Uploaded October 2026 | Updated November 2026, 1 day ago
Learn more about PyTorch here → https://ibm.biz/~puJVVvyrT
One .item() call can cost more performance than you think. ⚡ Antoni Virosi Martin explains how scalar reads can trigger graph breaks and synchronization, interrupt model compilation, and leave accelerators waiting on Python. Learn why these hidden bottlenecks matter for faster ML workloads.
00:00 - The Hidden Cost of Reading a Single Tensor Value
00:13 - How Torch.compile Optimizes Your Model
00:35 - Why Accelerators Work Ahead of Python
00:48 - The .item() Problem: Python Scalar Reads
01:12 - How Graph Breaks Fragment Optimized Regions
01:56 - Synchronization: When Python Forces the Chip to Wait
02:55 - The Real Performance Penalty: Empty Accelerator Queues
03:03 - How to Avoid These Slowdowns in Compiled Models
Continue your education in the course Compilation for Dataflow on Coursera here → https://ibm.biz/~yeC2veTHU
AI news moves fast. Sign up for a monthly newsletter for AI updates from IBM → https://ibm.biz/~6cdGktmYf
AI was used in the creation of the transcript and metadata for this video.
#aimodels #machinelearning #pytorch #ai
---------------------------------------------------------------------------------------------------------
Find us on YouTube:
🔵 IBM Technology: youtube.com/@IBMTechnology
🔵 IBM: youtube.com/@IBM
🔵 IBM Developer: youtube.com/@IBMDeveloperAdvocates
Learn more about PyTorch here → https://ibm.biz/~puJVVvyrT
One .item() call can cost more performance than you think. ⚡ Antoni Virosi Martin explains how scalar reads can trigger graph breaks and synchronization, interrupt model compilation, and leave accelerators waiting on Python. Learn why these hidden bottlenecks matter for faster ML workloads.
00:00 - The Hidden Cost of Reading a Single Tensor Value
00:13 - How Torch.compile Optimizes Your Model
00:35 - Why Accelerators Work Ahead of Python
00:48 - The .item() Problem: Python Scalar Reads
01:12 - How Graph Breaks Fragment Optimized Regions
01:56 - Synchronization: When Python Forces the Chip to Wait
02:55 - The Real Performance Penalty: Empty Accelerator Queues
03:03 - How to Avoid These Slowdowns in Compiled Models
Continue your education in the course Compilation for Dataflow on Coursera here → https://ibm.biz/~yeC2veTHU
AI news moves fast. Sign up for a monthly newsletter for AI updates from IBM → https://ibm.biz/~6cdGktmYf
AI was used in the creation of the transcript and metadata for this video.
#aimodels #machinelearning #pytorch #ai
---------------------------------------------------------------------------------------------------------
Find us on YouTube:
🔵 IBM Technology: youtube.com/@IBMTechnology
🔵 IBM: youtube.com/@IBM
🔵 IBM Developer: youtube.com/@IBMDeveloperAdvocates










