For the complete documentation index, see llms.txt. Markdown versions of all pages are available by appending .md to any URL (e.g. /get-started.md).

Mojo struct

GPUInfo

struct GPUInfo

Comprehensive information about a GPU architecture.

This struct contains detailed specifications about GPU capabilities, including compute units, memory, thread organization, and performance characteristics.

Fields

• ​name (StaticString): The model name of the GPU.
• ​api (StaticString): The graphics/compute API the GPU is programmed through. Each vendor has its own API, so this also identifies the vendor: "cuda" for NVIDIA, "hip" for AMD, "metal" for Apple, and "none" for NoGPU. A stdlib plugin contributes its own name for hardware the stdlib has no built-in knowledge of. Comparisons must use these exact spellings.
• ​arch_name (StaticString): The architecture name of the GPU (e.g., sm_80, gfx942).
• ​compute (Float32): Compute capability version number for NVIDIA GPUs.
• ​version (StaticString): Version string of the GPU architecture.
• ​sm_count (Int): Number of streaming multiprocessors (SMs) on the GPU.
• ​warp_size (Int): Number of threads in a warp/wavefront.
• ​threads_per_multiprocessor (Int): Maximum number of threads per streaming multiprocessor.
• ​shared_memory_per_multiprocessor (Int): Size of shared memory available per multiprocessor in bytes.
• ​max_registers_per_block (Int): Maximum number of registers that can be allocated to a thread block.
• ​max_thread_block_size (Int): Maximum number of threads allowed in a thread block.

Implemented traits

AnyType, Copyable, Deinitable, Equatable, Movable, RegisterPassable, Writable

Methods

__eq__

def __eq__(self, other: Self) -> Bool

Checks if two GPUInfo instances represent the same GPU model.

Args:

• ​other (Self): Another GPUInfo instance to compare against.

Returns:

Bool: True if both instances represent the same GPU model.

target

def target(self) -> __mlir_type.`!kgen.target`

Gets the MLIR target configuration for this GPU.

Returns:

__mlir_type.`!kgen.target`: MLIR target configuration for the GPU.

from_target

static def from_target[target: __mlir_type.`!kgen.target`]() -> Self

Creates a GPUInfo instance from an MLIR target.

Parameters:

• ​target (__mlir_type.`!kgen.target`): MLIR target configuration.

Returns:

Self: GPU info corresponding to the target.

from_name

static def from_name[name: StringSpan[ImmStaticOrigin]]() -> Self

Creates a GPUInfo instance from a GPU architecture name.

Parameters:

• ​name (StringSpan[ImmStaticOrigin]): GPU architecture name (e.g., "sm_80", "gfx942").

Returns:

Self: GPU info corresponding to the architecture name.

from_family

static def from_family(family: AcceleratorArchitectureFamily, name: StringSpan[ImmStaticOrigin], api: StringSpan[ImmStaticOrigin], arch_name: StringSpan[ImmStaticOrigin], compute: Float32, version: StringSpan[ImmStaticOrigin], sm_count: Int) -> Self

Creates a GPUInfo instance using architecture family defaults.

This constructor simplifies GPU definition by inheriting common characteristics from an architecture family while allowing specific values to be overridden.

Args:

• ​family (AcceleratorArchitectureFamily): Architecture family providing default values.
• ​name (StringSpan[ImmStaticOrigin]): The model name of the GPU.
• ​api (StringSpan[ImmStaticOrigin]): The graphics/compute API supported by the GPU.
• ​arch_name (StringSpan[ImmStaticOrigin]): The architecture name of the GPU.
• ​compute (Float32): Compute capability version number.
• ​version (StringSpan[ImmStaticOrigin]): Version string of the GPU architecture.
• ​sm_count (Int): Number of streaming multiprocessors.

Returns:

Self: A fully configured GPUInfo instance.

write_to

def write_to(self, mut writer: T)

Writes GPU information to a writer.

Outputs all GPU specifications and capabilities to the provided writer in a human-readable format.

Args:

• ​writer (T): A Writer instance to output the GPU information.