pub struct Device {Show 21 fields
pub id: String,
pub name: String,
pub class: String,
pub arch: String,
pub compute_capability: String,
pub sms: u32,
pub memory_bytes: f64,
pub memory_type: String,
pub bandwidth_gbps: f64,
pub l2_mb: Option<f64>,
pub smem_per_sm_kb: u32,
pub cluster_max: u32,
pub tmem_per_sm_kb: u32,
pub mma_family: MmaFamily,
pub native: BTreeSet<Instr>,
pub peaks: Peaks,
pub measured: Option<String>,
pub usable_fraction: f64,
pub usable_why: String,
pub derived: Vec<String>,
pub sources: Vec<String>,
}Expand description
2026-09-30: One SKU.
Fields§
§id: String2026-09-30: h100-sxm.
name: String2026-09-30: Marketing name.
class: String2026-09-30: Kernel class, kernels/<class>/.
arch: String2026-09-30: Build arch (sm_90a), equal to the class’s HARDWARE.toml.
compute_capability: String2026-09-30: Compute capability.
sms: u322026-09-30: SM count.
memory_bytes: f642026-09-30: Visible device memory, bytes (memory_gib_visible, else the datasheet GB,
which tracks the visible GiB).
memory_type: String2026-09-30: Memory technology.
bandwidth_gbps: f642026-09-30: Datasheet DRAM bandwidth, GB/s.
l2_mb: Option<f64>2026-09-30: L2, MB, when published.
smem_per_sm_kb: u322026-09-30: Shared memory per SM, KB.
cluster_max: u322026-09-30: Largest portable thread-block cluster.
tmem_per_sm_kb: u322026-09-30: Tensor memory per SM, KB (tcgen05 accumulators); 0 without.
mma_family: MmaFamily2026-09-30: The family fast kernels are written for.
native: BTreeSet<Instr>2026-09-30: Instruction kinds executed natively (mma_family x the native_mma pairs).
peaks: Peaks2026-09-30: Dense datasheet peaks.
measured: Option<String>2026-09-30: Where measured achievable figures live, when measured.
usable_fraction: f642026-09-30: Share of memory the planner may use.
usable_why: String2026-09-30: Why that share.
derived: Vec<String>2026-09-30: Keys computed by the research, not published.
sources: Vec<String>2026-09-30: Source ids (S1, …), resolved in Registry::sources.