CUDA(3) — AABS

🔍Understanding Allocated Active Blocks per SM (AABS)

① What Is AABS?

AABS (Allocated Active Blocks per SM) refers to the number of thread blocks that are simultaneously allocated and actively running within a single Streaming Multiprocessor (SM).
It represents how the GPU hardware distributes parallel work units (blocks) across SMs.

Formula:
AABS = min(BlockLimit, RegisterLimit, SharedMemoryLimit, WarpLimit)

That means the number of active blocks per SM is determined by the most restrictive resource among the four limits.

---

② Resource Constraints

Figure 1. Example result of the ncu (Nsight Compute) command on NVIDIA RTX A6000 (GA102, Ampere Architecture).

Limiting Factor	Description	Example (A100)
Block limit	Architectural maximum number of blocks per SM	32 blocks
Register limit	Limited register file per SM, affected by registers per thread	65,536 registers total
Shared memory limit	SM shared memory divided among active blocks	164 KB total
Warp limit	Depends on block size (warps per block × blocks per SM ≤ 64)	64 warps

The smallest of these values defines how many blocks can coexist on an SM.

---

③ Relationship to Occupancy

Occupancy is the ratio of active warps to maximum warps per SM:

Occupancy = Active Warps / Max Warps per SM

Since each block contains several warps, the number of active blocks directly determines active warps — hence occupancy.

Registers / Shared Memory / Warp Limit / Block Limit
                ↓
     Allocated Active Blocks per SM
                ↓
         Active Warps per SM
                ↓
          Achieved Occupancy

In short: resource limits → blocks → warps → occupancy.

---

④ Example Calculation

For NVIDIA A100:

• Total registers: 65,536 per SM
• Shared memory: 164 KB per SM
• Max warps per SM: 64
• Block size: 256 threads (8 warps)
• Each thread uses 48 registers
• Each block uses 24 KB shared memory

Factor	Calculation	Limit
Registers	65,536 / (256×48) ≈ 5 blocks	5
Shared Memory	164 KB / 24 KB = 6 blocks	6
Warp	64 / 8 = 8 blocks	8
Block Limit	32 blocks	32
Result	min(5,6,8,32) = 5 blocks/SM	✅

Thus:
Active Warps = 5 × 8 = 40,
Occupancy = 40 / 64 = 62.5%

---

⑤ Why It Matters

Higher AABS values mean more concurrent blocks → more active warps → higher potential occupancy.
However, just like occupancy, more is not always better — excessive register or memory pressure may cause performance drops.

⚠️ Optimizing AABS involves balancing block size, register usage, and shared memory to maximize useful concurrency.

---

Summary

Concept	Description
AABS	Number of active blocks allocated per SM
Determined by	Minimum of {Block, Register, Shared Memory, Warp limits}
Related to	Occupancy (active warps / max warps)
Goal	Maximize concurrency without resource exhaustion
Tool	Nsight Compute → sm__active_blocks_per_multiprocessor

---

References

• NVIDIA CUDA Programming Guide, v12.4
• Nsight Compute Metrics: sm__active_blocks_per_multiprocessor, sm__warps_active.avg.pct_of_peak_sustained_active
• NVIDIA A100 Whitepaper — Architecture and Resource Partitioning