Memory-bandwidth efficiency measurement on Autoregressive decoding workload
64.08HBU (%)Mamba-2-JAX
Evaluation Results
| Method | Links | |
|---|---|---|
| Mamba-2-JAXModel=2.7B, Sequence Length=4096, Hardware=TPU v6e2026.03 | 64.08 | |
| Mamba-2-JAXModel=2.7B, Sequence Length=2048, Hardware=TPU v6e2026.03 | 64.06 | |
| Mamba-2-JAXModel=2.7B, Sequence Length=1024, Hardware=TPU v6e2026.03 | 64 | |
| Mamba-2-JAXModel=2.7B, Sequence Length=512, Hardware=TPU v6e2026.03 | 63.91 | |
| Mamba-2-JAXModel=2.7B, Sequence Length=256, Hardware=TPU v6e2026.03 | 63.74 | |
| Mamba-2-JAXModel=2.7B, Sequence Length=128, Hardware=TPU v6e2026.03 | 63.43 | |
| Mamba-2-JAXModel=780M, Sequence Length=4096, Hardware=TPU v6e2026.03 | 62.99 | |
| Mamba-2-JAXModel=780M, Sequence Length=2048, Hardware=TPU v6e2026.03 | 62.93 | |
| Mamba-2-JAXModel=780M, Sequence Length=1024, Hardware=TPU v6e2026.03 | 62.87 | |
| Mamba-2-JAXModel=780M, Sequence Length=512, Hardware=TPU v6e2026.03 | 62.8 | |
| Mamba-2-JAXModel=780M, Sequence Length=256, Hardware=TPU v6e2026.03 | 62.59 | |
| Mamba-2-JAXModel=780M, Sequence Length=128, Hardware=TPU v6e2026.03 | 62.07 | |
| Mamba-2-JAXModel=1.3B, Sequence Length=4096, Hardware=TPU v6e2026.03 | 61.87 | |
| Mamba-2-JAXModel=1.3B, Sequence Length=2048, Hardware=TPU v6e2026.03 | 61.86 | |
| Mamba-2-JAXModel=1.3B, Sequence Length=1024, Hardware=TPU v6e2026.03 | 61.77 | |
| Mamba-2-JAXModel=1.3B, Sequence Length=512, Hardware=TPU v6e2026.03 | 61.69 | |
| Mamba-2-JAXModel=1.3B, Sequence Length=256, Hardware=TPU v6e2026.03 | 61.55 | |
| Mamba-2-JAXModel=1.3B, Sequence Length=128, Hardware=TPU v6e2026.03 | 61.22 | |
| Mamba-2-JAXModel=370M, Sequence Length=2048, Hardware=TPU v6e2026.03 | 59.32 | |
| Mamba-2-JAXModel=370M, Sequence Length=4096, Hardware=TPU v6e2026.03 | 59.32 | |
| Mamba-2-JAXModel=370M, Sequence Length=1024, Hardware=TPU v6e2026.03 | 59.29 | |
| Mamba-2-JAXModel=370M, Sequence Length=512, Hardware=TPU v6e2026.03 | 59.07 | |
| Mamba-2-JAXModel=370M, Sequence Length=256, Hardware=TPU v6e2026.03 | 58.65 | |
| Mamba-2-JAXModel=370M, Sequence Length=128, Hardware=TPU v6e2026.03 | 57.88 | |
| Mamba-2-JAXModel=130M, Sequence Length=4096, Hardware=TPU v6e2026.03 | 53.32 | |
| Mamba-2-JAXModel=130M, Sequence Length=2048, Hardware=TPU v6e2026.03 | 53.26 | |
| Mamba-2-JAXModel=130M, Sequence Length=1024, Hardware=TPU v6e2026.03 | 53.13 | |
| Mamba-2-JAXModel=130M, Sequence Length=512, Hardware=TPU v6e2026.03 | 52.87 | |
| Mamba-2-JAXModel=130M, Sequence Length=256, Hardware=TPU v6e2026.03 | 52.29 | |
| Mamba-2-JAXModel=130M, Sequence Length=128, Hardware=TPU v6e2026.03 | 51.62 |