Struct CudaArchitecture 

pub struct CudaArchitecture {
    pub version: u32,
    pub tensor_cores: bool,
}

Fields

version: u32

tensor_cores: bool

Compute capability cannot say this: TU116 and TU117 report 7.5 like every other Turing, and mma.sync still runs on them, on the FP16 pipeline at a fiftieth the rate.

Implementations

impl CudaArchitecture

pub fn has_tensor_cores(version: u32, name: &str) -> bool

Decided by the marketing name because CUDA exposes no attribute for it. No GTX-branded Turing has tensor cores and every Turing that does shipped under another brand, so the two coincide exactly at 7.5. The professional TU117 parts (T400 to T1000) are not caught.

Trait Implementations

impl Architecture for CudaArchitecture

fn warp_size(&self) -> u32

fn is_wmma_capable(&self) -> bool

fn is_mfma_capable(&self) -> bool

fn get_version(&self) -> u32

impl Debug for CudaArchitecture

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl Display for CudaArchitecture

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.