# NOTE: Metrics representing a superset of other metrics must be declared first # if they appear in the same group (displayRows). Otherwise different metrics will # dominate in the group row depending on zoom level if overlay display mode is used. version: 5 metricSets: - alias: gb20x name: General Metrics for NVIDIA GB20x (any frequency) chips: [GB202, GB203, GB205] displayRows: - name: $root:Activity type: overlay displayHeightScale: 0.5 metrics: - GR Active - name: $root:Memory type: overlay displayHeightScale: -0.5 metrics: - PCIe RX Throughput - PCIe TX Throughput - name: GPC Clock Frequency type: overlay metrics: - GPC Clock Frequency - name: SYS Clock Frequency type: overlay metrics: - SYS Clock Frequency - name: GPU Active description: The graphics or copy engine is active type: overlay metrics: - Sync Copy Engine Active - Async Copy Engine Active 0 - GR Active - name: SMs Active type: overlay metrics: - SMs Active - name: SM Instructions type: overlay metrics: - SM Issue - Tensor Active - name: SM Warp Occupancy type: stacked displayHeightScale: 3.0 metrics: - Vertex/Tess/Geometry Warps in Flight - Pixel Warps in Flight - Compute Warps in Flight - name: DRAM Bandwidth type: stacked metrics: - DRAM Read Bandwidth - DRAM Write Bandwidth - name: PCIe Bandwidth type: overlay metrics: - PCIe RX Throughput - PCIe TX Throughput metrics: - name: GPC Clock Frequency id: gpc__cycles_elapsed description: The frequency of the GPC graphics clock. In NVIDIA GPU Specs this is the Boost Clock and Base Clock. In nvidia-smi this is the "gr", "graphics", "Graphics Clock". color: '#8BDDB4' subMetrics: - name: MHz label: $freq suffix: .avg.per_second multiplier: 1.0e-06 - name: SYS Clock Frequency id: sys__cycles_elapsed description: The frequency of the SYS clock. The SYS clock is the clock used for the copy engines, compute and graphics command processor, and the hardware performance monitor streaming unit. color: '#118E44' schedulingRule: migOptional subMetrics: - name: MHz label: $freq suffix: .avg.per_second multiplier: 1.0e-06 - name: Sync Copy Engine Active id: gpu__engine_cycles_active_any_syncce color: '#B2B233' subMetrics: - name: Throughput % label: $pct suffix: .avg.pct_of_peak_sustained_elapsed - name: Cycles Active suffix: .sum - name: Async Copy Engine Active 0 id: gpu__scheduler_engine_asyncce0_cycles_active color: '#FFA700' subMetrics: - name: Throughput % label: $pct suffix: .avg.pct_of_peak_sustained_elapsed - name: Cycles Active suffix: .sum - name: GR Active id: gr__cycles_active description: Number of cycles the graphics/compute engine is active. The graphics and compute engine are active if there is any work in the graphics pipe or if the compute pipe is processing a dispatch. color: '#ED6A32' subMetrics: - name: Throughput % label: $pct suffix: .sum.pct_of_peak_sustained_elapsed - name: SMs Active id: sm__cycles_active description: The ratio of cycles SMs had at least 1 warp in flight to the number of cycles color: '#FFA5A5' subMetrics: - name: Throughput % label: $pct suffix: .avg.pct_of_peak_sustained_elapsed - name: SM Issue id: sm__inst_executed_realtime description: SM instruction issue rate. Each SM can issue 4 instructions/cycle (1 instruction/SMSP/cycle). color: '#98C4DD' subMetrics: - name: Throughput % label: $pct suffix: .avg.pct_of_peak_sustained_elapsed - name: Tensor Active id: sm__pipe_tensor_cycles_active_realtime description: Cycles the tensor pipe is active color: '#B3FF70' subMetrics: - name: Throughput % label: $pct suffix: .avg.pct_of_peak_sustained_elapsed - name: Vertex/Tess/Geometry Warps in Flight id: tpc__warps_active_shader_vtg_realtime description: Total number of graphics vertex, tesselation, geometry, and mesh shader warps in flight color: '#FFA5A5' subMetrics: - name: Throughput % label: $pct suffix: .avg.pct_of_peak_sustained_elapsed - name: Avg suffix: .avg multiplier: 0.5 - name: Avg Warps per Cycle suffix: .avg.per_cycle_elapsed multiplier: 0.5 - name: Pixel Warps in Flight id: tpc__warps_active_shader_ps_realtime description: Total number of pixel shader warps in flight color: '#8BDDB4' subMetrics: - name: Throughput % label: $pct suffix: .avg.pct_of_peak_sustained_elapsed - name: Avg suffix: .avg multiplier: 0.5 - name: Avg Warps per Cycle suffix: .avg.per_cycle_elapsed multiplier: 0.5 - name: Compute Warps in Flight id: tpc__warps_active_shader_cs_realtime description: Total number of compute shader warps in flight color: '#FFD179' subMetrics: - name: Throughput % label: $pct suffix: .avg.pct_of_peak_sustained_elapsed - name: Avg suffix: .avg multiplier: 0.5 - name: Avg Warps per Cycle suffix: .avg.per_cycle_elapsed multiplier: 0.5 - name: DRAM Read Bandwidth id: dram__read_throughput description: The ratio of cycles the DRAM interface was active reading data to the elapsed cycles in the same period as a percentage color: '#FFA5A5' subMetrics: - name: Throughput % label: $pct suffix: .avg.pct_of_peak_sustained_elapsed - name: DRAM Write Bandwidth id: dram__write_throughput description: The ratio of cycles the DRAM interface was active writing data to the elapsed cycles in the same period as a percentage color: '#98C4DD' subMetrics: - name: Throughput % label: $pct suffix: .avg.pct_of_peak_sustained_elapsed - name: PCIe RX Throughput id: pcie__read_bytes description: Number of bytes received by GPU (includes protocol) color: '#FF3D00' schedulingRule: migOptional subMetrics: - name: Throughput % label: $pct suffix: .avg.pct_of_peak_sustained_elapsed - name: PCIe TX Throughput id: pcie__write_bytes description: Number of bytes transmitted by GPU (includes protocol) color: '#00CFCF' schedulingRule: migOptional subMetrics: - name: Throughput % label: $pct suffix: .avg.pct_of_peak_sustained_elapsed