View source for CUDA

[[File:Gt200die-big.jpg|right|thumb|A "Fermi" GT200 die]]
==Hardware==
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
nvidia 0000:07:00.0: enabling device (0000 -> 0003)
nvidia 0000:07:00.0: PCI INT A -> GSI 21 (level, low) -> IRQ 21
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>
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.

==CUDA model==
===Host===
* A host contains zero or more CUDA-capable devices (emulation must be used if zero devices are available).
* It can run multiple CUDA processes, each composed of one or more host threads.
* A given host thread can execute code on only one device at once.
* Multiple host threads can execute code on the same device.
===Device===
* A device packages a streaming processor array (SPA), a memory interface, and possibly memory (''global memory''. ''device memory'').
** In CUDA terminology, an ''integrated'' (vs ''discrete'') device does not have its own global memory.
** Specially-prepared global memory is designated ''constant memory'', and can be [[#Streaming Multiprocessors|cached]].
* Pinned (''locked'') host memory avoids a bounce buffer, accelerating 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.
* 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)
* Write-combining memory (configured via [[MTRR|MTRRs]] or [[Page Attribute Tables|PATs]]) avoids PCI snoop requirements and maximizes linear throughput
** Subtle side-effects; not to be used glibly or carelessly!
* Distributes work at ''block'' granularity to [[#Texture Processing Cluster|Texture Processing Clusters]] (TPCs).
====Texture Processing Cluster====
[[#Streaming Multiprocessors|Streaming Multiprocessors]] (SMs) are grouped into TPCs. Each TPC contains some number of SMs and a single texture processing unit, including a few filters and a cache for texture memory. The details of these texture caches have not generally been publicized, but NVIDIA optimization guides confirm 1- and 2-dimensional spatial caching to be in effect.

===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).
** 8K registers for compute capability <= 1.1, otherwise
** 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.
** 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
** 256 registers per block for compute capability <= 1.1, otherwise
** 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.
** 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>).
* Thus, a given thread's &lt;blockId X threadId&gt; dyad is unique across the grid. All the threads of a block share a blockId, and corresponding threads of various blocks share a threadId.
* Each time the kernel is instantiated, new grid and block dimensions may be provided.
* A block's threads, starting from threadId 0, are broken up into contiguous '''warps''' having some '''warp size''' number of threads.
* Distributes out-of-order work at ''warp'' granularity across [[#Stream Processor|SPs]].
** One program counter per warp -- divergence within warp leads to serialization.
** Divergence is trivially supported with a per-warp stack; warps reconverge at [[Compiler Design|immediate post-dominators]] of branches
* Supports some maximum number of blocks and threads (~8 and ~768 on G80).
====Block sizing====
'''FIXME: review/verify this!'''

How tightly can we bound the optimal block size '''T''', given a warp size ''w''? The number of threads per block ought almost always be a multiple of ''w'', both to:
* facilitate coalescing (coalescing requirements are related to ''w/2''), and
* maximize utilization of SPs within warp-granular scheduling.
A SM has ''r'' registers and ''s'' words of shared memory, allocated per-block (see [[#Streaming Multiprocessors|above]]). Assuming that ''w'' threads can be supported (i.e., that none requires more than ''r/w'' registers or ''s/w'' words of shared memory), the most obvious lower bound is ''w'' itself. The most obvious upper bound, assuming arbitrary available work, is the greatest multiple of ''w'' supported by hardware (and, obviously, the SDK). A block must be scheduled to an SM, which requires:
* registers sufficient to support the block,
* shared memory sufficient to support the block,
* 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).
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'''
* larger maximum possible kernels (an absolute limit exists on grid dimensions)
* better if data can be reused among threads (e.g. in tiled matrix multiply)
Motivations for smaller blocks include:
* freedom in the ''t'' dimension exposes parallelism until ''t'' >= ''b'' * '''T'''
* freedom in the ''r'' and ''s'' dimensions exposes parallelism until ''r'' >= ''b'' * '''T''' * '''Thr<sub>reg</sub>''' or ''s'' >= ''b'' * '''Blk<sub>shmem</sub>'''.
* cheaper per-block operations(?) (<tt>__syncthreads()</tt>, voting, etc)
* support for older hardware and SDKs
* fairer distribution among SMs and thus possibly better utilization, lower latency
** relative speedup tends to 0 as work grows arbitrarily on finite SMs
** relative speedup tends to ''1/Frac<sub>par</sub>'' on infinitely many SMs
We can now optimize occupancy for a specific {''t'', ''b'', ''r'' and ''s''}, assuming ''t'' to be a multiple of both ''w'' and ''b'':
* Let '''T''' = ''t'' / ''b''. '''T''' is thus guaranteed to be the smallest multiple of ''w'' such that ''t'' == ''b'' * '''T'''.
* Check the ''r'' and ''w'' conditions. '''FIXME: handle reduction'''
* '''FIXME: handle very large (''external'') kernels'''
Optimizing for ranges of hardware values is left as an exercise for the reader. Occupancy is only worth optimizing if the number of warps are insufficient to hide latencies. It might be possible to eliminate latencies altogether by reusing data throughout a block via shared memory; if the algorithm permits, this is almost certainly a net win. In that case, we likely want to maximize '''Blk<sub>shmem</sub>'''. A more advanced theory would incorporate the arithmetic intensity of a kernel...'''FIXME'''

===Stream Processor===
* 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!
** 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. Full predication.
{| border="1"
! Memory type
! PTX name
! Sharing
! Kernel access
! Host access
! Cache location
! Adddressable
|-
| Registers
| .reg
| Per-thread
| Read-write
| None
| None
| No
|-
| Special registers
| .sreg
| varies
| Read-only
| None
| None
| No
|-
| Local memory
| .local
| Per-thread
| Read-write
| None
| None
| Yes
|-
| Shared memory
| .shared
| Per-block
| Read-write
| None
| None
| Yes
|-
| Global memory
| .global
| Global
| Read-write
| Read-write
| '''1.x''': None
'''2.0+''': L1 on SM, L2 on TPC(?)
| Yes
|-
| Constant memory
| .const
| Per-grid
| Read
| Read-write
| Stream multiprocessor
| Yes
|-
| Texture memory
| .tex
| Global
| Read
| Read-write
| Texture processing cluster
| texture API
|-
| Parameters (to grids or functions)
| .param
| Per-grid (or per-thread)
| Read-only (or read-write)
| None
| None
| Yes (or restricted)
|-
|}

===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 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"
! Revision
! Changes
|-
| 1.1
|
* Atomic ops on 32-bit global integers.
* Breakpoints and other debugging support.
|- 
| 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.
|-
| 1.3
|
* Double-precision floating point at 32 cycles per operation.
|-
| 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

|-
| 3.5
| 
* 255 registers per thread
* "CUDA Dynamic Parallelism", the ability to spawn threads from within device code
* ''PTX 3.1''
** A funnel shift instruction, <tt>shf</tt>
** Loading read-only global data through the non-coherent texture cache, <tt>ld.global.nc</tt>
** 64-bit atomic/reduction operators extended to {or, xor, and, integer min, integer max}
** Mipmap type support
** Indirect texture/surface support
** Extends generic addressing to include the const state space

|-
| 7.0
|
* ''PTX 6.3''
* Tensor cores
* Independent thread scheduling

|-
| 7.5
|
* ''PTX 6.4''
* Integer matrix multiplication in tensor cores
|-
|}

==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>

==Building CUDA Apps==
===nvcc flags===
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===
This assumes use of the SDK's common.mk, as recommended by the documentation.
* Add the library path to LD_LIBRARY_PATH, assuming CUDA's been installed to a non-standard directory.
* Set the <tt>CUDA_INSTALL_PATH</tt> and <tt>ROOTDIR</tt> (yeargh!) if outside the SDK.
* I keep the following in <tt>bin/cudasetup</tt> of my home directory. Source it, using sh's <tt>. cudasetup</tt> syntax:
<pre>CUDA="$HOME/local/cuda/"

export CUDA_INSTALL_PATH="$CUDA"
export ROOTDIR="$CUDA/C/common/"
if [ -n "$LD_LIBRARY_PATH" ] ; then
	export "LD_LIBRARY_PATH=$LD_LIBRARY_PATH:$CUDA/lib64"
else
	export "LD_LIBRARY_PATH=$CUDA/lib64"
fi

unset CUDA</pre>
* Set EXECUTABLE in your Makefile, and include <tt>$CUDA_INSTALL_PATH/C/common/common.mk</tt>
====Unit testing====
The <tt>DEFAULT_GOAL</tt> special variable of [[GNU Make]] can be used:
<pre>.PHONY: test
.DEFAULT_GOAL:=test

include $(CUDA_INSTALL_PATH)/C/common/common.mk

test: $(TARGET)
        $(TARGET)</pre>

==Libraries==
Two mutually exclusive means of driving CUDA are available: the "Driver API" and "C for CUDA" with its accompanying <tt>nvcc</tt> compiler and runtime. The latter (<tt>libcudart</tt>) is built atop the former, and requires its <tt>libcuda</tt> library.
===Undocumented Functions===
The following unlisted functions were extracted from 3.0's libcudart.so using <tt>objdump -T</tt>:<pre>00000000000097d0 g    DF .text	000000000000020e  Base        __cudaRegisterShared
0000000000005410 g    DF .text	0000000000000003  Base        __cudaSynchronizeThreads
0000000000009e60 g    DF .text	0000000000000246  Base        __cudaRegisterVar
000000000000a0b0 g    DF .text	0000000000000455  Base        __cudaRegisterFatBinary
00000000000095c0 g    DF .text	000000000000020e  Base        __cudaRegisterSharedVar
0000000000005420 g    DF .text	0000000000000002  Base        __cudaTextureFetch
000000000000a510 g    DF .text	00000000000009dd  Base        __cudaUnregisterFatBinary
00000000000099e0 g    DF .text	000000000000024e  Base        __cudaRegisterFunction
0000000000005820 g    DF .text	000000000000001c  Base        __cudaMutexOperation
0000000000009c30 g    DF .text	000000000000022e  Base        __cudaRegisterTexture</pre>

==deviceQuery info==
* Memory shown is that amount which is free; I've substituted total VRAM.
* 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!]


{| border="1"
! Device name
! Memory
! MP's
! Cores
! Shmem/block
! Reg/block
! Warp size
! Thr/block
! Texalign
! Clock
! C+E?
! Integrated?
! Shared maps?
|-
! COLSPAN="13" style="background:#eebeb6;" | Compute capability 7.0
|-
| Tesla V100 
| 16GB
| 84
| 5376/2688/672
| 
|
|
| 
| 
| 1.53GHz
| Yes
| No
| Yes
|-
! COLSPAN="13" style="background:#8070D8;" | Compute capability 3.0
|-
| GeForce GTX 680
| 1.5GB
| 8
| 1536
| 
|
|
|
|
|
| Yes
| No
| Yes
|-
! COLSPAN="13" style="background:#ffdead;" | Compute capability 2.1
|-
| GeForce GTX 560 Ti
|
|
|
|
|
|
|
|
|
|
|
|
|-
| GeForce GTX 550 Ti
|
|
|
|
|
|
|
|
|
|
|
|
|-
| GeForce GTX 460
| 1GB
| 7
| 224
| 48k
| 32k
| 32
| 1024
| 512b
| 1.35GHz
| Yes
| No
| Yes
|-
| GeForce GTS 450
|
|
|
|
|
|
|
|
|
|
|
|
|-
! COLSPAN="13" style="background:#ffdead;" | Compute capability 2.0
|-
| GeForce GTX 580
| 1.5GB
| 16
| 512
|
|
| 32
| 1024
|
| 1.544GHz
| Yes
| No
| Yes
|-
| Tesla C2050 (*CB)
| 3GB
| 14
| 448
| 48k
| 32k
| 32
| 1024
| 512b
| 1.15GHz
| Yes
| No
| Yes
|-
| Tesla C2070 (*CB)
| 6GB
| 14
| 448
| 48k
| 32k
| 32
| 1024
| 512b
| 1.15GHz
| Yes
| No
| Yes
|-
| GeForce GTX 480
| 1536MB
| 15
| 480
| 
| 
| 
| 
| 
| 
| 
| 
|
|-
| GeForce GTX 470
| 1280MB
| 14
| 448
| 
| 
| 
| 
| 
| 
| 
| 
|
|-
! COLSPAN="13" style="background:#efefef;" | Compute capability 1.3
|-
| Tesla C1060
| 4GB
| 30
| 240
| 16384b
| 16384
| 32
| 512
| 256b
| 1.30GHz
| Yes
| No
| Yes
|-
| GeForce GTX 295
| 1GB
| 30
| 240
| 16384b
| 16384
| 32
| 512
| 256b
| 1.24GHz
| Yes
| No
| Yes
|-
| GeForce GTX 285
| 1GB
| 30
| 240
| 16384b
| 16384
| 32
| 512
| 256b
| 1.48GHz
| Yes
| No
| Yes
|-
| GeForce GTX 280
| 1GB
| 30
| 240
| 16384b
| 16384
| 32
| 512
| 256b
| 1.30GHz
| Yes
| No
| Yes
|-
| GeForce GTX 260
| 1GB
| 27
| 216
| 16384b
| 16384
| 32
| 512
| 256b
| 1.47GHz
| Yes
| No
| Yes
|-
! COLSPAN="13" style="background:#efefef;" | Compute capability 1.2
|-
| GeForce GT 360M
| 1GB
| 12
| 96
| 16384b
| 16384
| 32
| 512
| 256b
| 1.32GHz
| Yes
| No
| Yes
|-
| GeForce 310
| 512MB
| 2
| 16
| 16384b
| 16384
| 32
| 512
| 256b
| 1.40GHz
| Yes
| No
| Yes
|-
| GeForce 240 GT
| 1GB
| 12
| 96
| 16384b
| 16384
| 32
| 512
| 256b
| 1.424GHz
| Yes
| No
| Yes
|-
! COLSPAN="13" style="background:#efefef;" | Compute capability 1.1
|-
| ION
| 256MB
| 2
| 16
| 16384b
| 8192
| 32
| 512
| 256b
| 1.1GHz
| No
| Yes
| Yes
|-
| Quadro FX 570
| 256MB
| 2
| 16
| 16384b
| 8192
| 32
| 512
| 256b
| 0.92GHz
| Yes
| No
| No
|-
| GeForce GTS 250 (*JR)
| 1G
| 16
| 128
| 16384b
| 8192
| 32
| 512
| 256b
| 1.84GHz
| Yes
| No
| No
|-
| GeForce 9800 GTX
| 512MB
| 16
| 128
| 16384b
| 8192
| 32
| 512
| 256b
| 1.67GHz
| Yes
| Yes
| Yes
|-
| GeForce 9600 GT
| 512MB
| 8
| 64
| 16384b
| 8192
| 32
| 512
| 256b
| 1.62GHz,
1.50GHz
| Yes
| No
| No
|-
| GeForce 9400M
| 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==
* The nouveau Wiki's [http://nouveau.freedesktop.org/wiki/CUDA CUDA page]
** [http://nouveau.freedesktop.org/wiki/HonzaHavlicek Honza Havlicek]'s guide to NVIDIA architecture
** [http://nouveau.freedesktop.org/wiki/ContextSwitching Context switching] and the PSWITCH instruction
* 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]
* My [[CUBAR]] tools and reverse-engineered [[libcudest]]
[[CATEGORY: GPGPU]]