d****e
发帖数: 14 | 1 We test their Fermi with our optimized code and found no performance
improved.
But the price is 3 times higher than Tesla.
Now my company IT guys think Fermi is a joke. |
b******t
发帖数: 965 | 2 you need to tune the code for the hardware
【在 d****e 的大作中提到】
: We test their Fermi with our optimized code and found no performance
: improved.
: But the price is 3 times higher than Tesla.
: Now my company IT guys think Fermi is a joke.
|
p****t
发帖数: 11416 | 3 虽然我觉得Fermi的设计比较失败，但是性能完全没有提升的最大可能
是你们的code写得不对
【在 d****e 的大作中提到】
: We test their Fermi with our optimized code and found no performance
: improved.
: But the price is 3 times higher than Tesla.
: Now my company IT guys think Fermi is a joke.
|
d****e
发帖数: 14 | 4 no, we made a lot of effort to push the code optimized and run fastest among
all the competing companies. After the optimization on Fermi, we found no
more performance improvement. Our application needs a lot of stencil
computation. |
b******t
发帖数: 965 | 5 stencil is a little bit regular,the added L1/L2 cache does not help much,
shared memory can do most of the job
another possibility, the code is bottomnecked by pci-e transfer or global
memory bandwidth
among
【在 d****e 的大作中提到】
: no, we made a lot of effort to push the code optimized and run fastest among
: all the competing companies. After the optimization on Fermi, we found no
: more performance improvement. Our application needs a lot of stencil
: computation.
|
d****e
发帖数: 14 | 6 For our application, the memory usage is fully optimized. We already test L1
/L2 is useless here.
【在 b******t 的大作中提到】
: stencil is a little bit regular,the added L1/L2 cache does not help much,
: shared memory can do most of the job
: another possibility, the code is bottomnecked by pci-e transfer or global
: memory bandwidth
:
: among
|
r****t
发帖数: 10904 | 7 至少 compute capability 2.0 多了很多东西吧。没比较过 performance 不知道。。
。其实我还没有用过 shared memory, 全用 constant 和 register |
