Name: NVIDIA Quantum CS8500 Modular InfiniBand Switch
Brand: NVIDIA
SKU: MCS8500

NVIDIA

NVIDIA Quantum CS8500 Modular InfiniBand Switch

A 29U modular HDR InfiniBand chassis. 800 x 200G ports (or 1,600 x HDR100), 320 Tb/s capacity, closed-loop liquid cooling, hot-swap leaf and spine blades.

800 x 200Gb/s HDR ports — 1,600 ports with HDR100
320 Tb/s aggregate bidirectional capacity
29U modular chassis, closed-loop liquid cooling
40 QSFP56 ports per leaf blade, hot-swap

Models in this family

CS8500 MCS8500 CS8500-WL MCS8500-WL

Get a Quote +90 212 483 70 33

Authorised dealer supply · manufacturer warranty · price and lead time per configuration

The CS8500 is the modular-chassis counterpart of the HDR generation. It carries 800 ports of 200Gb/s in a single 29U enclosure and delivers 320 Tb/s of bidirectional capacity. On servers with ConnectX-6 adapters, splitting the ports down to HDR100 raises that count to 1,600.

The real gain of the modular design is not port count but cabling. In an equivalent spine layer built from fixed-configuration switches, every leaf-to-spine link is a separate optical module and cable. In the CS8500 those links run across the chassis backplane; only the cables going to compute nodes leave the enclosure. Along with the capital cost, the failure surface shrinks too.

The chassis is populated with leaf (MQM8510-H) and spine (MQM8520-H) blades, each carrying 40 QSFP56 ports. The management module (MMB8500) runs a subnet manager capable of handling fabrics up to 2,000 nodes.

One point deserves attention: cooling. The CS8500 is cooled only by a closed liquid loop — there is no air-cooled variant. If your data centre has a chilled-water line, a liquid-to-liquid CDU is used; if not, a liquid-to-air AHX unit. That decision has to be made before ordering, together with the mechanical plant and rack layout.

Specifications

Specification	CS8500 MCS8500	CS8500-WL MCS8500-WL
Connectivity and capacity
HDR 200G ports	800	800
HDR100 100G ports (split)	1,600	1,600
Aggregate capacity (bidirectional)	320 Tb/s	320 Tb/s
Connector	QSFP56	QSFP56
Ports per leaf blade	40 x QSFP56 HDR 200G	40 x QSFP56 HDR 200G
Modular structure
Chassis	29U modular (excl. CDU/AHX)	29U modular (excl. CDU/AHX)
Leaf blade	MQM8510-H — 40 QSFP56 ports	MQM8510-H — 40 QSFP56 ports
Spine blade	MQM8520-H — 40 QSFP56 ports	MQM8520-H — 40 QSFP56 ports
Management module	MMB8500 — x86 dual core	MMB8500 — x86 dual core
Hot-swap blades and modules	Yes	Yes
Fabric capabilities
Onboard subnet manager (up to 2,000 nodes)	Yes	Yes
Adaptive routing	Yes	Yes
Congestion control	Yes	Yes
SHARP	Yes	Yes
UFM agent	Yes	Yes
Virtual lanes (VL)	9 (8 data + 1 management)	9 (8 data + 1 management)
IBTA compliance	1.21 and 1.3	1.21 and 1.3
Physical, cooling and power
Dimensions (H x D x W)	1,289 x 770 x 428 mm	1,289 x 770 x 428 mm
Cooling	Closed-loop liquid	Closed-loop liquid, AHX hose included
Cooling unit	Liquid-to-liquid CDU (900-00796)	Liquid-to-air AHX (AHX-22KW-350MM)
Power supplies	9 (N+1), up to 16 (N+N)	9 (N+1), up to 16 (N+N)
Power supply module	MTDF-PS-B — 3 kW AC	MTDF-PS-B — 3 kW AC
Operating temperature	Air 0-35 °C, water 5-40 °C	Air 0-35 °C, water 5-40 °C

Specifications follow the manufacturer datasheet. The exact configuration is confirmed in your quote.

Connectivity Solutions

Consolidating the HDR spine into one chassis

In a few-hundred-node HDR cluster, a spine layer built from fixed switches needs dozens of boxes and hundreds of optical links. The CS8500 moves that layer inside one chassis; leaf-to-spine links run across the backplane, so only node cables leave the enclosure.

Goes with QSFP56 200G HDR transceivers

Reaching 1,600 ports with HDR100

On servers with ConnectX-6 adapters, 200G ports split into two two-lane 100G ports and a single chassis serves 1,600 nodes. Where 100Gb/s per node is enough, this lowers the cost per port considerably.

Goes with InfiniBand adapters

Liquid cooling and mechanical plant

The chassis is cooled only by a closed liquid loop; there is no air-cooled alternative. If your data centre has a chilled-water line, a liquid-to-liquid CDU is used; if not, a liquid-to-air AHX unit. This decision has to be made before ordering, together with rack layout and plant.

Goes with HPC Infrastructure Deployment

Features

320 Tb/s in one chassis

All 800 ports of 200G run non-blocking. Compared with an equivalent fat-tree built from fixed switches, the modular design needs both less rack space and markedly fewer interconnect cables.

Quiet, efficient liquid cooling

Closed-loop liquid cooling removes fan noise and lowers the energy spent on cooling. The chassis ships with either a liquid-to-air (AHX) or liquid-to-liquid (CDU) unit to match the data centre.

Fully redundant, serviceable design

Leaf and spine blades, management modules, power supplies and cooling units are hot-swappable. N+N power redundancy is supported up to 16 power supplies; a single component failure does not stop the fabric.

Onboard subnet manager

The x86 management module runs a subnet manager capable of handling fabrics up to 2,000 nodes. With the UFM agent, cluster health and performance telemetry are monitored centrally.