CUDA / PTX Microbenchmark

SM_120 Microarchitecture Measured

Cycle-level characterization of an NVIDIA Blackwell GPU, synthesized from inline-PTX microbenchmarks.

Microarchitecture Diagram

Simplified SM block diagram and memory hierarchy, annotated with measured latency/throughput values. Peak issue throughput comes from the SMSP count and saturated instruction benchmarks.

Instruction Fetch / Decode 1 instruction cache line = 64 B Warp Scheduler 4 SMSPs / SM · 12 warps/SMSP SMSP 0 issue: 1 warp inst/cyc I-buffer FFMA 8 SMSP 1 issue: 1 warp inst/cyc I-buffer IMAD 4 SMSP 2 issue: 1 warp inst/cyc SMSP 3 issue: 1 warp inst/cyc max issue = 4 warp inst/cyc/SM Backend Execution Pipes (per SM) FP32 Pipe FMA/ADD/MUL lat 4.5 cyc thp inst/cyc peak TFLOP/s INT32 Pipe IADD3/LEA lat 1.8 IMAD/IDP4A lat 4.5 thp inst/cyc fast IADD3 FP64 Pipe ADD/MUL/FMA lat 62.9 cyc thp inst/cyc DMMA lat 558 SFU Pipe rsqrt/sin/cos/… lat 18–46 cyc thp inst/cyc Tensor Core MMA lat cyc FP16 thp TFLOP/s MMA+FFMA overlap Global Mem Pipe LDG thp STS thp DRAM lat cyc Shared / Atomic Shared lat cyc Shared BW Global atom lat Memory Hierarchy Registers 65536 / SM ~0 cyc Shared 100 KB / SM lat — · BW — L1 Cache ~48 KB / SM lat — · BW — L2 Cache 32 MB lat — · BW — HBM / DRAM 16 GB lat — · BW — PCIe Host BW — Measured on · SM count · Core clock MHz

Topology

Device limits and on-chip resources reported by the CUDA driver and measured microbenchmarks.

Pipeline Depth & ILP

ILP sweep for FFMA and IADD3: latency at ILP=1 approximates pipeline depth; saturation reveals issue throughput.

Co-Issue

Wall-clock cycle measurement of FMA-only, IMAD-only, and interleaved FMA+IMAD dependent chains. The overlap ratio estimates independent pipe dispatch.

Tensor Memory Accelerator (TMA)

Latency of a 1D cp.async.bulk.tensor global-to-shared load (16-byte tile) measured with an mbarrier.

Clock

Memory Hierarchy

Latencies are measured with dependency-chain or pointer-chase kernels; bandwidths use streaming copies.

Memory Subsystem Probes

Cache-line detection, coalescing behavior, and async global-to-shared copy latency.

Instructions

Latency is the dependent-chain cost per instruction; throughput is the saturated chip-level rate.

More Instructions

Logic, shift, permute, integer division, 64-bit integer, and type-conversion latencies.

Instruction Throughput

Saturated instruction throughput per SM, measured with ILP=4 independent accumulators and reported in instructions per cycle.

Tensor Core

MMA latency/throughput for various shapes and data types via inline PTX.

Tensor Core More Shapes

Additional MMA shapes such as FP64 DMMA.

Atomics

Barriers & Synchronization

Warp Shuffle

Warp Collectives & Divergence

Latency of warp-level collectives and cost of branch divergence.

PCIe & Kernel Launch

Thread-Block Cluster