Register Pressure

Register pressure is the colorful name for when the register file becomes a bottleneck. The root cause: registers in Parallel Thread eXecution (PTX) are virtual and unlimited, but the register files of a Streaming Multiprocessor (SM) are physical and limited.

How it limits performance#

• The register space each thread uses is set by the kernel's Streaming ASSembler (SASS) code.
• All threads in a thread block are scheduled onto the same SM, so a block's total register space also depends on the kernel launch configuration.
• The more space each block needs, the fewer blocks fit on an SM. That reduces occupancy and makes it harder to hide latency.

Key idea Every register a thread uses is multiplied across every thread on the SM, so a few extra registers per thread can cost whole thread blocks of occupancy.

Hardware responses#

This SemiAnalysis article explains how register pressure relates to key features in recent SM architectures:

• Asynchronous copies (added in Ampere)
• The Tensor Memory Accelerator (TMA, added in Hopper)
• Tensor memory (added in Blackwell)

On CPUs#

CPUs face register pressure too. Similar register bottlenecks limit how far loops can be strip-mined during auto-vectorization.

Related#

• Occupancy — what register pressure reduces.
• Register file — the physical resource under pressure.
• Latency hiding — harder with fewer resident warps.