dankwiki

CUDA: Difference between revisions

← Older edit

Dank (talk | contribs)

2 user groups
11,302 edits
"65K" ?!
 
(73 intermediate revisions by the same user not shown)
Line 1: Line 1:
[[File:Gt200die-big.jpg|right|thumb|A "Fermi" GT200 die]]
[[File:Gt200die-big.jpg|right|thumb|A "Fermi" GT200 die]]
==Hardware==
==Hardware==
NVIDIA maintains a list of [http://www.nvidia.com/object/cuda_learn_products.html supported hardware]. For actual hardware, you'll need the "nvidia.ko" kernel module. Download the <tt>nvidia-kernel-source</tt> and <tt>nvidia-kernel-common</tt> packages, unpack <tt>/usr/src/nvidia-kernel.tar.bz2</tt>, and run <tt>make-kpkg modules_image</tt>. Install the resulting .deb, and modprobe nvidia. You'll see something like this in dmesg output:<pre>nvidia: module license 'NVIDIA' taints kernel.
NVIDIA maintains a list of [http://www.nvidia.com/object/cuda_learn_products.html supported hardware]. You'll need the "nvidia.ko" kernel module. On [[Debian]], use the <tt>nvidia-kernel-dkms</tt> package to build a module appropriate for your kernel (and automatically rebuild it upon kernel upgrades). You can also download the <tt>nvidia-kernel-source</tt> and <tt>nvidia-kernel-common</tt> packages, unpack <tt>/usr/src/nvidia-kernel.tar.bz2</tt>, and run <tt>make-kpkg modules_image</tt>. Install the resulting .deb, and modprobe nvidia. You'll see something like this in dmesg output:<pre>nvidia: module license 'NVIDIA' taints kernel.
Disabling lock debugging due to kernel taint
Disabling lock debugging due to kernel taint
nvidia 0000:07:00.0: enabling device (0000 -> 0003)
nvidia 0000:07:00.0: enabling device (0000 -> 0003)
Line 7: Line 7:
nvidia 0000:07:00.0: setting latency timer to 64
nvidia 0000:07:00.0: setting latency timer to 64
NVRM: loading NVIDIA UNIX x86_64 Kernel Module  190.53  Wed Dec  9 15:29:46 PST 2009</pre>
NVRM: loading NVIDIA UNIX x86_64 Kernel Module  190.53  Wed Dec  9 15:29:46 PST 2009</pre>
Once the module is loaded, CUDA should be able to find the device. See [[CUDA#deviceQuery_Output|below]] for sample outputs.
Once the module is loaded, CUDA should be able to find the device. See [[CUDA#deviceQuery_Output|below]] for sample outputs. Each device has a [[CUDA#Compute_Capabilities|compute capability]], though this does not encompass all differentiated capabilities (see also <tt>deviceOverlap</tt> and <tt>canMapHostMemory</tt>...). Note that "emulation mode" has been removed as of CUDA Toolkit Version 3.1.
===Emulation===
Otherwise, there's emulation...
<pre>[recombinator](0) $ ~/local/cuda/C/bin/linux/emurelease/deviceQuery
CUDA Device Query (Runtime API) version (CUDART static linking)
There is no device supporting CUDA.


Device 0: "Device Emulation (CPU)"
  CUDA Driver Version:                          2.30
  CUDA Runtime Version:                          2.30
  CUDA Capability Major revision number:        9999
  CUDA Capability Minor revision number:        9999
  Total amount of global memory:                4294967295 bytes
  Number of multiprocessors:                    16
  Number of cores:                              128
  Total amount of constant memory:              65536 bytes
  Total amount of shared memory per block:      16384 bytes
  Total number of registers available per block: 8192
  Warp size:                                    1
  Maximum number of threads per block:          512
  Maximum sizes of each dimension of a block:    512 x 512 x 64
  Maximum sizes of each dimension of a grid:    65535 x 65535 x 1
  Maximum memory pitch:                          262144 bytes
  Texture alignment:                            256 bytes
  Clock rate:                                    1.35 GHz
  Concurrent copy and execution:                No
  Run time limit on kernels:                    No
  Integrated:                                    Yes
  Support host page-locked memory mapping:      Yes
  Compute mode:                                  Default (multiple host threads can use this device simultaneously)
Test PASSED</pre>
Each device has a '''compute capability''', though this does not encompass all differentiated capabilities (see also <tt>deviceOverlap</tt> and <tt>canMapHostMemory</tt>...).
==CUDA model==
==CUDA model==
===Host===
===Host===
Line 53: Line 22:
** Larger one-time setup cost due to device register programming for DMA transfers.
** Larger one-time setup cost due to device register programming for DMA transfers.
** This memory will be unswappable -- allocate only as much as is needed.
** This memory will be unswappable -- allocate only as much as is needed.
* Pinned memory can be mapped directly into CUDAspace on ''integrated'' devices or in the presence of some IOMMUs.
* Pinned memory can be mapped directly into CUDAspace on ''integrated'' devices or in the presence of some [[IOMMU|IOMMUs]].
** "Zero (explicit)-copy" interface (can never hide all bus delays)
** "Zero (explicit)-copy" interface (can never hide all bus delays)
* Write-combining memory (configured via [[MTRR|MTRRs]] or [[Page Attribute Tables|PATs]]) avoids PCI snoop requirements and maximizes linear throughput
* Write-combining memory (configured via [[MTRR|MTRRs]] or [[Page Attribute Tables|PATs]]) avoids PCI snoop requirements and maximizes linear throughput
Line 63: Line 32:
===Streaming Multiprocessor===
===Streaming Multiprocessor===
* Each SM has a register file, fast local (''shared'') memory, a cache for constant memory, an instruction cache (ROP), a multithreaded instruction dispatcher, and some number of [[#Stream Processor|Stream Processors]] (SPs).
* Each SM has a register file, fast local (''shared'') memory, a cache for constant memory, an instruction cache (ROP), a multithreaded instruction dispatcher, and some number of [[#Stream Processor|Stream Processors]] (SPs).
** 8192 registers for compute capability <= 1.1, otherwise
** 8K registers for compute capability <= 1.1, otherwise
** 16384 for compute capability <= 1.3
** 16K for compute capability <= 1.3, otherwise
** 32K for compute capability <= 2.1, otherwise
** 64K through at least compute capability 3.5
* A group of threads which share a memory and can "synchronize their execution to coördinate accesses to memory" (use a [[barrier]]) form a '''block'''. Each thread has a ''threadId'' within its (three-dimensional) block.
* A group of threads which share a memory and can "synchronize their execution to coördinate accesses to memory" (use a [[barrier]]) form a '''block'''. Each thread has a ''threadId'' within its (three-dimensional) block.
** For a block of dimensions &lt;D<sub>x</sub>, D<sub>y</sub>, D<sub>z</sub>&gt;, the threadId of the thread having index &lt;x, y, z&gt; is (x + y * D<sub>x</sub> + z * D<sub>y</sub> * D<sub>x</sub>).
** For a block of dimensions &lt;D<sub>x</sub>, D<sub>y</sub>, D<sub>z</sub>&gt;, the threadId of the thread having index &lt;x, y, z&gt; is (x + y * D<sub>x</sub> + z * D<sub>y</sub> * D<sub>x</sub>).
* Register allocation is performed per-block, and rounded up to the nearest
* Register allocation is performed per-block, and rounded up to the nearest
** 256 registers per block for compute capability <= 1.1, otherwise
** 256 registers per block for compute capability <= 1.1, otherwise
** 512 registers per block for compute capability <= 1.3.
** 512 registers per block for compute capability <= 1.3
* A group of blocks which share a kernel form a '''grid'''. Each block (and each thread within that block) has a ''blockId'' within its (two-dimensional) grid.
* A group of blocks which share a kernel form a '''grid'''. Each block (and each thread within that block) has a ''blockId'' within its (two-dimensional) grid.
** For a grid of dimensions &lt;D<sub>x</sub>, D<sub>y</sub>&gt;, the blockId of the block having index &lt;x, y&gt; is (x + y * D<sub>x</sub>).
** For a grid of dimensions &lt;D<sub>x</sub>, D<sub>y</sub>&gt;, the blockId of the block having index &lt;x, y&gt; is (x + y * D<sub>x</sub>).
Line 90: Line 61:
* that the total number of threads not exceed some limit ''t'' (likely bounding the divergence-tracking stacks), and
* that the total number of threads not exceed some limit ''t'' (likely bounding the divergence-tracking stacks), and
* that the total number of blocks not exceed some limit ''b'' (likely bounding the warp-scheduling complexity).
* that the total number of blocks not exceed some limit ''b'' (likely bounding the warp-scheduling complexity).
A given SM, then, supports '''T''' values through the minimum of {''r''/'''Thr<sub>reg</sub>''', ''s''/'''Blk<sub>shmem</sub>''', and ''t''}; as the block requires fewer registers and less shared memory, the upper bound converges to ''t''. Motivations for larger blocks include:
A given SM, then, supports '''T''' values through the minimum of {''r''/'''Thr<sub>reg</sub>''', ''s''/'''Blk<sub>shmem</sub>''', and ''t''}; as the block requires fewer registers and less shared memory, the upper bound converges to ''t''.
 
Motivations for larger blocks include:
* freedom in the ''b'' dimension exposes parallelism until ''t'' <= ''b'' * '''T'''
* freedom in the ''b'' dimension exposes parallelism until ''t'' <= ''b'' * '''T'''
* larger maximum possible kernels (an absolute limit exists on grid dimensions)
* larger maximum possible kernels (an absolute limit exists on grid dimensions)
Line 110: Line 83:
===Stream Processor===
===Stream Processor===
* In-order, multithreaded processor: memory latencies can be hidden only by TLP, not ILP.
* In-order, multithreaded processor: memory latencies can be hidden only by TLP, not ILP.
** '''UPDATE''' Vasily Volkov's awesome GTC 2010 paper, "[http://www.cs.berkeley.edu/~volkov/volkov10-GTC.pdf Better Performance at Lower Occupancy]", ''destroys'' this notion.
*** Really. Go read Vasily's paper. It's better than anything you'll find here.
** Arithmetic intensity and parallelism are paramount!
** Arithmetic intensity and parallelism are paramount!
** Memory-bound kernels require sufficiently high ''occupancy'' (the ratio of concurrently-running warps to maximum possible concurrent warps (as applied, usually, to [[#Streaming Multiprocessor|SMs]])) to hide latency.
** Memory-bound kernels require sufficiently high ''occupancy'' (the ratio of concurrently-running warps to maximum possible concurrent warps (as applied, usually, to [[#Streaming Multiprocessor|SMs]])) to hide latency.
* No branch prediction or speculation (and thus also no pipeline flushes on mispredicted branches).
* No branch prediction or speculation. Full predication.
{| border="1"
{| border="1"
! Memory type
! Memory type
Line 159: Line 134:
| Read-write
| Read-write
| Read-write
| Read-write
| None
| '''1.x''': None
'''2.0+''': L1 on SM, L2 on TPC(?)
| Yes
| Yes
|-
|-
Line 189: Line 165:


===Compute Capabilities===
===Compute Capabilities===
The original public CUDA revision was 1.0, implemented on the NV50 chipset corresponding to the GeForce 8 series. Compute capability, formed of a non-negative major and minor revision number, can be queried on CUDA-capable cards. All revisions thus far have been backwards-compatible.
The original public CUDA revision was 1.0, implemented on the NV50 chipset corresponding to the GeForce 8 series. Compute capability, formed of a non-negative major and minor revision number, can be queried on CUDA-capable cards. All revisions thus far have been fowards-compatible, though recent CUDA toolkits will not generate code for CC1 or 2.
 
{| border="1" class="wikitable"
! Resource
! 1.0 SM
! 1.1 SM
! 1.2 SM
! 1.3 SM
! 2.0 SM
! 2.1 SM
! 3.0 SMX
! 3.5 SMX
! 7.0 SM
! 7.5 SM
|-
|CUDA cores
|8
|8
|8
|8
|32
|48
|192
|192
|64/32<br/>64/8
|64/2<br/>64/8
|-
|Schedulers
|1
|1
|1
|1
|2
|2
|4
|4
|4
|4
|-
|Insts/sched
|1
|1
|1
|1
|1
|2
|2
|2
|1
|1
|-
|Threads
|768
|768
|1K
|1K
|1536
|1536
|2K
|2K
|2K
|1K
|-
|Warps
|24
|24
|32
|32
|48
|48
|64
|64
|64
|32
|-
|Blocks
|8
|8
|8
|8
|8
|8
|16
|16
|32
|16
|-
|32-bit regs
|8K
|8K
|16K
|16K
|32K
|32K
|64K
|64K
|64K
|64K
|-
|Examples
|G80
|G9x
|GT21x
|GT200
|GF110
|GF10x
|GK104
|GK110
|GV100
|TU10x
|-
|}
{| border="1"
{| border="1"
! Revision
! Revision
Line 195: Line 282:
|-
|-
| 1.1
| 1.1
| Atomic ops on 32-bit global integers. Breakpoints and other debugging support.
|
* Atomic ops on 32-bit global integers.
* Breakpoints and other debugging support.
|-  
|-  
| 1.2
| 1.2
| Atomic ops on 64-bit global integers and 32-bit shared integers. 32 warps (1024 threads) and 16K registers per multiprocessor (MP). Vote instructions. Three MPs per Texture Processing Cluster (TPC). Relaxed memory coalescing constraints.
|
* Atomic ops on 64-bit global integers and 32-bit shared integers.
* 32 warps (1024 threads) and 16K registers per multiprocessor (MP).
* Vote instructions.
* Three MPs per Texture Processing Cluster (TPC).
* Relaxed memory coalescing constraints.
|-
|-
| 1.3
| 1.3
| Double-precision floating point at 32 cycles per operation.
|
* Double-precision floating point at 32 cycles per operation.
|-
|-
| 2.0
| 2.0
| Atomic addition on 32-bit global and shared FP. 48 warps (1536 threads), 48K shared memory banked 32 ways, and 32K registers per MP. 512K local memory per thread. <tt>__syncthreads_{count,and,or}()</tt>, <tt>__threadfence_system()</tt>, and <tt>__ballot()</tt>. 1024 threads per block and <tt>blockIdx.{x,y}</tt> values ranging through 1024. Larger texture references. PTX 2.0.
|
* 32 cores per SM
* 4 SFUs
* Atomic addition on 32-bit global and shared FP.
* 48 warps (1536 threads), 48K shared memory banked 32 ways, and 32K registers per MP.
* 512K local memory per thread.
* <tt>__syncthreads_{count,and,or}()</tt>, <tt>__threadfence_system()</tt>, and <tt>__ballot()</tt>.
* 1024 threads per block and <tt>blockIdx.{x,y}</tt> values ranging through 1024.
* Larger texture references.
* ''PTX 2.0''
** Efficient uniform addressing (<tt>ldu</tt>)
** Unified address space: <tt>isspacep</tt>/<tt>cvta</tt>
** Prefetching: <tt>prefetch</tt>/<tt>prefetchu</tt>
** Cache modifiers on loads and stores: <tt>.ca</tt>, <tt>.cg</tt>, <tt>.cs</tt>, <tt>.lu</tt>, <tt>.cv</tt>
** New integer ops: <tt>popc</tt>/<tt>clz</tt>/<tt>bfind</tt>/<tt>brev</tt>/<tt>bfe</tt>/<tt>bfi</tt>
** Video ops: <tt>vadd</tt>, <tt>vsub</tt>, <tt>vabsdiff</tt>, <tt>vmin</tt>, <tt>vmax</tt>, <tt>vshl</tt>, <tt>vshr</tt>, <tt>vmad</tt>, <tt>vset</tt>
** New special registers: <tt>nsmid</tt>, <tt>clock64</tt>, ...).
|-
|-
|}
| 2.1
|
* 48 cores per SM
* 8 SFUs per SM, 8 TFUs per ROP
* 2 warp schedulers per SM, capable of issuing two instructions per clock
|-
| 3.0
|
* 192 cores per SMX
* 32 SFUs per SMX, 32 TFUs per ROP
* 4 warp schedulers per SMX, capable of issuing two instructions per clock
* Double-precision instructions can be paired with non-DP
** Previously, double-precision instructions couldn't be paired with anything
* ''PTX 3.0''
** <tt>madc</tt> and <tt>mad.cc</tt> instructions
** Cubemaps and cubearrays for the <tt>tex</tt> instruction
** 3D surfaces via the <tt>suld.b.3d</tt> and <tt>sust.b.3d</tt> instructions
** <tt>pmevent.mask</tt> to trigger multiple performance counters
** 64-bit grid IDs
** 4 more performance counters, for a total of 8
** DWARF debugging symbols support


==Installation on [[Debian]]==
|-
[http://packages.debian.org/sid/libdevel/libcuda1-dev libcuda-dev] packages exist in the <tt>non-free</tt> archive area, and supply the core library <tt>libcuda.so</tt>. Together with the upstream toolkit and SDK from NVIDIA, this provides a full CUDA development environment for 64-bit Debian Unstable systems. I installed CUDA 2.3 on 2010-01-25 (hand-rolled 2.6.32.6 kernel, built with gcc-4.4). This machine did not have CUDA-compatible hardware (it uses [[Intel 965]]).
| 3.5
* Download the Ubuntu 9.04 files from NVIDIA's "[http://www.nvidia.com/object/cuda_get.html CUDA Zone]".
|
* Run the toolkit installer (<tt>sh cudatoolkit_2.3_linux_64_ubuntu9.04.run</tt>)
* 255 registers per thread
** For a user-mode install, supply <tt>$HOME/local</tt> or somesuch
* "CUDA Dynamic Parallelism", the ability to spawn threads from within device code
<pre>* Please make sure your PATH includes /home/dank/local/cuda/bin
* ''PTX 3.1''
* Please make sure your LD_LIBRARY_PATH
** A funnel shift instruction, <tt>shf</tt>
*   for 32-bit Linux distributions includes /home/dank/local/cuda/lib
** Loading read-only global data through the non-coherent texture cache, <tt>ld.global.nc</tt>
*   for 64-bit Linux distributions includes /home/dank/local/cuda/lib64
** 64-bit atomic/reduction operators extended to {or, xor, and, integer min, integer max}
* OR
** Mipmap type support
*   for 32-bit Linux distributions add /home/dank/local/cuda/lib
** Indirect texture/surface support
*   for 64-bit Linux distributions add /home/dank/local/cuda/lib64
** Extends generic addressing to include the const state space
* to /etc/ld.so.conf and run ldconfig as root


* Please read the release notes in /home/dank/local/cuda/doc/
|-
| 7.0
|
* ''PTX 6.3''
* Tensor cores
* Independent thread scheduling


* To uninstall CUDA, delete /home/dank/local/cuda
|-
* Installation Complete</pre>
| 7.5
* Run the SDK installer (<tt>sh cudasdk_2.3_linux.run</tt>)
|
** I just installed it to the same directory as the toolkit, which seems to work fine.
* ''PTX 6.4''
<pre>========================================
* Integer matrix multiplication in tensor cores
|-
|}


Configuring SDK Makefile (/home/dank/local/cuda/shared/common.mk)...
==PTX==
===Syntax Coloring===
[[File:ptxcolor.png|thumb|right|PTX with syntax coloring]]
I've got a [[vim]] syntax coloring file for PTX/NVIR/SASS at https://raw.github.com/dankamongmen/dankhome/master/.vim/syntax/nvir.vim. It operates by coloring all registers congruent to some integer mod 10 the same color:
<pre>syn match asmReg0 "v\?R[0-9]*0\(\.B\|\.F\|\.U\?\(I\|L\)\|\([^0-9]\)\@=\)"
syn match asmReg1 "v\?R[0-9]*1\(\.B\|\.F\|\.U\?\(I\|L\)\|\([^0-9]\)\@=\)"
syn match asmReg2 "v\?R[0-9]*2\(\.B\|\.F\|\.U\?\(I\|L\)\|\([^0-9]\)\@=\)"
syn match asmReg3 "v\?R[0-9]*3\(\.B\|\.F\|\.U\?\(I\|L\)\|\([^0-9]\)\@=\)"
syn match asmReg4 "v\?R[0-9]*4\(\.B\|\.F\|\.U\?\(I\|L\)\|\([^0-9]\)\@=\)"
syn match asmReg5 "v\?R[0-9]*5\(\.B\|\.F\|\.U\?\(I\|L\)\|\([^0-9]\)\@=\)"
syn match asmReg6 "v\?R[0-9]*6\(\.B\|\.F\|\.U\?\(I\|L\)\|\([^0-9]\)\@=\)"
syn match asmReg7 "v\?R[0-9]*7\(\.B\|\.F\|\.U\?\(I\|L\)\|\([^0-9]\)\@=\)"
syn match asmReg8 "v\?R[0-9]*8\(\.B\|\.F\|\.U\?\(I\|L\)\|\([^0-9]\)\@=\)"
syn match asmReg9 "v\?R[0-9]*9\(\.B\|\.F\|\.U\?\(I\|L\)\|\([^0-9]\)\@=\)"
syn match asmPReg "P[0-9]\([0-9]*\)\(\.B\|\.F\|\.U\?\(I\|L\)\|\([^0-9]\)\@=\)"
syn match asmBB "BB[0-9][0-9]*\(_\d\d*\)\?"
syn match asmBBNew "BB-\d\d*"
syn match nvirNT ".NEXT_TRUE.*"
syn match nvirNF ".NEXT_FALSE.*"
syn match hexconst "0x\x\+\(\.F\|\.U\?\(I\|L\)\)\?"
syn match spreg "\(ctaid\|ntid\|tid\|nctaid\).\(x\|y\|z\)"</pre>


========================================
* Please make sure your PATH includes /home/dank/local/cuda/bin
* Please make sure your LD_LIBRARY_PATH includes /home/dank/local/cuda/lib
* To uninstall the NVIDIA GPU Computing SDK, please delete /home/dank/local/cuda
* Installation Complete</pre>
==Building CUDA Apps==
==Building CUDA Apps==
===nvcc flags===
===nvcc flags===
* <tt>-ptax-options=-v</tt> displays per-thread register usage
Pass flags to <tt>ptxas</tt> via -X:
* <tt>-X -v</tt> displays per-thread register usage
* <tt>-X -abi=no</tt> disables the PTX ABI, saving registers but taking away your stack
* <tt>-dlcm={cg,cs,ca}</tt> modifies cache behavior for loads
* <tt>-dscm={cw,cs}</tt> modifies cache behavior for stores
===SDK's common.mk===
===SDK's common.mk===
This assumes use of the SDK's common.mk, as recommended by the documentation.
This assumes use of the SDK's common.mk, as recommended by the documentation.
Line 284: Line 438:


==deviceQuery info==
==deviceQuery info==
===Compute capability 2.0===
* Memory shown is that amount which is free; I've substituted total VRAM.
===Compute capability 1.3===
* Most CUDA devices can switch between multiple frequencies; the "Clock rate" output ought be considered accurate only at a given moment, and the outputs listed here are merely illustrative.
* Three device modes are currently supported:
** 0: Default (multiple applications can use the device)
** 1: Exclusive (only one application may use the device; other calls to <tt>cuCtxCreate</tt> will fail)
** 2: Disabled (no applications may use the device; all calls to <tt>cuCtxCreate</tt> will fail
* The mode can be set using <tt>nvidia-smi</tt>'s -c option, specifying the device number via -g.
* A run time limit is activated by default if the device is being used to drive a display.
* Please feel free to [mailto:nickblack@acm.org send me output!]




====Tesla C1060====
{| border="1"
<pre>Device 0: "Tesla C1060"
! Device name
  CUDA Driver Version:                          2.30
! Memory
  CUDA Runtime Version:                          2.30
! MP's
  CUDA Capability Major revision number:        1
! Cores
  CUDA Capability Minor revision number:        3
! Shmem/block
  Total amount of global memory:                4294705152 bytes
! Reg/block
  Number of multiprocessors:                    30
! Warp size
  Number of cores:                              240
! Thr/block
  Total amount of constant memory:              65536 bytes
! Texalign
  Total amount of shared memory per block:      16384 bytes
! Clock
  Total number of registers available per block: 16384
! C+E?
  Warp size:                                    32
! Integrated?
  Maximum number of threads per block:           512
! Shared maps?
  Maximum sizes of each dimension of a block:    512 x 512 x 64
|-
  Maximum sizes of each dimension of a grid:    65535 x 65535 x 1
! COLSPAN="13" style="background:#eebeb6;" | Compute capability 7.0
  Maximum memory pitch:                          262144 bytes
|-
  Texture alignment:                            256 bytes
| Tesla V100
  Clock rate:                                    1.30 GHz
| 16GB
  Concurrent copy and execution:                Yes
| 84
  Integrated:                                    No
| 5376/2688/672
  Support host page-locked memory mapping:       Yes</pre>
|
Other clock rates: 1.44 GHz
|
 
|
====GeForce GTX 295====
|
<pre>Device 1: "GeForce GTX 295"
|
  CUDA Driver Version:                          2.30
| 1.53GHz
  CUDA Runtime Version:                          2.30
| Yes
  CUDA Capability Major revision number:        1
| No
  CUDA Capability Minor revision number:        3
| Yes
  Total amount of global memory:                939261952 bytes
|-
  Number of multiprocessors:                    30
! COLSPAN="13" style="background:#8070D8;" | Compute capability 3.0
  Number of cores:                              240
|-
  Total amount of constant memory:              65536 bytes
| GeForce GTX 680
  Total amount of shared memory per block:      16384 bytes
| 1.5GB
  Total number of registers available per block: 16384
| 8
  Warp size:                                    32
| 1536
  Maximum number of threads per block:          512
|
  Maximum sizes of each dimension of a block:    512 x 512 x 64
|
  Maximum sizes of each dimension of a grid:    65535 x 65535 x 1
|
  Maximum memory pitch:                          262144 bytes
|
  Texture alignment:                            256 bytes
|
  Clock rate:                                    1.24 GHz
|
  Concurrent copy and execution:                Yes
| Yes
  Integrated:                                    No
| No
  Support host page-locked memory mapping:      Yes</pre>
| Yes
====GeForce GTX 285====
|-
<pre>Device 0: "GeForce GTX 285"
! COLSPAN="13" style="background:#ffdead;" | Compute capability 2.1
  CUDA Capability Major revision number:        1
|-
  CUDA Capability Minor revision number:        3
| GeForce GTX 560 Ti
  Total amount of global memory:                1073414144 bytes
|
  Number of multiprocessors:                    30
|
  Number of cores:                              240
|
  Total amount of constant memory:              65536 bytes
|
  Total amount of shared memory per block:      16384 bytes
|
  Total number of registers available per block: 16384
|
  Warp size:                                    32
|
  Maximum number of threads per block:          512
|
  Maximum sizes of each dimension of a block:    512 x 512 x 64
|
  Maximum sizes of each dimension of a grid:    65535 x 65535 x 1
|
  Maximum memory pitch:                          262144 bytes
|
  Texture alignment:                            256 bytes
|
  Clock rate:                                    1.48 GHz
|-
  Concurrent copy and execution:                Yes
| GeForce GTX 550 Ti
  Integrated:                                    Yes
|
  Support host page-locked memory mapping:      Yes</pre>
|
 
|
====GeForce GTX 280====
|
<pre>Device 0: "GeForce GTX 280"
|
  CUDA Driver Version:                          2.30
|
  CUDA Runtime Version:                         2.30
|
  CUDA Capability Major revision number:        1
|
  CUDA Capability Minor revision number:        3
|
  Total amount of global memory:                1073020928 bytes
|
  Number of multiprocessors:                    30
|
  Number of cores:                              240
|
  Total amount of constant memory:              65536 bytes
|-
  Total amount of shared memory per block:      16384 bytes
| GeForce GTX 460
  Total number of registers available per block: 16384
| 1GB
  Warp size:                                    32
| 7
  Maximum number of threads per block:          512
| 224
  Maximum sizes of each dimension of a block:    512 x 512 x 64
| 48k
  Maximum sizes of each dimension of a grid:    65535 x 65535 x 1
| 32k
  Maximum memory pitch:                          262144 bytes
| 32
  Texture alignment:                            256 bytes
| 1024
  Clock rate:                                    1.30 GHz
| 512b
  Concurrent copy and execution:                Yes
| 1.35GHz
  Integrated:                                    No
| Yes
  Support host page-locked memory mapping:      Yes</pre>
| No
 
| Yes
====GeForce GTX 260====
|-
<pre>Device 0: "GeForce GTX 260"
| GeForce GTS 450
  CUDA Driver Version:                          2.30
|
  CUDA Runtime Version:                          2.30
|
  CUDA Capability Major revision number:        1
|
  CUDA Capability Minor revision number:        3
|
  Total amount of global memory:                938803200 bytes
|
  Number of multiprocessors:                    27
|
  Number of cores:                              216
|
  Total amount of constant memory:              65536 bytes
|
  Total amount of shared memory per block:      16384 bytes
|
  Total number of registers available per block: 16384
|
  Warp size:                                    32
|
  Maximum number of threads per block:          512
|
  Maximum sizes of each dimension of a block:    512 x 512 x 64
|-
  Maximum sizes of each dimension of a grid:    65535 x 65535 x 1
! COLSPAN="13" style="background:#ffdead;" | Compute capability 2.0
  Maximum memory pitch:                          262144 bytes
|-
  Texture alignment:                            256 bytes
| GeForce GTX 580
  Clock rate:                                    1.47 GHz
| 1.5GB
  Concurrent copy and execution:                Yes
| 16
  Integrated:                                    No
| 512
  Support host page-locked memory mapping:      Yes</pre>
|
===Compute capability 1.2===
|
====GeForce GTS 360M (PCIe x16)====
| 32
<pre>Device 0: "GeForce GTS 360M"
| 1024
  CUDA Driver Version:                          3.0
|
  CUDA Capability Major revision number:        1
| 1.544GHz
  CUDA Capability Minor revision number:        2
| Yes
  Total amount of global memory:                1073020928 bytes
| No
  Number of multiprocessors:                    12
| Yes
  Number of cores:                              96
|-
  Total amount of constant memory:              65536 bytes
| Tesla C2050 (*CB)
  Total amount of shared memory per block:      16384 bytes
| 3GB
  Total number of registers available per block: 16384
| 14
  Warp size:                                    32
| 448
  Maximum number of threads per block:          512
| 48k
  Maximum sizes of each dimension of a block:    512 x 512 x 64
| 32k
  Maximum sizes of each dimension of a grid:    65535 x 65535 x 1
| 32
  Maximum memory pitch:                          2147483647 bytes
| 1024
  Texture alignment:                            256 bytes
| 512b
  Clock rate:                                    1.32 GHz
| 1.15GHz
  Concurrent copy and execution:                Yes
| Yes
  Integrated:                                    No
| No
  Support host page-locked memory mapping:      Yes</pre>
| Yes
====GeForce 310 (PCIe x16)====
|-
<pre>Device 0: "GeForce 310"
| Tesla C2070 (*CB)
  CUDA Driver Version:                          3.0
| 6GB
  CUDA Runtime Version:                          2.30
| 14
  CUDA Capability Major revision number:        1
| 448
  CUDA Capability Minor revision number:        2
| 48k
  Total amount of global memory:                536084480 bytes
| 32k
  Number of multiprocessors:                    2
| 32
  Number of cores:                              16
| 1024
  Total amount of constant memory:              65536 bytes
| 512b
  Total amount of shared memory per block:      16384 bytes
| 1.15GHz
  Total number of registers available per block: 16384
| Yes
  Warp size:                                    32
| No
  Maximum number of threads per block:          512
| Yes
  Maximum sizes of each dimension of a block:    512 x 512 x 64
|-
  Maximum sizes of each dimension of a grid:    65535 x 65535 x 1
| GeForce GTX 480
  Maximum memory pitch:                          262144 bytes
| 1536MB
  Texture alignment:                            256 bytes
| 15
  Clock rate:                                    1.40 GHz
| 480
  Concurrent copy and execution:                Yes
|
  Integrated:                                    No
|
  Support host page-locked memory mapping:      Yes</pre>
|
====GeForce 240 GT====
|
<pre>Device 0 GeForce GT 240
|
  CUDA Driver Version:                          3.0
|
  CUDA Runtime Version:                          2.30
|
  CUDA Capability Major revision number:        1
|
  CUDA Capability Minor revision number:        2
|
  Total amount of global memory:                1073414144 bytes
|-
  Number of multiprocessors:                    12
| GeForce GTX 470
  Number of cores:                              96
| 1280MB
  Total amount of constant memory:              65536 bytes
| 14
  Total amount of shared memory per block:      16384 bytes
| 448
  Total number of registers available per block: 16384
|
  Warp size:                                    32
|
  Maximum number of threads per block:          512
|
  Maximum sizes of each dimension of a block:    512,512,64
|
  Maximum sizes of each dimension of a grid:    65535,65535,1
|
  Maximum memory pitch:                          262144 bytes
|
  Texture alignment:                            256 bytes
|
  Clock rate:                                    1.424 GHz
|
  Concurrent copy and execution:                Yes
|
  Integrated:                                    No
|-
  Support host page-locked memory mapping:      Yes</pre>
! COLSPAN="13" style="background:#efefef;" | Compute capability 1.3
 
|-
===Compute capability 1.1===
| Tesla C1060
====ION====
| 4GB
<pre>Device 0: "ION"
| 30
  CUDA Driver Version:                          2.30
| 240
  CUDA Runtime Version:                          2.30
| 16384b
  CUDA Capability Major revision number:        1
| 16384
  CUDA Capability Minor revision number:        1
| 32
  Total amount of global memory:                268435456 bytes
| 512
  Number of multiprocessors:                    2
| 256b
  Number of cores:                              16
| 1.30GHz
  Total amount of constant memory:              65536 bytes
| Yes
  Total amount of shared memory per block:      16384 bytes
| No
  Total number of registers available per block: 8192
| Yes
  Warp size:                                    32
|-
  Maximum number of threads per block:          512
| GeForce GTX 295
  Maximum sizes of each dimension of a block:    512 x 512 x 64
| 1GB
  Maximum sizes of each dimension of a grid:    65535 x 65535 x 1
| 30
  Maximum memory pitch:                          262144 bytes
| 240
  Texture alignment:                            256 bytes
| 16384b
  Clock rate:                                    1.10 GHz
| 16384
  Concurrent copy and execution:                No
| 32
  Integrated:                                    Yes
| 512
  Support host page-locked memory mapping:      Yes</pre>
| 256b
====Quadro FX 570====
| 1.24GHz
<pre>Device 1: "Quadro FX 570"
| Yes
  CUDA Driver Version:                          2.30
| No
  CUDA Runtime Version:                          2.30
| Yes
  CUDA Capability Major revision number:        1
|-
  CUDA Capability Minor revision number:        1
| GeForce GTX 285
  Total amount of global memory:                268107776 bytes
| 1GB
  Number of multiprocessors:                    2
| 30
  Number of cores:                              16
| 240
  Total amount of constant memory:              65536 bytes
| 16384b
  Total amount of shared memory per block:      16384 bytes
| 16384
  Total number of registers available per block: 8192
| 32
  Warp size:                                    32
| 512
  Maximum number of threads per block:          512
| 256b
  Maximum sizes of each dimension of a block:    512 x 512 x 64
| 1.48GHz
  Maximum sizes of each dimension of a grid:    65535 x 65535 x 1
| Yes
  Maximum memory pitch:                          262144 bytes
| No
  Texture alignment:                            256 bytes
| Yes
  Clock rate:                                    0.92 GHz
|-
  Concurrent copy and execution:                Yes
| GeForce GTX 280
  Integrated:                                    No
| 1GB
  Support host page-locked memory mapping:      No</pre>
| 30
 
| 240
====GeForce 9800 GTX/9800 GTX+====
| 16384b
<pre>Device 0: “GeForce 9800 GTX/9800 GTX+”
| 16384
  CUDA Capability Major revision number:        1
| 32
  CUDA Capability Minor revision number:        1
| 512
  Total amount of global memory:                536608768 bytes
| 256b
  Number of multiprocessors:                    16
| 1.30GHz
  Number of cores:                              128
| Yes
  Total amount of constant memory:              65536 bytes
| No
  Total amount of shared memory per block:      16384 bytes
| Yes
  Total number of registers available per block: 8192
|-
  Warp size:                                    32
| GeForce GTX 260
  Maximum number of threads per block:          512
| 1GB
  Maximum sizes of each dimension of a block:    512 x 512 x 64
| 27
  Maximum sizes of each dimension of a grid:    65535 x 65535 x 1
| 216
  Maximum memory pitch:                          262144 bytes
| 16384b
  Texture alignment:                            256 bytes
| 16384
  Clock rate:                                    1.67 GHz
| 32
  Concurrent copy and execution:                Yes
| 512
  Integrated:                                    Yes
| 256b
  Support host page-locked memory mapping:      Yes</pre>
| 1.47GHz
====GeForce 9600 GT====
| Yes
<pre>Device 0: “GeForce 9600 GT”
| No
  CUDA Driver Version:                          2.30
| Yes
  CUDA Runtime Version:                          2.30
|-
  CUDA Capability Major revision number:        1
! COLSPAN="13" style="background:#efefef;" | Compute capability 1.2
  CUDA Capability Minor revision number:        1
|-
  Total amount of global memory:                536543232 bytes
| GeForce GT 360M
  Number of multiprocessors:                    8
| 1GB
  Number of cores:                              64
| 12
  Total amount of constant memory:              65536 bytes
| 96
  Total amount of shared memory per block:      16384 bytes
| 16384b
  Total number of registers available per block: 8192
| 16384
  Warp size:                                    32
| 32
  Maximum number of threads per block:          512
| 512
  Maximum sizes of each dimension of a block:    512 x 512 x 64
| 256b
  Maximum sizes of each dimension of a grid:    65535 x 65535 x 1
| 1.32GHz
  Maximum memory pitch:                          262144 bytes
| Yes
  Texture alignment:                            256 bytes
| No
  Clock rate:                                    1.50 GHz
| Yes
  Concurrent copy and execution:                Yes
|-
  Integrated:                                    No
| GeForce 310
  Support host page-locked memory mapping:      No</pre>
| 512MB
Other clock rates: 1.62GHz.
| 2
 
| 16
====GeForce 9400M====
| 16384b
<pre>Device 0: "GeForce 9400M"
| 16384
  Major revision number:                        1
| 32
  Minor revision number:                        1
| 512
  Total amount of global memory:                266010624 bytes
| 256b
  Number of multiprocessors:                    2
| 1.40GHz
  Number of cores:                              16
| Yes
  Total amount of constant memory:              65536 bytes
| No
  Total amount of shared memory per block:      16384 bytes
| Yes
  Total number of registers available per block: 8192
|-
  Warp size:                                    32
| GeForce 240 GT
  Maximum number of threads per block:          512
| 1GB
  Maximum sizes of each dimension of a block:    512 x 512 x 64
| 12
  Maximum sizes of each dimension of a grid:    65535 x 65535 x 1
| 96
  Maximum memory pitch:                          262144 bytes
| 16384b
  Texture alignment:                            256 bytes
| 16384
  Clock rate:                                    0.80 GHz
| 32
  Concurrent copy and execution:                No</pre>
| 512
====GeForce 8800 GTS 512====
| 256b
<pre>Device 0: "GeForce 8800 GTS 512"
| 1.424GHz
  Major revision number:                        1
| Yes
  Minor revision number:                        1
| No
  Total amount of global memory:                536150016 bytes
| Yes
  Number of multiprocessors:                    16
|-
  Number of cores:                              128
! COLSPAN="13" style="background:#efefef;" | Compute capability 1.1
  Total amount of constant memory:              65536 bytes
|-
  Total amount of shared memory per block:      16384 bytes
| ION
  Total number of registers available per block: 8192
| 256MB
  Warp size:                                    32
| 2
  Maximum number of threads per block:          512
| 16
  Maximum sizes of each dimension of a block:    512 x 512 x 64
| 16384b
  Maximum sizes of each dimension of a grid:    65535 x 65535 x 1
| 8192
  Maximum memory pitch:                          262144 bytes
| 32
  Texture alignment:                            256 bytes
| 512
  Clock rate:                                    1.62 GHz
| 256b
  Concurrent copy and execution:                Yes
| 1.1GHz
  Integrated:                                    No
| No
  Support host page-locked memory mapping:      No</pre>
| Yes
 
| Yes
====GeForce 8600 GT====
|-
<pre>Device 0: "GeForce 8600 GT"
| Quadro FX 570
  Major revision number:                        1
| 256MB
  Minor revision number:                        1
| 2
  Total amount of global memory:                268107776 bytes
| 16
  Total amount of constant memory:              65536 bytes
| 16384b
  Total amount of shared memory per block:      16384 bytes
| 8192
  Total number of registers available per block: 8192
| 32
  Warp size:                                    32
| 512
  Maximum number of threads per block:          512
| 256b
  Maximum sizes of each dimension of a block:    512 x 512 x 64
| 0.92GHz
  Maximum sizes of each dimension of a grid:    65535 x 65535 x 1
| Yes
  Maximum memory pitch:                          262144 bytes
| No
  Texture alignment:                            256 bytes
| No
  Clock rate:                                    1674000 kilohertz</pre>
|-
====GeForce 8600M GT====
| GeForce GTS 250 (*JR)
<pre>Device 0: "GeForce 8600M GT"
| 1G
  CUDA Driver Version:                          2.30
| 16
  CUDA Runtime Version:                          2.30
| 128
  CUDA Capability Major revision number:        1
| 16384b
  CUDA Capability Minor revision number:        1
| 8192
  Total amount of global memory:                267714560 bytes
| 32
  Number of multiprocessors:                    4
| 512
  Number of cores:                              32
| 256b
  Total amount of constant memory:              65536 bytes
| 1.84GHz
  Total amount of shared memory per block:      16384 bytes
| Yes
  Total number of registers available per block: 8192
| No
  Warp size:                                    32
| No
  Maximum number of threads per block:          512
|-
  Maximum sizes of each dimension of a block:    512 x 512 x 64
| GeForce 9800 GTX
  Maximum sizes of each dimension of a grid:    65535 x 65535 x 1
| 512MB
  Maximum memory pitch:                          262144 bytes
| 16
  Texture alignment:                            256 bytes
| 128
  Clock rate:                                    0.95 GHz
| 16384b
  Concurrent copy and execution:                Yes
| 8192
  Integrated:                                    No
| 32
  Support host page-locked memory mapping:      No</pre>
| 512
 
| 256b
====PNY GeForce 8400 GS (PCI)====
| 1.67GHz
<pre>Device 0: "GeForce 8400 GS"
| Yes
  CUDA Driver Version:                          2.30
| Yes
  CUDA Runtime Version:                          2.30
| Yes
  CUDA Capability Major revision number:        1
|-
  CUDA Capability Minor revision number:        1
| GeForce 9600 GT
  Total amount of global memory:                536608768 bytes
| 512MB
  Number of multiprocessors:                    1
| 8
  Number of cores:                              8
| 64
  Total amount of constant memory:              65536 bytes
| 16384b
  Total amount of shared memory per block:      16384 bytes
| 8192
  Total number of registers available per block: 8192
| 32
  Warp size:                                    32
| 512
  Maximum number of threads per block:          512
| 256b
  Maximum sizes of each dimension of a block:    512 x 512 x 64
| 1.62GHz,
  Maximum sizes of each dimension of a grid:    65535 x 65535 x 1
1.50GHz
  Maximum memory pitch:                          262144 bytes
| Yes
  Texture alignment:                            256 bytes
| No
  Clock rate:                                    1.40 GHz
| No
  Concurrent copy and execution:                No
|-
  Integrated:                                    No
| GeForce 9400M
  Support host page-locked memory mapping:      No</pre>
| 256MB
| 2
| 16
| 16384b
| 8192
| 32
| 512
| 256b
| 0.88GHz
| No
| No
| No
|-
| GeForce 8800 GTS 512
| 512MB
| 16
| 128
| 16384b
| 8192
| 32
| 512
| 256b
| 1.62GHz
| Yes
| No
| No
|-
| GeForce 8600 GT
| 256MB
| 4
| 32
| 16384b
| 8192
| 32
| 512
| 256b
| 0.95GHz
| Yes
| No
| No
|-
| GeForce 9400M
| 512MB
| 1
| 8
| 16384b
| 8192
| 32
| 512
| 256b
| 1.40GHz
| No
| No
| No
|-
|}
(*CB) Thanks to Cameron Black for this submission!
(*JR) Thanks to Javier Ruiz for this submission!


==See Also==
==See Also==
Line 663: Line 881:
* The [http://code.google.com/p/gpuocelot/ gpuocelot] project, hosted on Google Code.
* The [http://code.google.com/p/gpuocelot/ gpuocelot] project, hosted on Google Code.
* The NVIDIA [http://developer.nvidia.com/object/gpucomputing.html GPU Developer Zone]
* The NVIDIA [http://developer.nvidia.com/object/gpucomputing.html GPU Developer Zone]
* My [[CUBAR]] tools
* My [[CUBAR]] tools and reverse-engineered [[libcudest]]
[[CATEGORY: GPGPU]]
Retrieved from "https://nick-black.com/dankwiki/index.php?title=CUDA"