SSDi8 Achieves Efficient 8-bit Quantization for Mamba-2 State Space Models

Hyunwoo Kim, Byoungchan Ko, Minseok Kang, Minwoo Kim, Dongjin Lee, Jaehoon Lee, Sungroh Yoon, Dahuin Jung· August 25, 2026 View original

Key takeaways

  • SSDi8 is the first 8-bit quantization framework for Mamba-2's Structured State Space Duality (SSD).
  • It decouples multiplications and reuses quantized activations for significant latency reduction.
  • The framework maintains FP16-comparable accuracy while achieving up to 1.4x speedup.
  • SSDi8 enables efficient deployment of Mamba-2 models on resource-constrained devices.

Who benefits

Edge AIAutomotiveRoboticsConsumer ElectronicsCloud Computing

Summary

SSDi8 is the first post-training 8-bit quantization framework specifically designed for Mamba-2's Structured State Space Duality (SSD) architecture. It maintains FP16-comparable accuracy while significantly reducing memory and latency overhead, enabling efficient deployment of these advanced sequence models.

Mamba, a state space architecture, has emerged as a strong alternative to Transformers for efficient long-range dependency modeling. Its successor, Mamba-2, further enhances this with Structured State Space Duality (SSD), integrating both recurrent and attention modes for improved efficiency and scalability. However, this architectural complexity introduces substantial memory and latency demands, highlighting the need for specialized compression. SSDi8 is introduced as the first post-training 8-bit quantization framework tailored for SSD. Its core innovation lies in reformulating the architecture to decouple element-wise and matrix multiplications, allowing for the reuse of quantized activations across different modules. This design significantly reduces latency. To maintain high accuracy, SSDi8 adaptively quantizes channel-varying activations at optimal points and leverages the intrinsic dimensional decomposition of SSD, accounting for distinct outlier distributions. It also incorporates an error correction term based on per-channel error statistics. Comprehensive experiments show SSDi8 achieves accuracy comparable to FP16 while delivering up to 1.4x speedup in W4A8 and W8A8 settings, proving its robustness for resource-constrained environments like the Orin NX device.

Why it matters

For AI engineers and product developers, SSDi8 offers a crucial method to deploy advanced Mamba-2 models more efficiently on edge devices or in resource-constrained environments. This enables faster inference and lower operational costs without significant accuracy loss.

How to implement this in your domain

  1. 1Evaluate the performance and resource consumption of Mamba-2 models in your current AI inference pipelines.
  2. 2Investigate post-training quantization techniques like SSDi8 to optimize model deployment on edge or resource-limited hardware.
  3. 3Experiment with 8-bit quantization for state space models to achieve speedups and memory reductions.
  4. 4Benchmark quantized models on target hardware (e.g., Orin NX) to validate real-world performance gains and accuracy.

Original post by Hyunwoo Kim, Byoungchan Ko, Minseok Kang, Minwoo Kim, Dongjin Lee, Jaehoon Lee, Sungroh Yoon, Dahuin Jung

"arXiv:2608.21952v1 Announce Type: new Abstract: Recent advances in sequence modeling have highlighted Mamba as a state space architecture offering efficient long-range dependency modeling and providing a viable alternative to Transformers. Building upon this, Mamba-2 introduces t…"

View on X

Originally posted by Hyunwoo Kim, Byoungchan Ko, Minseok Kang, Minwoo Kim, Dongjin Lee, Jaehoon Lee, Sungroh Yoon, Dahuin Jung on X · view source

Want to go deeper?

Turn these trends into skills with Learnijoy's hands-on AI & tech courses.

Explore courses