跳到主要内容
主页代理入口联系客服

Hardware guide

DMA Setup: How a Two-Machine Configuration Works

A DMA setup is not a single device — it is a topology: a host machine with a PCIe endpoint installed, an external link, and a second machine that receives and processes the data. Understanding the topology explains both why the approach works and why most setup complaints come from the link rather than the card.

Last reviewed: 2026-11-14Published by Heino HMA Hardware Lab

The link is the bottleneck, and the second machine is the real cost

PCIe bandwidth is not the constraint; the USB, Ethernet or fibre link between the two machines is. Most "slow" or "unstable" reports trace to link configuration, driver versions on the receiving machine, or thermal issues on the FPGA.

Machines required2 (host + analysis)
Host sidePCIe card installed, no driver
LinkUSB 3.x / GbE / fibre
Main bottleneckThe external link

The physical topology

The host machine holds the card in a PCIe slot — usually x1 or x4, sometimes an M.2 slot with an adapter. Nothing is installed in the host operating system: no driver, no service. The card enumerates as a PCIe endpoint and is otherwise inert from the OS perspective.

The card connects over USB, Ethernet or fibre to a second machine, which runs the client software that issues transactions and processes results. That second machine is a full computer: CPU, RAM, OS, and a network path to whatever the results feed into.

ComponentRequirementCommon mistake
Host PCIe slotFree x1/x4, or M.2 adapterSlot wired to chipset with IOMMU enforced
Host OSWindows 10/11 or LinuxInstalling a driver "to make it work"
LinkUSB 3.x / GbE / fibreUSB 2.0 port limiting throughput
Analysis machineSeparate physical computerUsing a VM on the same host

What the host machine actually needs

A free PCIe slot with adequate lanes. Chipset-attached slots usually work; CPU-attached slots behave slightly differently with respect to IOMMU grouping. An M.2 slot needs an adapter and consumes a storage port.

Enough physical clearance: the card plus a fibre transceiver can be taller than a standard low-profile bracket. This is the most common physical surprise.

BIOS settings that determine whether an IOMMU is enforcing. This is the single largest source of "it does not work" reports, and it is a configuration issue rather than a hardware one.

The failures people actually hit

Card does not enumerate: usually the slot, the adapter, or a firmware mismatch. Check Device Manager for an unknown PCI device before assuming the board is dead.

Enumerates but transfers nothing: IOMMU enforcing mappings. If VT-d or AMD-Vi is active and the OS has not mapped memory for the device, transactions fail by design.

Transfers but slowly: the external link. USB 2.0 instead of 3.x, a GbE link instead of fibre, or a hub in the path. Measure the link, not the card.

Works then stops: thermal. FPGAs in a passive low-profile card with no airflow will throttle. This is a physical installation problem, not a software one.

  • Check enumeration first, throughput second.
  • IOMMU state explains most failed transfers.
  • Thermal throttling explains most "works then stops".

Isolation: the part that matters most

For any legitimate use — forensics, driver development, anti-cheat research — the analysis and research environment must be physically separate from any machine that signs into a game account. This is not a formality; it is the difference between research and a rules violation when someone reviews what you did.

Do not use a virtual machine on the same host as the research target. Do not share disk images between the two environments: identifiers travel with images. Do not reuse the same Windows licence and profile across both.

  • Separate physical machines, not VMs.
  • Never share disk images across environments.
  • Document the topology and its purpose if the work is contracted.

How to measure whether your setup is working

Measure at each stage rather than end to end so you can find the bottleneck: link throughput with a standard network or USB benchmark, then transaction latency from the client software, then end-to-end results. A setup that shows full link speed but poor transaction latency has a client-side or firmware problem, not a link problem.

Frequently asked questions

01Do I need two computers for a DMA setup?

Yes, in the standard topology. The card reads the host machine and transmits over an external link to a second machine that processes the data.

02Does the host machine need a driver installed?

No. The card enumerates as a PCIe endpoint without a driver in the host OS. Installing a driver changes what the OS can see and is generally counterproductive.

03Why does the card enumerate but transfer nothing?

Almost always IOMMU enforcement. If VT-d or AMD-Vi is active and memory has not been mapped for the device, transactions are blocked at the hardware level.

04Why does throughput drop over time?

Usually thermal throttling on the FPGA, especially in a passive low-profile card with limited airflow. It is a physical installation issue rather than a software one.

05Can I run the analysis side in a VM?

Not for research that needs to be defensible. The analysis environment should be a separate physical machine, and disk images must not be shared between research and account machines.

Official references

  1. Microsoft: Kernel DMA Protection
  2. Microsoft: DMA remapping for drivers
  3. pcileech: project documentation
HMA 安装支持

售后支持已迁移

IM 售后频道已关闭,售后服务已转移至 Discord。

前往 Discord