400GbE for the Basement: The MikroTik CRS804-4DDQ-hRM Is the Cheap On-Ramp to a Terabit Backplane

Four ports of 400GbE for roughly the price of a mid-range GPU. The MikroTik CRS804-4DDQ-hRM is a half-width, rackmount L2/L3 switch that puts 1.6 Tbps of network bandwidth on your bench — ServeTheHome just published a full review (with the video to match) after running it as the RDMA backend for their 8x NVIDIA GB10 cluster. List price is $1,295; STH notes these "often sell at a street price of $1,100 or so," which works out to under $0.70 per Gbps — terabits at commodity prices. They bought theirs from B&H and note demand is high.

What you're actually buying//

• Four QSFP56-DD ports, each doing 400GbE (1.6 Tbps aggregate). Breakouts work — STH notes you can split them to 2x 200G and further — so four cages is a path to 400G, not a ceiling on lane count.
• Two 10Gbase-T management ports (yes, copper 10G for management — STH's own comment section called it out: "wild") plus a serial console port and front reset button.
• Redundant power: two hot-swappable Gospower 350 W PSUs. In practice the box draws about 29.3 W at idle, and even its 123 W max rating stays well inside what these supplies are built for.
• Two hot-swap fans, rack ears included, mounting kit for a standard 19" rack or side-by-side in half-rack units (MikroTik ships an actual dual-mount clip kit that slides two of these together into 1U).
• Form factor: only 218 mm wide but 387 mm deep — it's a desktop/half-width box with datacenter depth, so check your mini-rack before you order.

Management is the familiar MikroTik stack: CLI, WebFig web configurator, and WinBox on the desktop. It supports PFC, which matters if you care about RDMA over this thing (you will).

The catch that saves you money: buy QSFP56-DD optics, not QSFP112//

This is the detail that burns people who order from a spec sheet. These are QSFP56-DD cages — not the QSFP112 ports on an NVIDIA GB300 Station, and not OSFP like you'd find in Spectrum-X-class gear or ConnectX-8/9 cards. Same "400G" label, different physical connector. STH is blunt: if you are looking for DACs and optics, the QSFP56-DD is what you want. The good news is breakouts — 400G→2x200G or 400G→4x100G cables exist precisely so a four-port box can fan out to something that looks like a normal cluster.

How it actually works inside//

STH opened one up, and the silicon is the interesting part:

• Switch chip: Marvell 98DX7335A1 — a 5nm "Prestera" switch ASIC running as a 32-lane, 50G-per-lane device (4x 400G in this configuration). Getting a current-gen 5nm datacenter switch chip into a sub-$1,300 box is the whole trick.
• Management CPU: quad-core Annapurna Labs AL52400 with 4 GB of DRAM that looks ECC (five Nanya packages on board) — plus a separate 10GbE PHY for those two management RJ45 ports.
• The gotcha STH flags explicitly: there's only a 10 Gbps link between the Annapurna CPU and the switch chip. So no, this is not a 1.6 Tbps firewall box even though RouterOS has all the features for one — anything that traverses the CPU path tops out at 10G. Line-rate forwarding goes through the Marvell; management just watches over it on the side.
• Nice touch: decent-sized heatsink fins on the QSFP56-DD cages themselves, because hot-pluggable PAM4 optics run warm and most switches don't bother cooling them.

On the performance side, STH's numbers from Keysight gear: 1.567 Tbps in CyPerf (DR4-to-DR4) and 1.589 Tbps in IxNetwork (4x DR4 broken out to 16x DR1) — effectively line rate on a $1,100–$1,300 switch. And the real-world proof is that it's been carrying RDMA/NCCL traffic for their GB10 cluster for months, with PFC and MTU tuning done via RouterOS (their note that AI agents can now read MikroTik docs and configure all of this end-to-end is either a joke or prophecy — possibly both).

Who should buy one (and who shouldn't)//

Buy it if: you're building a local-AI cluster around GB10/GB300-class hardware and need an RDMA-capable backplane without writing a five-figure check; or you simply want the fastest switch money can buy per Gbps as a flex piece for a basement rack. Two of these with 400G→4x100G breakouts, run in parallel (one switch to port 0 on every node, the other to port 1), backplanes a 16-node cluster — that's the topology people are already drawing in STH's comments section.

Skip it if: you want more than four high-speed ports (this is a concentrator, not a fabric — scale topologies need real multi-plane gear); you're running a half-rack where 387 mm of depth won't fit; or your optics budget is fragile, because 400G DR4 optics are not cheap and the switch's price tag doesn't include them. It's also proprietary RouterOS/CRS firmware — no open hardware here (and if you've been following our RouterOS CVE coverage — we did a whole post on the no-fix brute-force one last week — give the management ports a firewall and a strong password; that platform has had a rough year).