Tile Lang Benchmark

Benchmark Summary

TileLang achieves exceptional performance across a variety of computational patterns. Below are selected results showcasing its capabilities:

• Flash Attention Performance on H100

  

• Matmul Performance on GPUs (RTX 4090, A100, H100, MI300X)

  

• Dequantize Matmul Performance on A100

  

Benchmark OP Set

---

Table 1: Matrix shapes in our benchmark

	V0	V1	V2	V3	V4	V5	V6	V7
m	1	1	1	1	1	1	1	1
n	16384	43008	14336	57344	14336	9216	36864	9216
k	16384	14336	14336	14336	57344	9216	9216	36864

	M0	M1	M2	M3	M4	M5	M6	M7
m	4096	4096	4096	4096	8192	8192	8192	8192
n	1024	8192	28672	8192	1024	8192	28672	8192
k	8192	8192	8192	28672	8192	8192	8192	28672

---

Table 2: FlashAttention shapes in our benchmark

	FA0	FA1	FA2	FA3	FA4
batch	1	1	1	1	1
nheads	32	32	32	32	32
seq_len	512	512	1024	1024	4096
head_dim	128	128	128	128	128
causal	true	false	true	false	true

---

Table 3: Linear Attention shapes in our benchmark

	CC0	CC1	CC2	CC3	CC4	CC5
batch	1	1	1	64	64	64
nheads	64	64	64	64	64	64
seq_len	1024	2048	8192	1024	2048	8192
head_dim	64	64	64	64	64	64
d_state	128	128	128	128	128	128

	CT0	CT1	CT2	CT3	CT4	CT5
batch	1	1	1	64	64	64
nheads	64	64	64	64	64	64
seq_len	1024	2048	8192	1024	2048	8192
head_dim	64	64	64	64	64	64
d_state	128	128	128	128	128	128

---