David Kanter (dkanter@realworldtech.com) on 4/15/12 wrote:
---------------------------
>Tom (noemail@yahoo.com) on 4/15/12 wrote:
>---------------------------
>>Tom wrote:
>>> When GPUs went from pixel shaders and vertex shaders to unified
>>> shaders, it solved the problem of balancing the load between the
>>> two types of calculations. Similarly, when there is a Tesla board and
>>> a graphics board in a system, the load balancing between these two
>>> boards is much easier when the two boards have the same type of GPU.
>>
>>David Kanter wrote:
>>> Why does this matter? Why would you ever want to load balance between a Tesla and graphics board?
>>
>>On a workstation with a Tesla board and a graphics board, you would want to balance
>>the load between these boards whenever doing GPU computing since graphics drawing
>>rarely uses 100% of the available performance of the graphics board. For example,
>>if graphics drawing was using 20% of the available performance of the graphics board,
>>you could use the Tesla board plus 80% of the cycles on the graphics board for GPU
>>computing. Here is an example of software where you would want to do this:
>>
>>http://www.ks.uiuc.edu/Research/gpu/#multigpu
>
>Honestly, this seems somewhat pointless. If your compute workload is intense, then
>you probably want to do things on a remote server with 2-4 GPUs and then you just
>want a small workstation card for visualization (heck you can probably just use an IGP).
>
>I don't think it's a useful case. Besides, there are no workstation cards based on the GK104, and there won't be.
>
>The workstation cards and Tesla's use the same GPU and it will be different than any consumer graphics part.
>
>David

While I agree that this is not so important, what is
is price performance, we have raised with NVidia a number
of times that our love of their systems drops every time
they partition off critical areas of performance into
Tesla/Quadro only, as it can seriously hurt their
price/performance metrics.

Paying the extra for Quadro/Tesla can make a lot of sense
for a single unit, but make zero sense when you are looking
at tens, hundreds, or thousands of units. A lot of the
advnatge of 'GPGPU' is that it rides on the back of
consumer scale hardware..

NVidia is making quite a lot of moves lately to partition
the two targets (professional and gaming..) and that is
causing a lot of rumblings for people to whom price
performance in compute matters more than anything else.

The DP performance hit (now a double hit) is only one of
them, the more insidious was the limitation of readback
performance in non-professional cards. Unfortunately ATI
seems to get getting on the same wagon.

Things like ECC etc are smoke really, we will always go with
the better price/performance as we (and many others) can
quickly post-verify results. It matters more for a few
target applications but not as much as many seem to assume.

NVidia of course wants to defend the huge extra profitability
of professional cards, but they are painting themselves
into a bit of a corner for some users, I suspect a medium
term backlash and rethink.