Torch Hammer benchmark report

ALCF Combined — ALCF Combined stress 1-GPU: Sunspot (XPU) vs Sirius (A100) — fleet_combined_stress
source: fleet_combined_stress GPU: Intel(R) Data Center GPU Max 1550 nodes: 2 GPUs: 2 benchmarks: 8
nodes
2
GPUs
2
benchmarks
8
outliers
5

(a) Batched GEMM (float32) [TFLOP/s]

fleet mean: 18.9 TFLOP/s | CV: 0.2% | mean power: 316W | max temp: 47 C
Fleet distribution of GPU mean performance (TFLOP/s)
0 0 0 0 1 GPU count 18.9–18.9 TFLOP/s: 1 GPUs 18.9–18.9 TFLOP/s: 1 GPUs 18.9 18.9 TFLOP/s mean
Fleet distribution of GPU mean power (W)
0 0 1 GPU count 316–317 W: 1 GPUs 316 W mean
Fleet distribution of GPU max temperature (°C)
0 0 1 GPU count 47.0–48.0 °C: 1 GPUs 47.5 °C mean
Fleet distribution of GPU mean SM utilization (%)
0 0 1 GPU count 100–101 %: 1 GPUs 100 % mean
Fleet distribution of GPU mean memory BW utilization (%)
0 0 1 GPU count 19.0–20.0 %: 1 GPUs 19.5 % mean
Fleet distribution of GPU mean GPU clock (MHz)
0 0 1 GPU count 1,410–1,411 MHz: 1 GPUs 1,410 MHz mean
node GPU 0 node mean vs fleet thrt
x1921c1s2b0n01919+0.1%
x3200c0s13b1n01919-0.1%

(b) Convolution (float32) [img/s]

fleet mean: 5,505,220.9 img/s | CV: 99.3% | mean power: 390W | max temp: 50 C
Fleet distribution of GPU mean performance (img/s)
0 0 0 0 1 GPU count 1.6M–5.5M img/s: 1 GPUs 5.5M–9.4M img/s: 1 GPUs 3.6M 7.4M img/s -15% +15% mean
Fleet distribution of GPU mean power (W)
0 0 1 GPU count 390–391 W: 1 GPUs 391 W mean
Fleet distribution of GPU max temperature (°C)
0 0 1 GPU count 50.0–51.0 °C: 1 GPUs 50.5 °C mean
Fleet distribution of GPU mean SM utilization (%)
0 0 1 GPU count 99.0–100 %: 1 GPUs 99.5 % mean
Fleet distribution of GPU mean memory BW utilization (%)
0 0 1 GPU count 72.0–73.0 %: 1 GPUs 72.5 % mean
Fleet distribution of GPU mean GPU clock (MHz)
0 0 1 GPU count 1,396–1,397 MHz: 1 GPUs 1,396 MHz mean
node GPU 0 node mean vs fleet thrt
x1921c1s2b0n01,640,9401,640,940-70.2%
x3200c0s13b1n09,369,5029,369,502+70.2%

(c) Einsum Attention (float32) [TFLOP/s]

fleet mean: 15.7 TFLOP/s | CV: 12.8% | mean power: 346W | max temp: 52 C
Fleet distribution of GPU mean performance (TFLOP/s)
0 0 0 0 1 GPU count 14.2–15.7 TFLOP/s: 1 GPUs 15.7–17.1 TFLOP/s: 1 GPUs 14.9 16.4 TFLOP/s mean
Fleet distribution of GPU mean power (W)
0 0 1 GPU count 346–347 W: 1 GPUs 346 W mean
Fleet distribution of GPU max temperature (°C)
0 0 1 GPU count 52.0–53.0 °C: 1 GPUs 52.5 °C mean
Fleet distribution of GPU mean SM utilization (%)
0 0 1 GPU count 100–101 %: 1 GPUs 100 % mean
Fleet distribution of GPU mean memory BW utilization (%)
0 0 1 GPU count 26.0–27.0 %: 1 GPUs 26.5 % mean
Fleet distribution of GPU mean GPU clock (MHz)
0 0 1 GPU count 1,410–1,411 MHz: 1 GPUs 1,410 MHz mean
node GPU 0 node mean vs fleet thrt
x1921c1s2b0n01414-9.0%
x3200c0s13b1n01717+9.0%

(d) Memory Traffic (float32) [GB/s]

fleet mean: 28.6 GB/s | CV: 4.7% | mean power: 194W | max temp: 51 C
Fleet distribution of GPU mean performance (GB/s)
0 0 0 0 1 GPU count 27.7–28.6 GB/s: 1 GPUs 28.6–29.6 GB/s: 1 GPUs 28.2 29.1 GB/s mean
Fleet distribution of GPU mean power (W)
0 0 1 GPU count 194–195 W: 1 GPUs 194 W mean
Fleet distribution of GPU max temperature (°C)
0 0 1 GPU count 51.0–52.0 °C: 1 GPUs 51.5 °C mean
Fleet distribution of GPU mean SM utilization (%)
0 0 1 GPU count 100–101 %: 1 GPUs 100 % mean
Fleet distribution of GPU mean memory BW utilization (%)
0 0 1 GPU count 52.0–53.0 %: 1 GPUs 52.5 % mean
Fleet distribution of GPU mean GPU clock (MHz)
0 0 1 GPU count 1,410–1,411 MHz: 1 GPUs 1,410 MHz mean
node GPU 0 node mean vs fleet thrt
x1921c1s2b0n02828-3.3%
x3200c0s13b1n03030+3.3%

(e) Heat Equation (float32) [MLUPS]

fleet mean: 13,056.2 MLUPS | CV: 26.5% | mean power: 289W | max temp: 44 C
Fleet distribution of GPU mean performance (MLUPS)
0 0 0 0 1 GPU count 11k–13k MLUPS: 1 GPUs 13k–15k MLUPS: 1 GPUs 12k 14k MLUPS -15% +15% mean
Fleet distribution of GPU mean power (W)
0 0 1 GPU count 289–290 W: 1 GPUs 290 W mean
Fleet distribution of GPU max temperature (°C)
0 0 1 GPU count 44.0–45.0 °C: 1 GPUs 44.5 °C mean
Fleet distribution of GPU mean SM utilization (%)
0 0 1 GPU count 99.0–100 %: 1 GPUs 99.5 % mean
Fleet distribution of GPU mean memory BW utilization (%)
0 0 1 GPU count 100–101 %: 1 GPUs 100 % mean
Fleet distribution of GPU mean GPU clock (MHz)
0 0 1 GPU count 1,410–1,411 MHz: 1 GPUs 1,410 MHz mean
node GPU 0 node mean vs fleet thrt
x1921c1s2b0n010,61310,613-18.7%
x3200c0s13b1n015,49915,499+18.7%

(f) Schrödinger Equation (float32) [iter/s]

fleet mean: 51.0 iter/s | CV: 21.6% | mean power: 201W | max temp: 44 C
Fleet distribution of GPU mean performance (iter/s)
0 0 0 0 1 GPU count 43.2–51.0 iter/s: 1 GPUs 51.0–58.8 iter/s: 1 GPUs 47.1 54.9 iter/s -15% +15% mean
Fleet distribution of GPU mean power (W)
0 0 1 GPU count 201–202 W: 1 GPUs 202 W mean
Fleet distribution of GPU max temperature (°C)
0 0 1 GPU count 44.0–45.0 °C: 1 GPUs 44.5 °C mean
Fleet distribution of GPU mean SM utilization (%)
0 0 1 GPU count 99.0–100 %: 1 GPUs 99.5 % mean
Fleet distribution of GPU mean memory BW utilization (%)
0 0 1 GPU count 100–101 %: 1 GPUs 100 % mean
Fleet distribution of GPU mean GPU clock (MHz)
0 0 1 GPU count 1,410–1,411 MHz: 1 GPUs 1,410 MHz mean
node GPU 0 node mean vs fleet thrt
x1921c1s2b0n04343-15.3%
x3200c0s13b1n05959+15.3%

(g) Atomic Contention (float32) [Gops/s]

fleet mean: 5.5 Gops/s | CV: 96.5% | mean power: 124W | max temp: 41 C
Fleet distribution of GPU mean performance (Gops/s)
0 0 0 0 1 GPU count 1.7–5.5 Gops/s: 1 GPUs 5.5–9.2 Gops/s: 1 GPUs 3.6 7.3 Gops/s -15% +15% mean
Fleet distribution of GPU mean power (W)
0 0 1 GPU count 124–125 W: 1 GPUs 124 W mean
Fleet distribution of GPU max temperature (°C)
0 0 1 GPU count 41.0–42.0 °C: 1 GPUs 41.5 °C mean
Fleet distribution of GPU mean SM utilization (%)
0 0 1 GPU count 99.0–100 %: 1 GPUs 99.5 % mean
Fleet distribution of GPU mean memory BW utilization (%)
0 0 1 GPU count 9.0–10.0 %: 1 GPUs 9.5 % mean
Fleet distribution of GPU mean GPU clock (MHz)
0 0 1 GPU count 1,410–1,411 MHz: 1 GPUs 1,410 MHz mean
node GPU 0 node mean vs fleet thrt
x1921c1s2b0n022-68.2%
x3200c0s13b1n099+68.2%

(h) Sparse MM (float32) [GFLOP/s]

fleet mean: 1,135.4 GFLOP/s | CV: 111.2% | mean power: 220W | max temp: 43 C
Fleet distribution of GPU mean performance (GFLOP/s)
0 0 0 0 1 GPU count 243–1,135 GFLOP/s: 1 GPUs 1,135–2,028 GFLOP/s: 1 GPUs 689 1,582 GFLOP/s -15% +15% mean
Fleet distribution of GPU mean power (W)
0 0 1 GPU count 220–222 W: 1 GPUs 221 W mean
Fleet distribution of GPU max temperature (°C)
0 0 1 GPU count 43.0–44.0 °C: 1 GPUs 43.5 °C mean
Fleet distribution of GPU mean SM utilization (%)
0 0 1 GPU count 100–101 %: 1 GPUs 100 % mean
Fleet distribution of GPU mean memory BW utilization (%)
0 0 1 GPU count 5.0–6.0 %: 1 GPUs 5.5 % mean
Fleet distribution of GPU mean GPU clock (MHz)
0 0 1 GPU count 1,410–1,411 MHz: 1 GPUs 1,410 MHz mean
node GPU 0 node mean vs fleet thrt
x1921c1s2b0n02,0282,028+78.6%
x3200c0s13b1n0243243-78.6%

Outliers (>15% from fleet mean)

x x3200c0s13b1n0:GPU0 -- Sparse MM (float32): 242.7 GFLOP/s (78.6% below fleet)
x x1921c1s2b0n0:GPU0 -- Convolution (float32): 1,640,940.0 img/s (70.2% below fleet)
x x1921c1s2b0n0:GPU0 -- Atomic Contention (float32): 1.7 Gops/s (68.2% below fleet)
x x1921c1s2b0n0:GPU0 -- Heat Equation (float32): 10,613.0 MLUPS (18.7% below fleet)
x x1921c1s2b0n0:GPU0 -- Schrödinger Equation (float32): 43.2 iter/s (15.3% below fleet)