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

Mojo function

clusters_per_wave

def clusters_per_wave[cluster_size: Int, sm_count: Int]() -> Int

Number of size-cluster_size thread-block clusters that fit on the target Blackwell datacenter GPU in ONE wave, honoring GPC co-residency.

A cluster's CTAs must all live inside a single GPC, so a size-C cluster occupies C SMs within one GPC and only floor(gpc_sm / C) such clusters fit per GPC. The usable count is therefore below the flat sm_count / C; it is the per-GPC histogram sum_g floor(gpc_sm[g] / C). GPC layout is not queryable (neither DeviceAttribute nor GPUInfo exposes it), so it is hardcoded per chip; the whole expression folds at comptime since both parameters are comptime.

B200 (148 SMs) has 11 GPCs with an irregular layout -- three 20s, four 18s, one 10, three 2s (2x the pair counts 10,10,10,9,9,9,9,5,1,1,1). B300 / Blackwell Ultra (160 SMs) is the full die with 8 GPCs of a UNIFORM 20 SMs each, so the histogram collapses to 8 * (20 // C).

Non-increasing in cluster_size, so scanning candidate sizes largest-first yields the largest that fits. Only B200 (148) and B300 (160) are modeled; the else branch is a comptime error on any other chip.

Returns:

Int