My travel BeamCase - 2x 3090 1x A3000 in Sub 16L
I was working on this for the last 8 weeks because I wanted to upgrade my 6L PC to be capable of running local AI.
Since I was coming from ITX I had to tinker a bit as I wanted to support dual GPU for the inference and a third GPU for Helper Models like Embedding or similar tasks in my daily work.
Let's start with the Specs.
- Motherboard: MSI Wifi Edge Z890i
- CPU: Intel Core Ultra 245K
- Cooler: Thermalright Assassin Mini 120
- RAM: 64GB Fury Beast 5600
- GPU: 2 x DELL OEM 3090 24GB VRAM NVLInked, power limited to 250w each
- GPU: 1x RTX A3000 12GB VRAM
- SlimSas: Splitter x16 > x8/x8
- SlimSas: 2x Backplane x8/x8
- M2 to PCIE Adapter
- NVME: 2x 4TB Lexar Ares
- HDD: 3x 2TB Seagate Barracuda 2.5
- HDD: 1x 5TB Seagate Barracuda 2.5
- SSD: 1x 8TB Samsung Sata SSD,
- SSD: 1x 1TB Kingston Sata M2 in 2.5 Adapter
- PSU: Lian Li sp1000p
- FAN: 3x 120mm Thermalright
- FAN: 3x 92mm Noctua
I had to keep the case as small as possible as this is basically my travel rig. So I have taken the Slim ATX version of the BeamCase and cut out a Slot that I did not need. This allowed me to have the case a total frame height of 330mm. The width is 160mm and the length is 300mm which comes from the minimum inner frame length of 270mm dictated by the 2 RTX 3090, plus 2times 15mm from the frame.
I did not want to use 3d printed ones, so I vibe coded panels as svg for laser cutting acrylic sheet.
For the Mainboard and the SlimSas Backplanes I have modified the new Frame element and added mount points for the backplanes. For the A3000 which is actually a mobile GPU on a pcie card, I created a holder that I can mount behind the PSU.
I operate this case in Thailand and even when its 36 Degrees in my house, I still have plenty of thermal headroom even when I train Loras on both 3090's
Under Heavy load, the 3090s have around 80 degrees Core temp, 100 degrees Junction Temp, the CPU is at 80 degrees under full multi core load and the A3000 is at roughly 90 degrees. This load lasted for several hours during one of my stress tests and it was not going above, so I think this is the thermal fingerprint under heavy load.
All in all I am very happy with the density and the capabilities of running powerful local coding agents with local memory and having the possibility to travel with it.