Does Tesla support atomic operations?
Since what version? (Like CUDA 1.1 hardware supports atomics)
All current Tesla cards are sm10: so there are no atomic operations.
Also, do not confuse the CUDA software/driver version (1.0, 1.1, 2.0) with the sm compute capability of the hardware. It is true that you need CUDA 1.1 or 2.0 to perform atomic operations, but only because they were introduced into the software API with 1.1. There is no compute capability sm20 (yet), and at least the CUDA 2.0 makes no mention of it.
That is what I had thought. But there is an ACM article in the March/April issue of the magazine that I interpreted to mean there is support for atomicAdd at least on the teslas.
Here is what the article state (quoting as is) “If we want the final value of the location pointed to by total to contain the total of all elements in the array, we must combine the partial sums of all the blocks in the grid. One strategy would be to have each block write its partial sum into a second array and then launch the reduction kernel again repeating the process until we had reduced the sequence to a single value. A more attractive alternative supported by the Tesla architecture is to use atomicAdd(), an efficient atomic read-modify-write primitive supported by the memory subsystem.”
This article begins on page 41 of the issue and this quote is on page 52 second full paragraph. The article is written by nVidia people and not a 3d party article. This is what induced this post.
It would be cool if that is true because the only reason I am not using a Tesla right now is because it lacks the atomic ops. (Instead am using a 88 series card though I dont use its graphics abilities.)
Further clarifications would be awesome!
You are confusing the Tesla product with the Tesla architecture.
The current Tesla product (C870) has no support for atomic add.
Oh! I had not idea they are different. Thanks for that info!
So, what I infer is the 88 series has the ‘Tesla architecture’ but the Tesla c870 (or for that matter even the workstations?) do not have that architecture. Hence the 88 series supports the atomic ops, but not the tesla ‘products’ (870s).
G80, G84, G86, G92, G9x are all members of the Tesla architecture family.
Each chip has different compute capability: G80 has 1.0, G8x and G9x have 1.1.
The C/D/S 870 are all G80 based, so they do not support atomicAdd.
Thanks!