Intel QuickAssist Technology Wiki

Published by Daniel Nenni on 11-19-2026 at 2:34 am
Last updated on 11-19-2026 at 2:38 am

Intel QuickAssist Technology Hero Banner

Intel QuickAssist Technology (QAT) is a hardware accelerator for selected cryptographic and compression operations. It can free CPU cores for application work by handling operations such as AES-GCM, RSA, and DEFLATE. The application or an integrated library must submit requests to QAT. Effective support depends on the accelerator, firmware, driver, library version, and configured services. This page focuses on Linux deployment and validation. Intel overview · Algorithm matrix

Architecture

QAT is a lookaside accelerator. The application prepares buffers and submits a request through a library or driver. The device reads and writes mapped buffers using direct memory access (DMA). The host collects completion status and consumes the result. Applications can keep several operations in flight to amortize submission and completion overhead. Small requests and lightly loaded synchronous workloads may favor optimized CPU execution.

Layer Role
Application Selects operations, manages request lifetimes, and handles fallback and errors
Integration QATlib exposes the CPA API. QATzip, the OpenSSL integration, and DPDK expose interfaces for specific workloads
Driver and firmware Initialize the device, expose services and queues, and manage access to accelerator resources
Hardware Runs supported symmetric crypto, public-key crypto, and compression operations

A physical function (PF) is the PCI function used to manage device resources. With SR-IOV, it can expose virtual functions (VFs) for compatible user-space drivers or virtual machine assignment. A VF shares the underlying device’s resources, so benchmark with the expected number of tenants and service mix. The QATlib Linux path depends on supported PF and VF bindings, appropriate firmware, and the platform’s VT-d and SR-IOV settings. QATlib requirements · Linux QAT sysfs ABI

Workloads and integrations

Workload Integration Verify before deployment
TLS termination Intel’s OpenSSL QAT provider for compatible OpenSSL 3.x deployments, or the QAT Engine interface where supported Verify which handshake and record operations actually route to hardware. The project also supports optimized software execution
IPsec or packet processing DPDK QAT cryptodev poll-mode driver Check cipher, session, queue, and device support in the deployed DPDK release
Backup, storage, or data pipelines QATzip or DPDK compressdev Check codec, framing, compression and decompression support, and interoperability with downstream readers
Custom C or C++ integration QATlib CPA API and sample programs Check service capabilities and handle asynchronous completion, buffer ownership, and backpressure

AES-GCM and SHA-256 appear across the QAT 1.7, 1.8, and 2.0 columns of Intel’s cryptographic matrix. Compression support varies with hardware and format. For example, QATzip documents gzip/DEFLATE support on supported QAT 1.x and later devices and LZ4 framing from QAT 2.x. Confirm the direction and framing you actually use, since software fallback can differ between compression and decompression. Intel algorithm matrix · QAT Engine and provider · QATzip formats · DPDK documentation

Linux bring-up

  1. Confirm the CPU SKU or add-in adapter, PCI device ID, and hardware generation. Look up the corresponding driver, kernel, firmware, and user-space compatibility matrix. Intel documents integrated QAT on supported Xeon platforms and separate PCIe adapters. Intel quick start · Intel driver downloads
  2. Enable QAT as required by the server platform. For the QATlib/VFIO path, verify VT-d, SR-IOV, and the required Linux boot configuration. Install the matching distribution or vendor packages and firmware. QATlib requirements
  3. Inspect PCI enumeration, driver bindings, and kernel logs. Substitute the actual PF bus address in the example. The optional sysfs attributes depend on the driver, hardware, and kernel.
    lspci -Dnnk -d 8086:
    journalctl -k -b --no-pager | rg -i 'qat|firmware|vfio|iommu'
    
    pf=0000:00:00.0  # Example only. Replace with the QAT PF address.
    for attr in qat/state qat/cfg_services sriov_totalvfs sriov_numvfs; do
      path="/sys/bus/pci/devices/$pf/$attr"
      if [ -r "$path" ]; then
        printf '%s: ' "$attr"
        cat "$path"
      fi
    done
  4. Provision the service mix and VFs using the instructions for your driver stack. Where the deployment uses managed QATlib, verify the matching qatmgr service and device permissions. For VM passthrough, verify host VF assignment and guest-side enumeration. QATlib installation guide · Linux QAT sysfs ABI
  5. Run a sample for each service enabled on the device. Intel’s QATlib sample binaries may need to be built or installed separately. For compression, run ./dc_stateless_sample for a single round trip or ./cpa_sample_code runTests=32 with the sample corpus installed. Follow with an application-level round trip and use device counters or application telemetry to confirm hardware execution. QATlib installation and sample commands

Performance and rollout

Start with the existing optimized software path as the baseline. Keep algorithms, buffer sizes, datasets, concurrency, and output requirements consistent. For compression, match output format and record compression ratio. For TLS, measure handshakes per second separately from encrypted bulk throughput.

Measure What to compare
Throughput and tail latency Requests per second or GiB/s alongside p95 and p99 latency
CPU cost Core utilization and polling cost at equivalent throughput
Accelerator health Completed hardware operations, queue retries, errors, and resets
Fallback Software-executed requests during steady load, saturation, and device loss
Locality Performance with workers and buffers placed near the accelerator’s NUMA node

Sweep request size, queue depth, and worker count with representative traffic. Set a bounded retry or backpressure policy for full queues. Exercise restart and device-loss behavior in staging, then compare the canary against the baseline before expanding deployment. For direct asynchronous integrations, retain DMA-mapped buffers until completion and check the operation status before consuming output. QATlib documentation · QATlib installation guide

Troubleshooting

Symptom First checks
No QAT function appears in PCI inventory Confirm the SKU or adapter and check platform firmware settings
PF appears but the service fails to initialize Inspect boot logs for firmware or driver errors and compare versions with the compatibility matrix
PF works but the application finds no usable device Check enabled services, VF creation and assignment, ownership, VFIO permissions, and the managed service where used
Application succeeds but hardware counters stay flat Verify provider or library selection, operation capability, and software fallback
Throughput improves but p99 rises Check queue saturation, polling cadence, NUMA locality, and competing tenants
Compressed output fails to decode downstream Check codec and framing, then verify a round trip with the actual downstream reader

For escalation, capture the CPU SKU, QAT PCI IDs, PF/VF bindings, kernel and firmware versions, library version, service configuration, logs, and the smallest workload that reproduces the problem. Intel QAT support · QATlib installation and troubleshooting

Share this post via:

Comments

There are no comments yet.

You must register or log in to view/post comments.