Name: NVIDIA Quantum QM8700 / QM8790 InfiniBand Switch
Brand: NVIDIA
SKU: MQM8700-HS2F

NVIDIA

NVIDIA Quantum QM8700 / QM8790 InfiniBand Switch

A 1U InfiniBand switch family with 40 x 200G HDR ports. QM8700 managed, QM8790 unmanaged; 16 Tb/s capacity, sub-130ns latency, SHARP supported.

40 x 200Gb/s HDR QSFP56 ports — 80 x HDR100 with splitters
16 Tb/s switching capacity in a 1U chassis
Sub-130ns port-to-port latency, 7.2 billion packets/s
Onboard subnet manager for fabrics up to 2,000 nodes (QM8700)

Models in this family

QM8700 MQM8700-HS2F QM8790 MQM8790-HS2F

Get a Quote +90 212 483 70 33

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

QM8700 and QM8790 are 1U HDR switches built on the NVIDIA Quantum silicon. Both offer 40 ports of 200Gb/s QSFP56 and deliver sub-130 nanosecond port-to-port latency at 16 Tb/s of non-blocking capacity. The only difference is management: the QM8700 ships with an onboard subnet manager, while the QM8790 leaves management to an external subnet manager in the fabric.

The HDR generation still forms the backbone of most existing HPC clusters in Türkiye. Because 200Gb/s is usually sufficient for classic MPI-heavy simulation workloads, expanding the current HDR fabric can be markedly more economical than moving to NDR.

The QM8700’s notable capability is HDR100 support: 200G ports split into two separate two-lane 100G ports, connecting up to 80 nodes with ConnectX-6 adapters. In dense racks holding more than 40 servers, and in clusters that want to scale without moving to a three-level fat-tree, this halves the number of switches required.

Specifications

Specification	QM8700 MQM8700-HS2F	QM8790 MQM8790-HS2F
Connectivity and capacity
Ports	40 x HDR 200G (QSFP56)	40 x HDR 200G (QSFP56)
With splitters	80 x HDR100 100G	80 x HDR100 100G
Switching capacity	16 Tb/s	16 Tb/s
Port-to-port latency	Sub-130ns	Sub-130ns
Packet rate	7.2 billion packets/s	7.2 billion packets/s
Management	Managed — CLI, WebUI, SNMP, JSON	Unmanaged — external subnet manager
Fabric capabilities
Onboard subnet manager (up to 2,000 nodes)	Yes	No
Adaptive routing	Yes	Yes
Congestion control	Yes	Yes
SHARP	Yes	Yes
VL2VL mapping	Yes	Yes
UFM agent	Yes	Yes
Virtual lanes (VL)	9 (8 data + 1 management)	9 (8 data + 1 management)
MTU	256 B - 4 KB	256 B - 4 KB
IBTA compliance	1.21 and 1.3	1.21 and 1.3
Physical and electrical
Chassis	1U, 19" rack	1U, 19" rack
Power supplies	2 (1+1 redundant), hot-swap, 1100 W AC	2 (1+1 redundant), hot-swap, 1100 W AC
Fans	N+1 redundant, hot-swap	N+1 redundant, hot-swap
Input voltage	100-127 / 200-240 VAC, 50-60 Hz	100-127 / 200-240 VAC, 50-60 Hz
Airflow	P2C (-HS2F) or C2P (-HS2R)	P2C (-HS2F) or C2P (-HS2R)
Operating temperature	0-40 °C	0-40 °C
Management ports	RJ45 100/1000, RS232, USB	RJ45 100/1000, RS232, USB

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

Connectivity Solutions

Doubling port density with HDR100

The QM8700 can split each 200G port into two separate two-lane 100G ports. On servers with ConnectX-6 adapters that means one switch reaching 80 nodes — halving the switch count in dense racks with more than 40 servers and in mid-sized clusters that want to avoid a three-level fat-tree.

Goes with InfiniBand adapters

QSFP56 optics and cable selection

Passive copper DAC is used inside the rack, active optical cable or a QSFP56 transceiver plus trunk cable between racks. In a splitter configuration the cable type must match the port-split setting; we verify this in the quote.

Goes with QSFP56 200G HDR transceivers HDR AOC cables

Expanding an existing HDR cluster

When adding capacity to a running HDR fabric, the first switch is chosen managed (QM8700) and the ones added later unmanaged (QM8790). If a subnet manager is already running in the fabric, newly added switches need not be managed — which cuts the cost of expansion noticeably.

Goes with HPC Network Design

Features

Onboard subnet manager

The x86 CPU inside the QM8700 runs a subnet manager capable of handling fabrics up to 2,000 nodes. No separate management server is required; MLNX-OS provides full chassis management over CLI, WebUI, SNMP and JSON.

SHARP — collectives aggregated in the network

Part of MPI and SHMEM collective operations is reduced on the switch, so data does not travel repeatedly between endpoints. Large vector reductions are aggregated at wire speed, which benefits machine learning workloads directly.

Modern topology support

Static and adaptive routing, congestion control and enhanced VL mapping support SlimFly, Dragonfly+ and multi-dimensional torus topologies alongside fat-tree.

Bidirectional airflow option

The P2C (connector-side outlet) and C2P (connector-side inlet) variants are chosen according to how the switch faces inside the rack. The wrong direction breaks hot-aisle / cold-aisle containment — settle this together with the rack layout before ordering.