Initial commit

results/h100_sdpa/inference_NVIDIA_H100_sdpa.json

@@ -0,0 +1,37 @@
+{
+  "model_name": "Qwen/Qwen3-4B",
+  "gpu_name": "NVIDIA H100",
+  "attention_implementation": "sdpa",
+  "num_requests": 10,
+  "prompt_length": 512,
+  "generation_length": 100,
+  "prefill": {
+    "stage_name": "prefill",
+    "duration_ms": 253.97859653458,
+    "tokens_processed": 5120,
+    "tokens_per_second": 20159.179040517676,
+    "energy_joules": 0.0,
+    "energy_per_token": 0.0,
+    "avg_power_watts": 0.0,
+    "peak_memory_gb": 46.01458740234375,
+    "avg_gpu_util_percent": 48.8
+  },
+  "decode": {
+    "stage_name": "decode",
+    "duration_ms": 23519.252635538578,
+    "tokens_processed": 1000,
+    "tokens_per_second": 42.51835785330007,
+    "energy_joules": 4544.901999980211,
+    "energy_per_token": 4.544901999980211,
+    "avg_power_watts": 192.5432634001641,
+    "peak_memory_gb": 46.01458740234375,
+    "avg_gpu_util_percent": 48.8
+  },
+  "e2e_latency_ms": 2377.323123207316,
+  "e2e_tokens_per_second": 257.43240118504923,
+  "e2e_energy_joules": 4544.901999980211,
+  "e2e_energy_per_token": 0.7426310457484006,
+  "ttft_ms": 25.397859653458,
+  "itl_ms": 23.519252635538578,
+  "timestamp": 1769149269.5228984
+}

results/h100_sdpa/pretrain_NVIDIA_H100_sdpa.json

@@ -0,0 +1,47 @@
+{
+  "model_name": "Qwen/Qwen3-4B",
+  "gpu_name": "NVIDIA H100",
+  "attention_implementation": "sdpa",
+  "batch_size": 3,
+  "sequence_length": 2048,
+  "num_steps": 10,
+  "forward": {
+    "stage_name": "forward",
+    "duration_ms": 1790.2467511594296,
+    "tokens_processed": 61440,
+    "tokens_per_second": 34319.29143857359,
+    "energy_joules": 981.029000043869,
+    "energy_per_token": 0.01596726888092235,
+    "avg_power_watts": 520.9058508009567,
+    "peak_memory_gb": 76.45208740234375,
+    "avg_gpu_util_percent": 100.0
+  },
+  "backward": {
+    "stage_name": "backward",
+    "duration_ms": 3854.5540031045675,
+    "tokens_processed": 61440,
+    "tokens_per_second": 15939.587290906931,
+    "energy_joules": 1953.71099999547,
+    "energy_per_token": 0.03179868164055127,
+    "avg_power_watts": 491.5443624439596,
+    "peak_memory_gb": 76.45208740234375,
+    "avg_gpu_util_percent": 100.0
+  },
+  "optimizer": {
+    "stage_name": "optimizer",
+    "duration_ms": 899.9840868636966,
+    "tokens_processed": 61440,
+    "tokens_per_second": 68267.87372886644,
+    "energy_joules": 365.9209999740124,
+    "energy_per_token": 0.005955745442285358,
+    "avg_power_watts": 377.8756124501158,
+    "peak_memory_gb": 76.45208740234375,
+    "avg_gpu_util_percent": 100.0
+  },
+  "total_duration_ms": 6544.784841127694,
+  "total_tokens": 61440,
+  "total_tokens_per_second": 9387.627170553957,
+  "total_energy_joules": 3300.6610000133514,
+  "total_energy_per_token": 0.053721695963758975,
+  "timestamp": 1769149234.99943
+}