Magzimilian commited on
Commit
6657987
·
1 Parent(s): 133b3c7

Parakeet: correct AGX numbers, add L4 device + peak-memory metric

Browse files
Files changed (2) hide show
  1. config.json +3 -2
  2. data/Parakeet-TDT-0.6B-v3.csv +4 -2
config.json CHANGED
@@ -351,7 +351,8 @@
351
  "accuracy_higher_is_better": false,
352
  "default_device": "agx_orin",
353
  "experiment_setup": {
354
- "agx_orin": "Values from the published model-card benchmark card: 'trtexec --fp16' baseline vs Embedl-optimized 'embedl int8', batch 1. Latency derived as 1000/FPS."
 
355
  },
356
  "accuracy_file": "data/acc-Parakeet-TDT-0.6B-v3.csv",
357
  "accuracy_title": "Open ASR Leaderboard WER (%)"
@@ -419,4 +420,4 @@
419
  }
420
  },
421
  "accuracy_title": "Accuracy"
422
- }
 
351
  "accuracy_higher_is_better": false,
352
  "default_device": "agx_orin",
353
  "experiment_setup": {
354
+ "agx_orin": "Measurement setup (from the model-card benchmark tables): encoder only, static (1, 3000, 128) log-mel input (30 s window), batch 1, trtexec GPU compute time (--noDataTransfers, CUDA Graph + Spin Wait, 1000 iterations, 2 s warm-up). Baseline is trtexec --fp16 of the unmodified encoder; Embedl INT8 uses calibrated Q/DQ (--fp16 --int8). Peak memory = engine + per-context activation pool reported by trtexec. Latency shown as 1000/FPS. Device: Jetson AGX Orin, TensorRT 10.3 (JetPack 6).",
355
+ "l4": "Measurement setup (from the model-card benchmark tables): encoder only, static (1, 3000, 128) log-mel input (30 s window), batch 1, trtexec GPU compute time (--noDataTransfers, CUDA Graph + Spin Wait, 1000 iterations, 2 s warm-up). Baseline is trtexec --fp16 of the unmodified encoder; Embedl INT8 uses calibrated Q/DQ (--fp16 --int8). Peak memory = engine + per-context activation pool reported by trtexec. Latency shown as 1000/FPS. Device: NVIDIA L4, TensorRT 10.16."
356
  },
357
  "accuracy_file": "data/acc-Parakeet-TDT-0.6B-v3.csv",
358
  "accuracy_title": "Open ASR Leaderboard WER (%)"
 
420
  }
421
  },
422
  "accuracy_title": "Accuracy"
423
+ }
data/Parakeet-TDT-0.6B-v3.csv CHANGED
@@ -1,3 +1,5 @@
1
  model_family,model,type,batch,device,res,ctx,lat,ips,mem
2
- Parakeet-TDT-0.6B-v3,nvidia/parakeet-tdt-0.6b-v3,audio,1,agx_orin,N/A,,33.1,30.2,
3
- Parakeet-TDT-0.6B-v3,embedl/parakeet-tdt-0.6b-v3-quantized-tensorrt,audio,1,agx_orin,N/A,,33.6,29.8,
 
 
 
1
  model_family,model,type,batch,device,res,ctx,lat,ips,mem
2
+ Parakeet-TDT-0.6B-v3,nvidia/parakeet-tdt-0.6b-v3,audio,1,agx_orin,N/A,,35.6,28.1,1274
3
+ Parakeet-TDT-0.6B-v3,embedl/parakeet-tdt-0.6b-v3-quantized-tensorrt,audio,1,agx_orin,N/A,,34.4,29.1,806
4
+ Parakeet-TDT-0.6B-v3,nvidia/parakeet-tdt-0.6b-v3,audio,1,l4,N/A,,16.8,59.4,1280
5
+ Parakeet-TDT-0.6B-v3,embedl/parakeet-tdt-0.6b-v3-quantized-tensorrt,audio,1,l4,N/A,,14.6,68.6,862