CUDA benchmark shows 64M-number sum runs 20× faster on GPU than CPU

ctjlewis · x · 2026-07-27

D1 CUDA benchmark on adding 64 million numbers:

The post also notes that real memory bandwidth often lands well below the box number: a T4 rated at 320 GB/s hit 264 GB/s, and a MacBook rated at 120 GB/s hit 98 GB/s—both around 82% of spec.

The takeaway: you usually don’t get the headline bandwidth on the spec sheet because DRAM loses some capacity to refresh and read/write turnaround.

Original post →

More from Infra

Infra channel →