2U and 4U rackmount builds, sized to what you actually run. Tell us your workload and we build it to match.
★ 5-star across 683 Google reviews since 2018|3-Year Hardware Warranty
See it
The two sizes, side by side
Same build quality, two heights. Switch between them to see what your workload actually needs.
Both. One question decides it: do you need a graphics card?
4U — for a discrete GPU, a tower cooler or a 360 mm AIO. Runs AI and LLM rigs and render nodes.
2U — for storage, networking and virtualization with no GPU. Runs a LAN-shop server, NAS or firewall, and fits more units per cabinet.2U. 89 mm tall, so you fit more machines per cabinet. Air-cooled with integrated graphics — no discrete GPU means no heat to fight and a quieter rack. This is the box for work that doesn't need a graphics card: a diskless boot server for a LAN shop, a NAS or file server, a pfSense firewall, a Plex media server transcoding on the iGPU, or a Proxmox node for your self-hosted apps.4U. 178 mm tall — room for a full-height graphics card, a tower air cooler, or a 360 mm liquid cooler that holds up under sustained load. This is the box when the work needs a GPU: a local AI or LLM rig on an RTX 5090, a render or workstation node, or a rackmounted gaming PC.
Choosing a size
2U or 4U: which one fits your build
It comes down to two things: how much cooling you need, and whether you need a discrete graphics card. A 4U has the height for big coolers, liquid cooling and full graphics cards. A 2U doesn't, so it stays with lighter work. We match the size to your workload so you don't pay for space you won't use.
2U89 mm tall · K245F
70 mm cooler ceilingLow-profile coolers, lower-power CPUs.
No room for a radiatorAir cooling only.
Integrated graphicsNo discrete graphics card.
Good for networking, NAS, firewalls and light virtualization.
4U178 mm tall · M445F-17
150 mm cooler clearanceFull tower coolers, higher-power CPUs.
Fits a 360 AIOLiquid cooling for sustained loads.
Full-height graphics cardsDiscrete GPUs up to 360 mm.
Built for AI, GPU compute, rendering and heavy virtualization.
Need a graphics card? Go 4U. If not, a 2U does the same job for less.
See the full spec comparison
Spec
2U · K245F
4U · M445F-17
Height
89 mm
178 mm
Motherboard
ATX, MATX, ITX
ATX, MATX, ITX
Discrete graphics
Integrated only
Multiple full-height, ≤360 mm
Expansion slots
3 full-height + 4 half-height
7 full-height
Liquid cooling
No
360 AIO (also 240 / 120)
CPU cooler height
≤70 mm
≤150 mm
Stock fans
2× 80 mm
2× 120 mm front + 2× 80 mm rear mounts
Drive bays
3× 3.5" or 1× 2.5" + 2× 3.5"
3× 3.5" or 2× 3.5" + 2× 2.5"
Power supply
ATX
ATX, redundant optional
Material
1.2 mm galvanized steel
1.2 mm galvanized steel
Weight (net)
6.3 kg
9.0 kg
Best for
Networking, NAS, firewall, light VMs
AI, rendering, virtualization, heavy compute
For AI builds
How much GPU memory do you need?
Only relevant if you're running AI models. If that's you: a model has to fit into the graphics card's memory before it runs, so VRAM decides what you can run, and it's the easiest place to overspend. Here's roughly what each size handles.
Rough guide, 4-bit modelsVRAM ≈ billions of parameters × 0.6A 14B model needs about 9 GB. A 70B model needs about 42 GB. Leave headroom for long context.
GPU memory
Runs (4-bit inference)
Fine-tuning
Typical use
16 GB
Up to ~14B comfortably. 32B only with offload (slow).
LoRA on 7 to 8B
Chatbots, RAG, coding assistants
24 GB
Up to ~32B. 70B tight.
LoRA up to ~13B
Local dev, mid-size models
32 GB
32B with long context. 70B with offload.
LoRA up to ~30B
Single-card high end
48 GB
70B comfortably.
Full fine-tune of small models
70B-class inference on one card
96 GB+
70B at higher precision, 120B+ at 4-bit.
Serious fine-tuning
Production serving and training
Need more than one card can hold?A 4U has the slots and cooling for multiple full-height graphics cards, so you can add their memory together for the big models. The 2U uses integrated graphics, so AI builds are a 4U job.
These are practical guides for 4-bit models at moderate context. Exact fit depends on the model, quantization and context length.
What's inside
Built to take the parts you actually want
Standard ATX internals, not proprietary parts. Whatever we put in, you can service or upgrade later without going back to one vendor.
Expansion and compatibility
Takes a standard ATX, MATX or ITX board on a normal ATX power supply. Nothing proprietary, so you choose the parts.
4U: 7 full-height slots, graphics cards up to 360 mm
A crossbar braces heavy cards so they stay seated
2U: 3 full-height plus 4 half-height slots
Storage and drive bays
Room for spinning disks and SSDs in the same box, so you can run capacity and speed together.
3 × 3.5" drives, or 2 × 3.5" plus 2 × 2.5"
Mix RAID, cache and backup in one chassis
Cooling that holds under load
Heat is what kills a server that runs around the clock. Both sizes pull air front to back through a filtered intake you can clean in seconds.
2U: low-profile air cooling up to 70 mm
4U: tower coolers to 150 mm, or a 360 mm liquid cooler
Contact Us
1000+ Genuine 5* Reviews
Hit us with your hardest questions and we promise to answer you on the same day.
PS: We work all days of the week.
Why Mansa
We build the size you need, and we build it to last
We've built more than 13,000 machines since 2018, and the most common thing we do is talk people out of overspending. Tell us the workload. We size the chassis, cooling and graphics to match it, build it in 1 to 2 days, and back it for 3 years.
Right-sized, not overkillMost builds, we recommend less than people walk in asking for. You pay for the workload, not the spec sheet.
Out in 1 to 2 daysTell us the build today and most ship within one to two days, set up and ready to rack.
3-year guaranteeWe back every build for three years, with WhatsApp support that replies. Customers run our builds for six years and more.
Built clean, built genuineNeat cabling, genuine parts, original boxes. Assembled to run continuously, not just to boot once.
