GB10 Hardware Baseline — First Direct Measurements and Findings

Accelerated Computing DGX Spark / GB10 User Forum DGX Spark / GB10 Projects
,
15

Over the past week, I’ve been building gb10-kernel-probe to address a gap in GB10 / SM121a characterization tooling.

The tool runs sustained CUTLASS GEMM sweeps across tile and cluster-topology configurations while collecting hardware telemetry throughout execution.

Current sweep axes include:

  • threadblock tile shape

  • warp tile shape

  • pipeline stage depth

  • cluster topology

  • datatype

  • alignment

  • matrix layout

Telemetry captured per config includes:

  • TFLOPS

  • shared memory usage

  • occupancy

  • GPU temperature

  • power draw

  • SM clocks

  • PTX/kernel metadata

The sweep data is now exposing scheduling, thermal, power, and topology behavior during sustained tensor-core GEMM execution on GB10 systems.

New comparison data from two GB10 platforms:

  • ASUS GX10 (azampatti)

  • DGX Spark (dustin1925)

Important context:

  • azampatti ran the 48-config fast sweep

  • dustin1925 ran the full 96-config sweep (--full, all cluster shapes enabled)

=== STARTING CONDITIONS ===

azampatti (GX10):

  • Start temp: 56°C

  • Warm-start condition

dustin1925 (DGX Spark):

  • Start temp: 42°C

  • Cool-start condition

Despite the 14°C difference at sweep start, both systems converged near the same sustained operating region during tensor-core GEMM execution.

=== THERMAL BEHAVIOR ===

azampatti (GX10):

  • Rapid thermal rise

  • Plateau behavior near ~62°C

  • ~+6°C rise during 48-config sweep

dustin1925 (DGX Spark):

  • Gradual thermal accumulation

  • Stabilized near ~62-65°C

  • ~+20°C rise during full 96-config sweep

=== POWER / CLOCK BEHAVIOR ===

GX10:

  • Avg Power: ~68.4 W

  • Peak Power: ~76.9 W

DGX Spark:

  • Avg Power: ~67.7 W

  • Peak Power: ~81.4 W

Both systems maintained stable sustained power behavior throughout execution.

=== PERFORMANCE OBSERVATIONS ===

No sustained thermal or clock throttling was observed on either system.

One interesting result:
the highest throughput configuration did NOT correspond to the highest SM clocks.

Best config:

  • 13.35 TFLOPS @ 2294 MHz

Lowest config:

  • 3.97 TFLOPS @ 2398 MHz

For these GEMM kernels on GB10 / SM121a, tile shape, cluster topology, and occupancy behavior appear more influential than raw SM frequency alone.

=== CLUSTER TOPOLOGY RESULTS ===

64x64x32:

  • 1x1x1: 4.05 TFLOPS

  • 2x1x1: 3.99 TFLOPS

  • 2x2x1: 3.97 TFLOPS

The smaller tile regresses slightly as cluster size increases.

128x128x32:

  • 1x1x1: 13.20 TFLOPS

  • 2x1x1: 13.35 TFLOPS

  • 2x2x1: 13.10 TFLOPS

The larger tile benefits modestly from 2x1x1, then regresses again at 2x2x1.

So larger cluster topology is not acting as a universal throughput gain on GB10:

  • smaller tiles regress slightly

  • larger tiles benefit modestly from 2x1x1

  • larger cluster grouping does not consistently improve throughput

The analyzer layer is now exposing:

  • thermal trajectory

  • sustained power behavior

  • topology sensitivity

  • clock stability

  • platform convergence behavior

rather than raw benchmark numbers alone.

Huge thanks to:

  • azampatti for the GX10 sweep data

  • dustin1925 for the full sustained DGX Spark runs and validation work

Community-contributed runs are making it possible to build real comparative SM121a characterization data instead of isolated single-system observations.

Tooling + methodology:
https://github.com/parallelArchitect/gb10-kernel-probe