NVIDIA A100 Interconnect Matrix: HDR 200G Cables & Optics

Short answer: NVIDIA A100 systems connect to the cluster fabric through ConnectX-6 at 200Gb/s HDR InfiniBand (or 200GbE) on QSFP56 ports, switched by Quantum QM8700 / QM8790 HDR switches. In a rack, use 200G QSFP56 passive DACs (up to about 2 m). Beyond that, use 200G QSFP56 active optical cables or SR4 optics over MPO-12 multimode fiber. A switch port can also be split into two 100G HDR100 links with a breakout DAC or AOC. At Resilient Tec the MMA1T00-HS 200G SR4 optic is $239, the MCP1650-H001E30 1 m DAC is $87.45 and the MCP7H50 HDR100 breakout DAC is from $80. Prices and stock checked October 12, 2026.

Which Ampere system · Link-by-link matrix · Rules that prevent wrong orders · A40 and A100 GPUs and servers · Adding Hopper · FAQ · Shop all Ampere-generation parts

Which Ampere System Do You Have?

System Cluster (compute) ports What plugs in
DGX A100 8× single-port ConnectX-6 at 200G HDR for compute, plus dual-port ConnectX-6 for storage. All QSFP56. 200G QSFP56 DACs in-rack (MCP1650-H001E30, MCP1650-H002E26); MFS1S00-H005E AOC or MMA1T00-HS optics beyond 2 m.
HGX A100 servers (Dell, HPE, Supermicro and others) Usually one ConnectX-6 HDR / 200GbE QSFP56 adapter per GPU or per GPU pair. Same cables and optics as DGX A100. Check whether the adapter runs InfiniBand or Ethernet before ordering.
A100 PCIe, A40 and A30 servers Whatever adapter the server was built with: ConnectX-6 at 200G, or 100G ConnectX-5 / ConnectX-6 Dx. At 200G, the QSFP56 parts on this page. At 100G, see the 100G EDR parts.

Ampere is the last NVIDIA generation built on QSFP56 and HDR. Hopper moved to OSFP and 400G NDR, so very few Ampere cables carry forward unchanged.

Link-by-Link Matrix

Link Reach Cable or optic Price
ConnectX-6 server → Quantum HDR switch, in-rack copper Up to 2 m MCP1650-H001E30 200G QSFP56 DAC, 1 m · MCP1650-H002E26, 2 m $87.45 · $91 (quote)
ConnectX-6 server → switch, next rack Up to 5 m (this length) MFS1S00-H005E 200G QSFP56 active optical cable, 5 m $1,194.55
ConnectX-6 server → switch, longer runs Up to 100 m on OM4 MMA1T00-HS 200G QSFP56 SR4 optic at both ends + MPO-12 UPC multimode trunk $239 each
One switch port → two servers at HDR100, in-rack Up to 2 m MCP7H50-H001R30 200G → 2× 100G breakout DAC, 1 m · MCP7H50-H002R26, 2 m $80 · $92
One switch port → two servers at HDR100, next rack Up to 5 m (this length) MFS1S50-H005E 200G → 2× 100G splitter AOC, 5 m $989.89
Quantum-2 NDR switch → two ConnectX-6 HDR servers Up to 3 m (this length) MFA7U10-H003 OSFP → 2× 200G QSFP56 AOC $1,345.45

"Quote" means not currently in stock. Request a quote with your node count and we will price the cables and optics together.

Rules That Prevent Wrong Orders

  • HDR uses UPC MPO, NDR uses APC. The MPO-12 cords for MMA1T00-HS optics are flat-polished UPC. Do not buy the green APC trunks made for NDR optics for an HDR link, and do not expect your HDR cords to work when you move to Hopper.
  • QSFP56, not QSFP28 or OSFP. A 200G HDR port is QSFP56. A 100G QSFP28 cable will usually link at 100G at best; an OSFP cable will not fit.
  • HDR100 doubles the port count. A QM8700 has 40 HDR ports. Using splitters, it serves 80 nodes at 100G each. Both ends must be set for HDR100.
  • InfiniBand or Ethernet. ConnectX-6 VPI adapters run either. MCP1650-H cables are coded for HDR InfiniBand; for 200GbE ports, ask us for the V-coded version.
  • Passive copper stops at about 2 to 3 m. Past that, use an AOC or optics.

A40 and A100 GPUs and Servers

Item What it is Price
NVIDIA A40 48GB, refurbished Passive 300 W PCIe card, 48 GB ECC GDDR6. The most cost-effective 48 GB inference card. $5,490
NVIDIA A40 48GB, new Same card, new. $6,495
R740xd A40 build Dell PowerEdge R740xd with the A40, GPU enablement kit and power cable installed and tested. $10,959
AI Inference Rack, A40 Edition Turnkey A40 inference system sized for 70B models. $14,440 (quote)
Dell 575TK NVIDIA A100 80GB PCIe A100 80GB PCIe accelerator. $14,500 (quote)
Dell R7525 with 2× A100 Dual-A100 PowerEdge server. $57,643 (quote)

Measured A40 inference numbers: What can you run on a 48GB A40? More builds: GPU Servers & On-Prem AI Guide.

Adding Hopper Nodes to an Ampere Cluster

You do not have to replace the HDR fabric to add H100 or H200 nodes. A Quantum-2 NDR switch can drive two existing ConnectX-6 HDR nodes per port with the MFA7U10-H003 OSFP → 2× QSFP56 active optical cable, while new Hopper nodes connect at 400G. The full Hopper cabling is on the H100 / H200 interconnect matrix.

Frequently Asked Questions

What cable does an A100 cluster use? A100 systems use ConnectX-6 adapters at 200Gb/s HDR InfiniBand or 200GbE, with QSFP56 ports. In a rack, use a 200G QSFP56 passive DAC such as the MCP1650-H001E30. Beyond about 2 m, use a 200G QSFP56 active optical cable such as the MFS1S00-H005E, or MMA1T00-HS SR4 optics over MPO-12 multimode fiber.

Which switch does an A100 cluster use? DGX A100 and HGX A100 clusters are usually built on NVIDIA Quantum QM8700 or QM8790 HDR InfiniBand switches, which have 40 QSFP56 ports at 200G, or 80 ports at 100G HDR100 using splitter cables.

What is HDR100? HDR100 splits one 200G HDR switch port into two 100G links with a breakout DAC (MCP7H50) or splitter AOC (MFS1S50), so one switch can serve twice as many nodes at 100G each.

Can I reuse my HDR cables with H100 or H200? Mostly no. Hopper uses OSFP ports at 400G NDR and MPO-12/APC fiber, while HDR uses QSFP56 and MPO-12/UPC fiber. The exception is the MFA7U10-H003 AOC, which connects a Quantum-2 NDR switch port to two HDR nodes so existing A100 servers can join a Hopper fabric.

Does the A40 need InfiniBand? No. The A40 is a PCIe inference card and most A40 servers run standard Ethernet. InfiniBand matters when you cluster GPUs for multi-node training.

How much do A100 cluster cables cost at Resilient Tec? The MCP1650-H001E30 1 m 200G DAC is $87.45, the MMA1T00-HS 200G SR4 optic is $239, the MCP7H50 HDR100 breakout DAC starts at $80 and the MFS1S00-H005E 5 m AOC is $1,194.55. All are new with a lifetime warranty. (Prices checked October 12, 2026.)

Related Guides

Last verified October 12, 2026 against NVIDIA LinkX documentation. Every part ships from Knoxville, Tennessee with a lifetime warranty and free US shipping. Questions about a specific build? Talk to an engineer.