LLM solution or workstation.
Don’t buy if…
❌ You’re a gamer: Without a second stick of RAM by default, the A9 Max lacks the memory bandwidth to compete with machines half its cost.
And, breathe. I’ll get back to that in a minute. Still, for £1,600 in the UK or $1,800 US, it’s generally not a bad offering (memory aside). You get that 32 GB RAM capacity, a 2TB Kingston PCIe 4.0 SSD (it’s not the best), AMD’s Ryzen AI 9 HX 470 CPU, A Radeon 890M iGPU, and a hell of a lot of connectivity.
The chip itself is relatively new on the scene, only really landing with us in January of this year. Though it’s effectively a refreshed Strix Point (HX 370 series) chip, still built out of the same silicon architecture as its predecessor. You get those four Zen 5 cores, and eight Zen 5c cores for a fat 12/24 multi-threaded config, and it’s all still built on the back of TSMC’s 4 nm process as well. What’s different? Well… clock speeds have been bumped ever so slightly up (with an extra 100 MHz on boost), oh, and the “biggest” update comes in the form of the NPU, which has seen its performance flick up to 86 TOPS versus its predecessor’s 80 (huzzah, our AI datacenter worries are finally solved).
Connectivity is wild, too. There’s so much here if you’re looking for a lightweight workstation that’s not brimming with a dedicated GPU and full-fat processor. There’s no less than five USB 3.2 Gen2 Type-A, two USB 4.0 Type-Cs, twin 2.5 Gbe, along with support for twin HDMI 2.1 ports, as well as the other bells and whistles you’d expect. No Thunderbolt, of course, but you do get WiFi 7 and Bluetooth 5.4.
Alright, I’m going to give you some numbers now, and they’re quite clearly bad. But, it’s important that we all collectively have context on this before I act like a wailing harpie and impale this memory situation into the dirt with my shiny shiny words (honestly, it’s a miracle I get paid for this).
So, we test our SFF PCs (with iGPUs) at 1080p on the medium graphical preset. For this little wee dissection, I’m going to be using that Acemagic X5 I mentioned earlier for comparison. It has the same Radeon GPU, the same memory speed, memory capacity, a better SSD (albeit smaller), and practically the same CPU for what it’s worth (technically, the A9 Max has a better chip here). On average, across all our test titles, with and without upscaling, across five separate games, separate engines, separate developers, the A9 Max landed just 23.1 fps. The Retro X5, with its dual-channel memory, 47.6 fps.
Future
Future
Future
Previous
1 / 3
Next
1080p gaming performance
Black Myth Wukong (1080p Medium)Cyberpunk 2077 (1080p Medium)F1 24 (1080p Medium)Horizon Zero Dawn (1080p Original)Metro Exodus Enhanced (1080p High)
Avg FPS
1% Low FPS
Geekom A9 Max
9
7
Acemagic Retro X5
20
16
Black Myth Wukong (1080p Medium) Data
Product
Value
Geekom A9 Max
9 Avg FPS, 7 1% Low FPS
Acemagic Retro X5
20 Avg FPS, 16 1% Low FPS
Avg FPS
1% Low FPS
Geekom A9 Max
22
15
Acemagic Retro X5
38
22
Cyberpunk 2077 (1080p Medium) Data
Product
Value
Geekom A9 Max
22 Avg FPS, 15 1% Low FPS
Acemagic Retro X5
38 Avg FPS, 22 1% Low FPS
Avg FPS
1% Low FPS
Geekom A9 Max
30
24
Acemagic Retro X5
73
59
F1 24 (1080p Medium) Data
Product
Value
Geekom A9 Max
30 Avg FPS, 24 1% Low FPS
Acemagic Retro X5
73 Avg FPS, 59 1% Low FPS
Avg FPS
1% Low FPS
Geekom A9 Max
23
13
Acemagic Retro X5
43
32
Horizon Zero Dawn (1080p Original) Data
Product
Value
Geekom A9 Max
23 Avg FPS, 13 1% Low FPS
Acemagic Retro X5
43 Avg FPS, 32 1% Low FPS
Avg FPS
1% Low FPS
Geekom A9 Max
14
5
Acemagic Retro X5
23
8
Metro Exodus Enhanced (1080p High) Data
Product
Value
Geekom A9 Max
14 Avg FPS, 5 1% Low FPS
Acemagic Retro X5
23 Avg FPS, 8 1% Low FPS
You can clearly see how a memory bandwidth starved iGPU is hobbled by having single channel memory here.
1080p upscaled gaming performance
Black Myth Wukong | Upscaling (1080p Medium)Cyberpunk 2077 | Upscaling (1080p Medium)F1 24 | Upscaling (1080p Medium)Horizon Zero Dawn | Upscaling (1080p Original)
Avg FPS
1% Low FPS
Geekom A9 Max
14
11
Acemagic Retro X5
30
22
Black Myth Wukong | Upscaling (1080p Medium) Data
Product
Value
Geekom A9 Max
14 Avg FPS, 11 1% Low FPS
Acemagic Retro X5
30 Avg FPS, 22 1% Low FPS
Avg FPS
1% Low FPS
Geekom A9 Max
30
20
Acemagic Retro X5
52
27
Cyberpunk 2077 | Upscaling (1080p Medium) Data
Product
Value
Geekom A9 Max
30 Avg FPS, 20 1% Low FPS
Acemagic Retro X5
52 Avg FPS, 27 1% Low FPS
Avg FPS
1% Low FPS
Geekom A9 Max
30
26
Acemagic Retro X5
86
75
F1 24 | Upscaling (1080p Medium) Data
Product
Value
Geekom A9 Max
30 Avg FPS, 26 1% Low FPS
Acemagic Retro X5
86 Avg FPS, 75 1% Low FPS
Avg FPS
1% Low FPS
Geekom A9 Max
36
21
Acemagic Retro X5
63
44
Horizon Zero Dawn | Upscaling (1080p Original) Data
Product
Value
Geekom A9 Max
36 Avg FPS, 21 1% Low FPS
Acemagic Retro X5
63 Avg FPS, 44 1% Low FPS
The performance gap between the system with dual-channel memory and the one with single-channel memory doesn’t get any better with upscaling enabled.
Synthetic gaming performance
3DMark Time Spy3DMark StorageFFXIV Shadowbringers
Index score
GPU index score
CPU index score
Geekom A9 Max
2301
2035
8948
Acemagic Retro X5
3532
3153
11109
3DMark Time Spy Data
Product
Value
Geekom A9 Max
2301 Index score, 2035 GPU index score, 8948 CPU index score
Acemagic Retro X5
3532 Index score, 3153 GPU index score, 11109 CPU index score
Index score
Avg bandwidth (MB/s)
Access time (ųs)
Geekom A9 Max
1306
226.22
139
Acemagic Retro X5
2460
415.65
72
3DMark Storage Data
Product
Value
Geekom A9 Max
1306 Index score, 226.22 Avg bandwidth (MB/s), 139 Access time (ųs)
Acemagic Retro X5
2460 Index score, 415.65 Avg bandwidth (MB/s), 72 Access time (ųs)
Geekom A9 Max
38.96
Acemagic Retro X5
20.45
FFXIV Shadowbringers Data
Product
Value
Geekom A9 Max
38.96
Acemagic Retro X5
20.45
Contrary to what Geekom says, the synthetic performance difference is actually less than real-world in-game benchmarks.
It’s not a limitation that sits solely with gaming, either. In 7-Zip, compression and decompression were markedly slower, multi-core performance in CineBench was down too, and the only times it did beat out the HX 370 was in single-threaded tasks, where memory bandwidth is less of a thing. Not exactly a riveting score sheet then. Even 3DMark’s Storage benchmark came out significantly slower, which I can’t believe even for a second is entirely down to the drive being just that much sloppier than the EXT X200E found in the X5 is.
Previous
1 / 2
Next
Synthetic CPU performance
Cinebench R247-zip
Single-core index score
Multi-core index score
Geekom A9 Max
119
1101
Acemagic Retro X5
107
1169
Cinebench R24 Data
Product
Value
Geekom A9 Max
119 Single-core index score, 1101 Multi-core index score
Acemagic Retro X5
107 Single-core index score, 1169 Multi-core index score
Compression (GIPS)
Decompression (GIPS)
Geekom A9 Max
98.88
130.12
Acemagic Retro X5
110.45
131.4
7-zip Data
Product
Value
Geekom A9 Max
98.88 Compression (GIPS), 130.12 Decompression (GIPS)
Acemagic Retro X5
110.45 Compression (GIPS), 131.4 Decompression (GIPS)
This is likely where Geekom’s claims about limited performance difference ring true, as you can see both systems performing well.
Cooling performance
Thermal performance
Max CPU temp (°C)
Max CPU package power (W)
Geekom A9 Max
86
64.98
Acemagic Retro X5
85
69.98
Thermal performance Data
Product
Value
Geekom A9 Max
86 Max CPU temp (°C), 64.98 Max CPU package power (W)
Acemagic Retro X5
85 Max CPU temp (°C), 69.98 Max CPU package power (W)
This is likely where Geekom’s claims about limited performance difference ring true, as you can see both systems performing well.
Just to be doubly sure, I re-ran the whole benchmark suite, and triple-checked GPU frequencies, temps and power draw throughout that process too, noting down average frequencies as well (not typically something we report on for these machines). The X5 maxed out at 2,764 MHz and 2,100 MHz on average, at 60 degrees and with 50 W on max draw. The A9, 3,100 MHz at max, 2,300 MHz on average, with a top temp of 63 degrees and 55 W of power usage. By all accounts, then, the GPU itself was running at full tilt, but was just starved of memory bandwidth.
Geekom knows about this, too. In fact, I was pre-briefed that this was a decision it had made in advance. The justification being that it provides end-users with the option to expand their memory solution later down the line with an additional stick of 32 GB. Which, yeah, okay, I don’t necessarily disagree with.
Geekom also notes that, actually, the performance drops are only in synthetic benchmarks, and don’t affect real-world performance. Which I do disagree with.
Future
Future
Future
Future
You can quite clearly see that this does massively impact real-world results. In fact, more so than in the synthetic tests, by my calculations. What’s particularly frustrating about all this is that nowhere on the product pages (at least at time of writing) does it suggest that the A9 Max ships with a single stick. It’s all “dual-channel compatible”, “up to 128 GB”, or “32 GB RAM”, “High-speed memory”.
It’s a difficult thing to just accept, because the truth of it is the A9 Max is an outstanding piece of design. It’s cool, quiet, and it’s clear that the CPU in particular is seriously potent, particularly in localised AI applications if that’s your jam. The sad reality is, we’ve already seen what this hardware can do; it just came in a beige box at half the price.
