AI Infrastructure
Debug and Profile the GPU Code Behind AI
Linaro Forge brings the debugger and profiler used on the world’s supercomputers to the CUDA, HIP and NCCL code that AI training and inference depend on.
NCCL Profiling
Time in collective and point-to-point operations, and the transfer rates achieved, per GPU.
Visibility into Custom GPU Code
When your stack calls custom C++, CUDA or HIP kernels, or runs collectives across many GPUs, Linaro Forge shows what each process and GPU thread is doing. Linaro MAP also profiles Python scripts that call into native code.
Catch GPU Memory Errors
Linaro DDT detects out-of-bounds and misaligned accesses to GPU global memory in CUDA kernels, and shows the block and thread involved.
Debug Custom CUDA and HIP Kernels
Set breakpoints in kernels, step warps and wavefronts, and inspect GPU thread variables, with host and device code in one session. Kernels need debug information to debug at source level.
Find Where a Distributed Job Is Stuck
When an MPI or native GPU job hangs, attach Linaro DDT and use the parallel stack view to see which processes are waiting in which call.
Profile NCCL and GPU Utilization
Linaro MAP shows GPU utilization and memory use, NCCL time and rates, and host-to-device and peer-to-peer transfer rates over PCIe and NVLink.
Tell Us About Your GPU Cluster
We are optimizing Linaro Forge for AI workloads. Talk to our engineers about your stack and what you need from it.