AMD releases a 6-core proccessor for $150 dollars

[Dominus Atheos]

I've been trying all morning to come up with a justification for buying this but unfortunately I can't think of one.

If you can think of one, you should definitely get it. It's after a $50 rebate though.

[phongn]

As it happens, I may be getting one of their eight-core processors for virtualization purposes at work.

[Starglider]

The Opteron version has the additional bonus that you can make a 4 processor server without paying the ridiculous additional fee for quad-capable CPUs; or if you sacrifice some clockspeed, you can have 12 cores. 12 months ago I would have been really excited, but as of this GPU generation (58x0 / GTX 4x0) GP-GPU is far more exciting for all the compute-heavy stuff I want to do. I've been really eager to see if Bulldozer can bring AMD back to competitiveness at the high end for years. Now though I'm actually more excited about the prospect of the Northern Islands GPU series featuring low-level MIMD (for decently fine-grained branch performance).

[raptor3x]

The Opteron version has the additional bonus that you can make a 4 processor server without paying the ridiculous additional fee for quad-capable CPUs; or if you sacrifice some clockspeed, you can have 12 cores. 12 months ago I would have been really excited, but as of this GPU generation (58x0 / GTX 4x0) GP-GPU is far more exciting for all the compute-heavy stuff I want to do. I've been really eager to see if Bulldozer can bring AMD back to competitiveness at the high end for years. Now though I'm actually more excited about the prospect of the Northern Islands GPU series featuring low-level MIMD (for decently fine-grained branch performance).

GPGPU is such a pain in the ass if you need to use any kind of indirect addressing in your code, not to mention all the other hoops that there are to jump through.

[Starglider]

GPGPU is such a pain in the ass if you need to use any kind of indirect addressing in your code, not to mention all the other hoops that there are to jump through.

CUDA effectively has flat pointers over the GPU-accessible memory. OpenCL only has buffer-relative pointers, which is kind of annoying, but not really any worse than x86 programming back in the segmented memory days. I've heard about tricks for getting direct pointers in OpenCL, but I haven't tried any of them myself, since stuff like that may fail to work at the whim of the driver.

[raptor3x]

GPGPU is such a pain in the ass if you need to use any kind of indirect addressing in your code, not to mention all the other hoops that there are to jump through.

CUDA effectively has flat pointers over the GPU-accessible memory. OpenCL only has buffer-relative pointers, which is kind of annoying, but not really any worse than x86 programming back in the segmented memory days. I've heard about tricks for getting direct pointers in OpenCL, but I haven't tried any of them myself, since stuff like that may fail to work at the whim of the driver.

We tried using the Portland Group GPU enabled fortran compiler to accelerate the viscous stress calculation in our code and the best result we could get was only 80% as fast as the x86 code due to all the extra reordering and sorting necessary to port the routine over to the GPU. We were really excited when I got 30-40x acceleration out of one of my old solvers but the indirect addressing in the real code just killed us.

[RThurmont]

The problem I have with the new AMD multi-core CPUs, is that, exciting as they are by virtue of their core count (and thus multithread performance, which should be comparable to that of a first gen SPARC Niagara T1), and their low price vs. Intel, none of the brands I like are doing exciting or interesting things with them in terms of server hardware. The only modern AMD-based server that's even remotely interesting to me is the Sun x4640 (I own an x4600 M1, which uses a similiar architecture). To get a decent server with these units I'd probably have to go Supermicro (or maybe Dell or HP, in either case, yuck)

[Starglider]

We tried using the Portland Group GPU enabled fortran compiler to accelerate the viscous stress calculation in our code and the best result we could get was only 80% as fast as the x86 code due to all the extra reordering and sorting necessary to port the routine over to the GPU. We were really excited when I got 30-40x acceleration out of one of my old solvers but the indirect addressing in the real code just killed us.

Simple unconditional pointer following on its own shouldn't have that much impact, assuming your cache locality wasn't awful. Were you doing significant branching? As I mentioned earlier, I'm hot for Northern Islands because current GPUs are horrible at dynamic branching. If the branches aren't perfectly aligned on 64/32 (AMD/Nvidia) thread boundaries the different branches serialise, which quickly degrades the GPU into uselessness for nested conditionals. This is the single biggest problem for doing AI work on GPUs.