RELEASEMay 27, 2016
FOUNDRYTSMC
PROCESS16 nm
ARCHITECTUREPascal
ROPs64
TMUs160
SPs2,560CUDA
VRAM8GBGDDR5X / GDDR5
TDP180W
DIE FLOORPLANSCHEMATIC · NOT TO SCALE
GP104-895-A1 · 16 nm
area 314 mm²
- GPC / SM
- 4 / 20
- CUDA
- 2 560
- L2 cache
- 2 MB
- Memory controller
- 8 × 32-bit
- Media engine
- NVENC · NVDEC
SECTIONSclick a block or section
General
13Identity
ManufacturerNVIDIA
ArchitecturePascal
NameGP104-895-A1
CodenameNV134
Package
PackageFCBGA-2152
Package size37.5 mm × 37.5 mm
Die protectionBare Die
Silicon
FoundryTSMC
Process node16 nm
Process typeFinFET
Transistors7,200M
Density22.9M/mm²
Die size314mm²
Configuration
13Compute blocks
Compute units (SM/CU)20
GPCs / Shader engines4
TPCs / WGPs20
Execution units (FP32)2,560CUDA
TMUs160
ROPs64
Power
Max TDP180W
Specialised units
FP16 units20
FP64 units80
INT32 units2,560
SFUs640
Cache
L1 cache / Shared Memory per SM48KB
L2 cache2,048KB
Memory
03Max VRAM8GB
Max memory bus256bit
VRAM typeGDDR5XGDDR5
Clock speeds
03Base clock886MHz
Boost clock1,114MHz
Memory clock1,502MHz
APIs
11CUDA
6.1 (12.9)
Full
Vulkan
1.4
Full
OpenCL
3.0
Full
OpenGL
4.6
Full
DirectXFeature Level 12_1 (DirectX 12)
12.1
Full
Shader Model
6.8
Full
WDDMDESCRIPTIONWindows Display Driver Model — the Windows display driver model version
3.2
Full
VDPAUDESCRIPTIONVideo Decode and Presentation API for Unix — hardware decode feature set
Feature Set H
Full
DirectComputeDESCRIPTIONDirect3D compute shaders; the version is the compute shader model
5.0
Full
NVIDIA OptiXDESCRIPTIONNVIDIA ray-tracing API; the version is the newest OptiX SDK the chip supports
9.0
Full
DirectMLDESCRIPTIONMicrosoft machine-learning API on top of DirectX 12; runs on any DirectX 12 GPU
—
SUPPORT LEVEL✓ Full◐ Partial✕ None
Instructions
09IDP4ADESCRIPTION8-bit Integer Dot Product and Accumulate into 32 bits (PTX dp4a). From sm_61.
Full
IDP2ADESCRIPTION16-bit by 8-bit Integer Dot Product and Accumulate (PTX dp2a). From sm_61.
Full
FFMADESCRIPTIONSingle-Precision Fused Multiply-Add. From sm_20; sm_1x only had an unfused MAD.
Full
HFMA2DESCRIPTIONPacked fused multiply-add of two FP16 (or BF16) values per instruction (PTX fma.f16x2 / fma.bf16x2). FP16 from sm_53, BF16 from sm_80.FP16 (f16x2) only; packed BF16 (bf16x2) arrived with sm_80.
Partial
DFMADESCRIPTIONDouble-Precision Fused Multiply-Add. From sm_13; throughput depends on the number of FP64 units.
Full
LOP3.LUTDESCRIPTIONArbitrary three-input bitwise logic function from a truth table (Three-Input Bitwise Logic Operation). From sm_50.
Full
SHFLDESCRIPTIONExchanges registers between warp threads without shared memory (Warp Shuffle). From sm_30.
Full
ATOMG / ATOMSDESCRIPTIONHardware atomics on global (ATOMG) and shared (ATOMS) memory (PTX atom). Global from sm_11, shared from sm_12.
Full
MUFUDESCRIPTIONSpecial functions on the SFUs: sin, cos, ex2, lg2, rcp, rsqrt (Multi-Function Unit). On every CUDA architecture.
Full
SUPPORT LEVEL✓ Full◐ Partial✕ None
AI / Compute
21CUDA
Full
Vulkan
Full
OpenCL
Full
Floating-point formats
FP64Stream processors only
FP32Stream processors only
TF32None
FP16Stream processors only
BF16None
FP8None
MXFP8None
MXFP6None
FP4None
MXFP4None
NVFP4None
Integer formats
INT32Stream processors only
INT16Stream processors only
INT8Stream processors only
INT4None
INT2None
INT1None
Other
Async DMA / TMA✕ No
SUPPORT LEVEL✓ Full◐ Partial✕ None
Theoretical performance
04FP16 (half)1:64138.6GFLOPS
FP32 (single)8.873TFLOPS
FP64 (double)1:32277.3GFLOPS
INT8 (shaders)35.5TOPS
Media
04NVENCDESCRIPTIONHardware video encoder
6th Gen
NVDECDESCRIPTIONHardware video decoder
3rd Gen
PureVideo HDDESCRIPTIONNVIDIA hardware video decode feature set
VP8
AV1✕ No
Imagery
// 01// DIE FLOORPLANReference data. No buy prices, stock, or ads.UMTALE.LAB/DB · 81