GTX 460 superclocked

[Jason G]

Can someone explain to me ( in small words please )
just what i'm seeing and what it means ?

 


In a nutshell, the geometry listed for the 460  is around twice the computing capacity of the GTS250.  When drivers mature & we get a bit more development along Fermi lines going, numbers should reflect as such.

For more detail:
For the GTS 250 there are 16 multiprocessors, of 32 shaders each.  With a warp size of 32 threads, that means a kernel will generally execute a block per multiprocessor, each one having 8192 registers (which just like in CPU land, are faster storage for computation than memory).

The 460 also has 32 shaders per multiprocessor, but less multiprocressors.  That's not an error, it actually has 48, but the architecture uses the extra 16 for super-scalar execution, which is a hardware optimisation, so you treat it as 32 but things go ~50% faster at the same clock.    This has 32k registers per block, which is ~4x the 250.  Also, at a maximum of 1024 threads per block it means the multiprocessors can do about twice as much at once.  All told, that should make the 460 roughly equivalent to a GTX 285 in throughput ( or 2 x GTS 250s )

JAson

[Raistmer]

warp size remains same so the number of truly simultaneous threads.
Also, why so big downgrade in freq

[Claggy]

warp size remains same so the number of truly simultaneous threads.
Also, why so big downgrade in freq

Perhaps the GTX 460 is reporting GPU clock speed instead of shader clock speed?

Claggy

[Raistmer]

Perhaps the GTX 460 is reporting GPU clock speed instead of shader clock speed?

Claggy

Looks like that. At least from my experience with other NV and ATI cards.

[Ghost0210]

Looks like it's an issue with both 460 and 465 then
GPU-Z and Caps GPU Viewer report the following for me:

Core = 608 Mhz
Memory = 802 Mhz
Shader = 1215 Mhz

Nut nVidia control panel reads it as:

Core = 607 Mhz
Memory = 1604 Mhz
Shader = 1215 Mhz

When using GPU-Z etc, I usually just double the memory clock value to get the true figure?

[Raistmer]

Some counts DDR and some not? ...

[Ghost0210]

Some counts DDR and some not? ...

not a clue  , I've just put it down to one of nVidia's little quirks
Kinda of like Boinc reading it at 570 Gflops one day and 850 the next

[Jason G]

warp size remains same so the number of truly simultaneous threads.
Also, why so big downgrade in freq

The warp size has stayed the same, but is capable of super-scalar execution with additional execution units, for ~50% greater throughput, so = 48 threads, but treat them as 32.  That is extra compute units providing instruction level parallelism ( as was introduced in Intel CPUs with the Pentium Processor).   

The way clocks are reported & used in the architecture has changed, from two separate clocks to 1 core clock. 

The clock change is a significant one that will influence anything that tries to read clock frequencies.  The extra superscalar execution one boosts Cuda-Core performance by ~50% but will look like the chip has less processors  .   In both cases they are optimisations for increased throughput, the core change being super-scalar fairly obvious but confusing at the same time  , The clock one because the hardware is used as a poly-morph engine, so needs to be synchronised to the shaders (Less obvious, more confusing at the same time)

[Raistmer]

Good news - then I may hope NV will like float4 too as ATI does
[and it will save OpenCL kernels from rewriting for NV ]

[Jason G]

Good news - then I may hope NV will like float4 too as ATI does
[and it will save OpenCL kernels from rewriting for NV ]

Yes, though when it comes to that you may need to be backporting some of my experimental code, so don't get too complacent there 

[Raistmer]

Good news - then I may hope NV will like float4 too as ATI does
[and it will save OpenCL kernels from rewriting for NV ]

Yes, though when it comes to that you may need to be backporting some of my experimental code, so don't get too complacent there 

LoL

[TouchuvGrey]

Cuda 3.0 build of x32f (seeing as you're a squire).  Probably 2 instances per Fermi card at a time.

    What ( if anything ) do i need to add or modify in my config.xml file to get
2 instances running on my GTX460 ? Taking into account that the other card
is a GTS250

[Ghost0210]

you won't have to modify your app_info.xml at all, there's now a 32 and 64 bit installer in the Beta area that should do the job for you Big thanks to Jason, for that one! Although you should make sure that either you run your cahce down completely, or ensure that your plan_class and version_num match the following before you run the installer:
608 and cuda
609 and cuda23
610 and cuda_fermi
As me and SciManSteve found out, having the incorrect pairing of plan_class and version_num will trash your entire cache of tasks.
Or if your feeling really brave there is another way to make sure that you have the correct pairing, but its quite risky and could end up trashing your cache anyway

Think your struggle to get 2 wu's running running on the 250, never had one myself, but it wasn't worth it on my old 260, and unfortunately, there's no known way at the moment that I'm aware of to run 2 wu's on 1 card and only 1 wu on another

[Frizz]

What ( if anything ) do i need to add or modify in my config.xml file to get
2 instances running on my GTX460 ? Taking into account that the other card
is a GTS250

Doesn't work.

See http://setiathome.berkeley.edu/forum_thread.php?id=61184&nowrap=true#1027766

[EDIT] AFAIK there is only minimal performance gain running two or more tasks on a 460.

[Ghost0210]

Doesn't work.
See http://setiathome.berkeley.edu/forum_thread.php?id=61184&nowrap=true#1027766

Saw that thread, but haven't  been  over there since, did the GPU_RAM suggestion by Joe not work?

Edit, just read the thread, shame, I was thinking about plugging my 260 back in