LLVMpipe Scaling With Intel's Core i7 Gulftown

Written by Michael Larabel in Processors on 1 November 2010 at 08:03 AM EDT. Page 2 of 4. 15 Comments.

To no surprise considering the Core i7 970 offers up to 12 threads, 12MB of L3 cache, SSE 4.2 support, and other state-of-the-art features, with this Gulftown CPU are the best LLVMpipe numbers we have ever encountered. When utilizing all six cores, the Core i7 970 started out with an average frame-rate above 50 FPS (and at times during the timed demo had topped above 80 FPS) and then gradually dropped from there but even at 1024 x 768 the frame-rate was still above 30 FPS when using five cores or more.

In terms of scaling, when going from one to two cores, the frame-rate doubled at every step of the way except for 1920 x 1080. Simply taking advantage of one core on the Core i7 970 that is clocked at 3.20GHz with a turbo frequency of 3.46GHz was not enough to handle powering the LLVMpipe driver. Between two and three cores there was not too much of an increase, but between three and four cores there was a larger increase just as with the increase going from five to six cores being comparatively a larger gain than four to five cores. Utilizing all six cores, of course, was ideal for LLVMpipe where the best results were to be found. When enabling Hyper Threading to tap into all 12 threads of this LGA-1366 CPU, the performance actually dropped at the lower resolutions and at the higher resolutions, there was only an incremental advantage to Hyper Threading.

World of Padman uses the ioquake3 engine, like OpenArena, but it tends to be a bit more demanding on the graphics driver. With World of Padman, LLVMpipe on the Core i7 970 is able to deliver 30+ frames per second up through 1024 x 768. The performance scaling across cores is similar to OpenArena.

Urban Terror is more demanding than the two prior tests and here even at 640 x 480, Gulftown with LLVMpipe had a tough time even breaking 30 frames per second. The performance, however, did continue to scale -- albeit not linearly -- with the available cores/threads.


Related Articles