Both are single-port ConnectX-6 VPI cards on PCIe 4.0 x16. The HDAT runs 200Gb/s HDR or 200GbE; the ECAT stops at 100Gb/s HDR100 or 100GbE. On a 100G fabric the ECAT loses nothing. Building 200G? Only the HDAT works.
What Actually Differs Between HDAT and ECAT
The two cards share the same ConnectX-6 silicon, the same PCIe 4.0 x16 host interface and the same single QSFP56 cage. What changes is how many 50G PAM4 lanes the port drives. The HDAT drives four, for 200Gb/s. The ECAT drives two, for 100Gb/s. That is a hardware limit rather than a firmware setting, so no configuration file turns an ECAT into a 200G card.
| Spec | MCX653105A-HDAT | MCX653105A-ECAT |
|---|---|---|
| InfiniBand rate | 200Gb/s HDR | 100Gb/s HDR100 |
| Ethernet rate | 200GbE | 100GbE |
| Lanes in use | 4 x 50G PAM4 | 2 x 50G PAM4 |
| Host interface | PCIe 4.0 x16 | PCIe 4.0 x16 |
| Connector | QSFP56 | QSFP56 |
| Typical role | 200G compute or storage leaf, GPU nodes | 100G leaf, half of a split 200G port |
| Can it reach 200G | Yes | No |
HDR100 Is Half a Port, and That Changes Your Switch Plan
200G HDR is four 50G lanes in one link. HDR100 is two 50G lanes. That is why a single 200G switch port can be split with a 200G-to-2x100G DAC splitter and feed two 100G hosts, each on its own HDR100 link. It is the usual way to put 100G nodes on a 200G fabric without buying another switch.
The catch is contention. A split port gives each host half the port bandwidth, and the two halves share one physical port’s buffering. For storage nodes that burst and then idle, that is usually fine. For GPU nodes exchanging gradients all day, it is not: give those a dedicated 200G port and an HDAT.
Split behaviour is switch-specific. Send us the switch part number and we will confirm the split mode, cable type and port configuration before you order cards.
Will Your Server Actually Feed 200G?
A 200Gb/s link moves about 25 GB/s. A PCIe 4.0 x16 slot carries 31.5 GB/s in theory and closer to 25 GB/s in practice, so x16 is the floor for a 200G card and a x8 slot is not enough. In a x8 slot the card still links, but the host caps throughput well under 200G and you have paid for bandwidth the server cannot deliver.
PCIe 3.0 x16 is 15.75 GB/s raw and roughly 12 to 13 GB/s usable, which is a ceiling near 100G whatever card you install. If the server is PCIe 3.0 only, the ECAT is the honest choice; an HDAT in that slot runs at 100G-class throughput at best. We covered that slot penalty in detail in 9500-16i in a PCIe 3.0 slot.
Slot width is only half the question. The card should sit on the root complex of the CPU that owns the memory the job touches. On a two-socket server, a card in a slot wired to CPU 2 while the process runs on CPU 1 crosses the inter-socket link, and at 200G that link becomes the bottleneck. Check the slot-to-CPU map before you pick the slot.
When the ECAT Is the Right Buy
Three cases cover most builds. The fabric is 100G today, so the MCX653105A-ECAT negotiates 100GbE or HDR100 and nothing in the link sits idle. The switch port is already split, so each half is a 2-lane HDR100 link, which is exactly what the ECAT drives. Or the server is PCIe 3.0, and the slot caps you near 100G anyway.
Buy the MCX653105A-HDAT when the switch port is a full 200G port and the host has a free PCIe 4.0 x16 slot on the right CPU. If you need two 200G ports in one slot, that is a different card: the MCX653106A-HDAT.
One more case: if the node will move to 200G inside the life of the server, the HDAT avoids buying twice. The difference between the two cards is small next to pulling a server out of a rack to swap an adapter.
If the switch is a 32-port 100GbE QSFP28 unit such as the SN2700-CS2FC, every port is a single 100G link and there is nothing to split. In that fabric the ECAT is the natural fit, and a 100G ConnectX-5 card such as the MCX516A-CDAT does the same job. We compared those two 100G cards in MCX516A-CCAT vs CDAT, and what happens when a slower card meets a faster port in 25GbE on a 100GbE port.
What to Check on a Used ConnectX-6 Before You Install It
Five things decide whether a used card works in your rack.
- PSID and firmware branch. Cards built for a system vendor can carry a different PSID, and stock firmware for a retail card refuses to flash on them. Read the PSID first (flint -d <dev> q or mlxburn -info) and match the firmware branch to it. Ask us for the PSID of the card you are considering and we will send it.
- Cable rating. A 200G port needs a 200G-rated QSFP56 DAC, AOC or optic. A 100G QSFP28 cable in a QSFP56 port links at 100G only, which quietly turns an HDAT into an ECAT.
- Link type. VPI cards leave the bench configured for InfiniBand or Ethernet. Check the port type in firmware and set it to match the fabric before the card goes into production.
- Bracket and airflow. Full-height and low-profile brackets are not interchangeable between chassis. 200G optics also run hot, and a card in a low-airflow riser will throttle or drop links.
- Port count. Confirm single or dual port by part number, not by photo. The two-port cards look almost the same from the bracket side.
One last check that costs nothing: confirm the port comes up at the rate you paid for. A 200G card linked at 100G looks healthy in the OS and halves your fabric, and the cause is almost always the cable or the port setting.
Frequently Asked Questions
Do I need 200GbE or is 100GbE enough?
It depends on whether your workload is all-reduce bound. Tightly-coupled GPU training with heavy collective communication across many GPUs can be limited by per-GPU bandwidth, and 200GbE removes that ceiling. Storage, management and general-purpose traffic is comfortably served by 100GbE, where 200GbE adds cost without a measurable gain.
What is the difference between the MCX653105A HDAT and ECAT?
They are ConnectX-6 100GbE adapters distinguished by suffix, which describes the port configuration, the connector arrangement and the bracket. HDAT is typically the single-port variant and ECAT the dual-port. The functional difference is how many links you get per PCIe slot, and that is what decides which one fits your server.
Does a 200GbE adapter need a 200GbE switch?
To run at 200GbE, yes — the switch port has to terminate 200GbE. A 100GbE switch will down-negotiate the card to 100GbE, which still works and is a common way to stage a deployment. You can also run a 200GbE card on a 100GbE port and use the second port for a separate link, depending on the card.
What PCIe generation do I need for 200GbE?
PCIe 4.0 x16 in practice. At PCIe 3.0 x16 the host link carries around 15.75 GB/s, which is below what a 200GbE link wants, so the card would be limited by the host even if the network supports it. For 100GbE, PCIe 3.0 x16 is comfortable.
Is the extra cost of a dual-port card worth it?
Often yes, because a dual-port card uses one PCIe slot instead of two and gives you a single point of management and one SKU to stock. It does consume more lanes, so check the slot layout in your server. If you only need one link, a single-port card is the simpler and cheaper choice.
How do you verify a used ConnectX-6 adapter?
We clean, re-flash and bench-test on our own test line, then confirm the part number, port count, negotiated link rate and firmware level, and we read the error counters so a card that links but errors does not get listed. Ask for photos of the exact unit and its firmware level before you commit.
The ConnectX-6 models listed on zhongxindata.hk are quoted per part number, and each one is re-flashed and tested on our own test line before it leaves the bench. Availability is confirmed per model and per quantity.
Send the part number and quantity and we will check availability. WhatsApp +86 13068616914 · zhongxindata.hk
What the HDAT and ECAT variants are
- Both are ConnectX-6 100GbE adapters; the suffix distinguishes the port configuration and bracket
- HDAT is typically a single-port configuration and ECAT a dual-port one, which changes how they fit
- Check the exact suffix against the connector face before ordering
The 100GbE versus 200GbE decision
- 200GbE per GPU matters for tightly-coupled training with heavy all-reduce traffic
- For storage, management and general-purpose workloads, 100GbE is normally sufficient
- A 200GbE link also requires 200G-capable optics and a switch that terminates it
Matching the card to the server
- PCIe 4.0 x16 is the practical requirement for 200GbE; PCIe 3.0 caps you at 100GbE comfortably
- Port count and lane budget decide whether you can add a second card later
- On a used card, the firmware level and the negotiated link rate are part of our test
Related reading
- MCX653105A-ECAT 100GbE — the single-port variant
- MCX653106A-ECAT dual-port — the dual-port variant
- ConnectX-5 vs ConnectX-6 — generation context
- Network adapters — category hub
Frequently Asked Questions
Do I need 200GbE or is 100GbE enough?
It depends on whether your workload is all-reduce bound. Tightly-coupled GPU training with heavy collective communication across many GPUs can be limited by per-GPU bandwidth, and 200GbE removes that ceiling. Storage, management and general-purpose traffic is comfortably served by 100GbE, where 200GbE adds cost without a measurable gain.
What is the difference between the MCX653105A HDAT and ECAT?
They are ConnectX-6 100GbE adapters distinguished by suffix, which describes the port configuration, the connector arrangement and the bracket. HDAT is typically the single-port variant and ECAT the dual-port. The functional difference is how many links you get per PCIe slot, and that is what decides which one fits your server.
Does a 200GbE adapter need a 200GbE switch?
To run at 200GbE, yes — the switch port has to terminate 200GbE. A 100GbE switch will down-negotiate the card to 100GbE, which still works and is a common way to stage a deployment. You can also run a 200GbE card on a 100GbE port and use the second port for a separate link, depending on the card.
What PCIe generation do I need for 200GbE?
PCIe 4.0 x16 in practice. At PCIe 3.0 x16 the host link carries around 15.75 GB/s, which is below what a 200GbE link wants, so the card would be limited by the host even if the network supports it. For 100GbE, PCIe 3.0 x16 is comfortable.
Is the extra cost of a dual-port card worth it?
Often yes, because a dual-port card uses one PCIe slot instead of two and gives you a single point of management and one SKU to stock. It does consume more lanes, so check the slot layout in your server. If you only need one link, a single-port card is the simpler and cheaper choice.
How do you verify a used ConnectX-6 adapter?
We clean, re-flash and bench-test on our own test line, then confirm the part number, port count, negotiated link rate and firmware level, and we read the error counters so a card that links but errors does not get listed. Ask for photos of the exact unit and its firmware level before you commit.