top of page

Enabling Pipelining to Hide Latencies

  • htugrulbuyukisik
  • 14 Nis 2017
  • 1 dakikada okunur

Cekirdekler API makes it very easy to enable pipelining for computing. When enabled, it breaks a work into multiple-of-4 blobs, then pushes blobs serially or concurrently to devices such that stages of different pipelines get overlapped and hide each others latencies if they have comparable latencies.

Pipelining needs more CPU threads in-flight so it is advised to enable this feature on strong CPUs.

 
 
 

Comments

Why GPGPU? 

  • Offload image-resize to all GPUs and FPGAs so server feels more relaxed to host websites

  • Move compute-heavy sql table joins to C# side to let sql server handle the data-heavy parts.

  • Make particle physics programs performance-aware, even a mild overclock to one of GPUs will increase overall performance.

  • Write your own genuine kernel code to accomplish multi-GPU computing, easily without getting low-level on host side.

 UPCOMING VERSIONS: 

 

  • Device to device pipelining

 

  • Built-in image resizer functions.

 

  • Built-in matrix-multiplication functions.

 

​

 FOLLOW Cekirdekler: 
  • Twitter B&W
 RECENT POSTS: 
 SEARCH BY TAGS: 

© 2017 by Huseyin Tugrul BUYUKISIK. Proudly created with Wix.com

  • Twitter B&W
bottom of page