Processing Under Pressure: A Guide to Accelerating Capital Markets Workloads With CPU + GPU Compute
Learn how to identify, assess, and benchmark capital markets workloads that may benefit from selective GPU acceleration.
KDB-X · GPU Acceleration
The Performance KX Is Known For, Accelerated.
Keep working in q while KDB-X offloads compute-intensive joins, sorts, aggregations, and numerical operations to NVIDIA GPUs. In an equities TCA assessment, three workloads averaged 3.3× faster end to end, with 13.1× faster GPU-eligible processing and a 27× peak function speedup.
For Platform Owners
Drops Into the Stack You Already Run
On-premises, cloud, or containerized. KDB-X GPU Acceleration is available as a first-class q-native library. Teams continue writing q; GPU compute is introduced through targeted adaptation of performance-critical paths.
On-Premises, Cloud, or Containerized
Run KDB-X GPU on your infrastructure or in the cloud. Docker-native, NVIDIA-certified on A100, H100, L40, and B200 with CUDA 12.9.
Teams Continue Working in q
Use q-native GPU functions to selectively move compute-intensive operations onto NVIDIA GPUs — without requiring CUDA development.
Targeted Changes Where They Help
Existing applications and schemas remain in place. Performance-critical paths may require targeted changes; operations that don’t benefit from GPU execution stay on CPU.
How KDB-X GPU Acceleration Works
Three steps. Same q.
KDB-X GPU Acceleration doesn’t change how you write q. It changes where the heavy work runs — selectively, on NVIDIA GPUs, while the rest of your pipeline keeps moving.
Selective GPU Residency, End To End.
CPU and GPU memory are distinct. KDB-X GPU Acceleration gives you precise control over which data moves to the GPU and when — so every byte of PCIe bandwidth is spent on what the computation actually needs.
Accelerate the q Operations That Matter Most.
Use q-native capabilities for time-series joins, sorting, filtering, aggregation, matrix computation, and controlled movement between CPU and GPU memory.
Push Keys. Not Payload.
The single highest-leverage pattern in the .gpu namespace: move only the columns an operation needs — join keys, sort keys — and keep intermediate results GPU-resident. Wide payload columns never cross the bus.
Naive
/ whole tables cross the bus, both ways
T: .gpu.to t
Q: .gpu.to q
res: .gpu.aj[`sym`time; T; Q]
out: .gpu.from res / full table, pulled backFull table, both directions – every stage.
Optimized
/ push keys, not payload
T: update .gpu.to time, .gpu.to sym from t
Q: update .gpu.to time, .gpu.to sym from q
idx: .gpu.aj[`sym`time; T; Q] / stays GPU-resident
out: .gpu.from idx / index only, pulled backJoin-key columns only – result stays on device.
Bars are illustrative of relative bytes crossing the PCIe bus, not a measured benchmark.
