Linux PC build (2025)
Hello,
it’s me again. Some of you might remember me from this post, in which I was asking for feedback to build a Linux PC in 2025.
Stuff happened and I didn’t went through with it. So this still my first attempt at a build. Well now I’ve got time and want to try it again.
As you may notice, I’ve ditched the Z790-9 mother board in favor of a MSI PRO B650M-P. My dream of building a coreboot-system is officially dead, thus I decided to build an AMD-System.
Short Listing:
- CPU: AMD Ryzen 7 7700X 4.5 GHz 8-Core
- CPU-Cooler: Thermalright Assassin X 120 Refined SE 66.17 CFM CPU Cooler
- Mother board: MSI PRO B650M-P Micro ATX AM5 Motherboard
- Memory: Patriot Viper Venom 32 GB (2 x 16 GB) DDR5-6000 CL30 Memory
- Storage: Acer Predator GM7000 2 TB M.2-2280 PCIe 4.0 X4 NVME Solid State Drive
- GPU: Sapphire 21323-01-20G Radeon RX 7900 XT 20 GB Video Card
- Case: Zalman P10 MicroATX Mini Tower Case
- PSU: Thermaltake Toughpower GF1 (2024) 850 W 80+ Gold Certified Fully Modular ATX Power Supply
- Monitor: KOORUI 24E3 24.0" 1920 x 1080 165 Hz Monitor
If you notice anything wrong or have suggestions/improvements don’t hesitate to point them out.
Thanks in advance!!!
Specifications:
-
CPU: AMD Ryzen 7 7700X 4.5 GHz 8-Core Processor
- Boost Clock: 5.4 GHz
- TDP: 105 W
- L2 Chache: 8 MB
- L3 Chache: 32 MB
- Socket: AM5
- iGPU: Radeon
- Max. Memory: 128 GB
- Arch.: Zen 4
-
CPU-Cooler: Thermalright Assassin X 120 Refined SE 66.17 CFM CPU Cooler
- RPM: 1550
- Noise: 25.6 dB
- Height: 148 mm
- CPU-Socket: AM5 and others
-
Mother board: MSI PRO B650M-P Micro ATX AM5 Motherboard
- Chipset: AMD B650
- Memory:
- Type: DDR5
- Slots: 4
- Speed: DDR5-6000
- Max.: 128 GB
- PCIe x16 Slots: 1
- PCIe x1 Slots: 2
- M.2 Slots: 2x 2260/2280 M-key
- SATA 6.0 Gb/s Ports: 4
- Onboard Ethernet: 2.5 Gb/s Port (Realtek RTL8125BG)
- USB 2.0 Headers: 2
- USB 3.2 Gen 1 Headers: 1
- USB 3.2 Gen 2 Headers: 1
-
Memory: Patriot Viper Venom 32 GB (2 x 16 GB) DDR5-6000 CL30 Memory
- First Word Latency: 10 ns
- CAS Latency: 30
- Voltage: 1.35 V
- Timing: 30-40-40-76
-
Storage: Acer Predator GM7000 2 TB M.2-2280 PCIe 4.0 X4 NVME Solid State Drive
- Cache: 2048 MB
-
GPU: Sapphire 21323-01-20G Radeon RX 7900 XT 20 GB Video Card
- Memory: 20 GB
- Memory Type: GDDR6
- Core Clock: 2000 MHz
- Boost Clock: 2400 MHz
- Interface: PCIe x16
- Length: 276 mm
- TDP: 308 W
- Cooling: 3 Fans
- External Power: 2 x PCIe 8-pin
- DisplayPort: 2.1 Outputs 2
-
PSU: Thermaltake Toughpower GF1 (2024) 850 W 80+ Gold Certified Fully Modular ATX Power Supply
- Length: 140 mm
- Wattage: 850 W
- EPS 8-pin Connectors: 2
- PCIe 6+2-pin Connectors: 6
- SATA Connectors: 8
- AMP/Molex 4-pin Connectors: 4
-
Monitor: KOORUI 24E3 24.0" 1920 x 1080 165 Hz Monitor
- Size: 24"
- Resolution: 1920 x 1080
- Refresh Rate: 165 Hz
- Response Time: 1 ms
- Inputs:
- 1x HDMI 1.4
- 1x DisplayPort 1.2
What do you plan to do on that graphics card? Is that 20GB VRAM? That sounds nice, but not being a NVIDIA it lacks at least the CUDA cores which are necessary for many AI use cases which I have.
I think Ollama supported ROCM especially with the older AMD cards?
I have ollama running locally on my RX9070, I have to use kernel 6.14 since it’s such a new GPU.
The 16G VRAM means I can run decent models, faster than I can read… currently running gemma3:12b, it’s crazy fast.
How quantized? I don’t think 16GB of RAM is enough to run a full fat 12B model at FP16 but maybe I’m wrong.
Nvidia cards are just too expensive
q4_k_m
8.1GB file size
That’s great to know. Do you regret not having the performance of Nvidia? I would prefer not to buy green but I’m not sure from the comments about the delta in performance from ROCM vs CUDA
I don’t know, it is running much faster than I can read.
So I’m not sure why more performance would be needed, the only thing I was looking for was big VRAM, and AMD gives much more bang for you buck (especially in NZ). To get 16GB of VRAM on an NV card in NZ would have set me back an extra $800…not something I was willing to do.
Have you tried bigger models?
I tried Deepseek Coder 33b; it runs at around 2 words/second which is really slow.
Deepseek Coder V2 16b; seems to run as fast Gemma3
Tweakers just did some AI benchmarks for these cards https://tweakers.net/reviews/13150/ai-op-amd-en-nvidia-videokaarten-taalmodellen-upscaling-en-imagegeneration.html (you’ll need to translate it in your browser). The extra mem helps you fit bigger models, but performance is better for nvidia pretty much across the board.
I see. Thanks