Nvidia Announces CUDA 4.1 with LLVM Compiler

News
By Douglas Perry
published

Nvidia just released CUDA 4.1 Toolkit, which integrates, for the first time, the company's LLVM (Low Level Virtual Machine) compiler.

According to Nvidia, CUDA-based apps will gain about 10 percent performance as a result.

CUDA 4.1 also includes more than 1,000 new imaging and signal processing functions in the Performance Primitives (NPP) library, which now covers more than 3,200 functions in total. Nvidia claims that the NPP delivers 40 percent greater performance than Intel's IPP.

The Visual Profiler has been redesigned and now offers an automated expert system to give that provides step-by-step instructions to fine-tune CUDA code. Additionally, the new CUDA toolkit integrates version 2.1 of Parallel Nsight, a collection GPU developer tools for Visual Studio.

CUDA 4.1 can be downloaded from Nvidia's website.

Douglas Perry
See more GPUs News
16 Comments Comment from the forums
  • nikorr
    Good for folding : )
    Reply
  • bak0n
    Is it just my speakers or was he in need of a balancer of some sort for his voice. It was all over the place.
    Reply
  • blazorthon
    10% performance increase? Nothing to complain about there.
    Reply
  • DjEaZy
    ... GPU acceleration FTW!!!
    Reply
  • kronos_cornelius
    Nvidia make an Eclipse Plugin pleeeaseee!
    Reply
  • enewmen
    I thought the APU/GPU world was going towards OpenCL? Sure, nVidia can use CUDA experience as a competitive advantage. But I don't see any long term gains by heavy investing in CUDA specifically. What am I missing?
    Reply
  • lordstormdragon
    There are many functions and levels where OpenCL may not be the proper method to get the best results. OpenCL still lacks a vast array of functionality for GPU-based 3D rendering, for example. Until it catches up, which may or may not even happen at all, CUDA is the only viable cost-effective solution. Not that there's a lot of GPU rendering going on in the industry yet, but it's expanding for certain as Nvidia's solutions take hold in various pipelines.

    Thus, AMD is sometimes not even an option. Nvidia's own mental ray "iRay" (yes, I also hate the name) is CUDA-based, and there aren't many alternatives in the industry. And in some cases, there really doesn't need to be. AMD makes great cards too, but it would be impossible to recommend one for 3D creative content (Maya, 3DS Max, every CAD application) with any degree of honesty.

    Don't get me wrong, I'm not a brand fanatic. I do tend towards AMD CPUs, but Nvidia GPUs are the cornerstone of any nutritious CG artist.
    Reply
  • julianbautista87
    Hi. I'm a bit lost about this topic, aren't these cards mainly supposed to play videogames instead of these GPCPU applications? Because I bought the HD6850 instead of the gtx 460-765MB because of this (these 2 cards cost the same where I live) , the AMD's card doesn't have this feature, but I have never needed it...
    Reply
  • deanjo
    julianbautista87Hi. I'm a bit lost about this topic, aren't these cards mainly supposed to play videogames instead of these GPCPU applications? Because I bought the HD6850 instead of the gtx 460-765MB because of this (these 2 cards cost the same where I live) , the AMD's card doesn't have this feature, but I have never needed it...
    You are under a misguided assumption that consumer video cards are just for gamers. If that was the case there would be no need for the wide variety of discreet cards out there (or one could even argue no need for a gaming PC period since gaming can be all done on a console).
    Reply
  • madooo12
    blazorthon10% performance increase? Nothing to complain about there.and nVidia is complaining about the 30% increases of the GCN architecture
    Reply