Research map / Architectures and efficiency / Quantization
SSDi8: Accurate and Efficient 8-bit Quantization for State Space Duality
Hyunwoo Kim, Byoungchan Ko, Minseok Kang, Minwoo Kim, Dongjin Lee, Jaehoon Lee, Sungroh Yoon, Dahuin Jung
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SSDi8 instead of Quamba2 at W4A8 | Mamba-2 2.7B · six zero-shot tasks avg | accuracy: 62.1% -> 62.7% | Improved |
| Use SSDi8 instead of Quamba2 at W8A8 | Mamba-2 8B · six zero-shot tasks avg | accuracy: 69.8% -> 70.2% | Improved |
| Add SSD quantization to HadMamba (ablation) | Mamba-2 2.7B W4A8 · LAMBADA | accuracy: 51.2% -> 67.2% | Improved |
| Add mean correction (ablation) | SSDi8 W4A8, Mamba-2 2.7B · LAMBADA | accuracy: 67.2% -> 67.4% | Improved |
| Quantize B and C once and reuse (ablation) | SSD W4A8 with ChunkState Q(X) · SSD latency | latency (ms): 8.58 -> 8.05 | Improved |
| Add ChunkBMM quantization on top of reformulation and persistent INT8 (ablation) | SSDi8 W4A8 · SSD latency | latency (ms): 8.00 -> 6.53 | Improved |
| Use sparse-aware reformulation instead of quantizing X alone (ablation) | SSDi8 W4A8 · WikiText2 | perplexity: 9.35 -> 9.36 | No clear effect |
| Quantize X and State per head (H,G) together with P instead of per-T (ablation) | SSD W4A8 · WikiText2 | perplexity: 11.97 -> 7.42 | Improved |
| Naively quantize SSD per-tensor after in-projection (ablation) | Mamba-2 2.7B W4A8 · six zero-shot tasks avg | accuracy: 63.6% -> 58.4% | Hurt |
| Apply INT8 to the SSD path only | Nemotron-H-8B-Reasoning · SSD module latency | latency (ms): 19.834 -> 9.156 | Improved |
| Apply INT8 to the SSD path only | Nemotron-H-8B-Reasoning · average zero-shot accuracy | accuracy: 73.1% -> 73.0% | No clear effect |
| Use SSDi8 instead of Quamba2 on edge device | Mamba-2 2.7B W8A8 · Orin NX SSD latency L=2048 | latency (ms): 249.29 -> 217.69 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- Quamba2 — Quamba: A Post-Training Quantization Recipe for Selective State Space Models (ICLR 2025) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Quamba: A Post-Training Quantization Recipe for Selective State Space Models (ICLR 2025 · similarity 0.88)
- Quamba2: A Robust and Scalable Post-training Quantization Framework for Selective State Space Models (ICML 2025 · similarity 0.86)
- ScaleBITS: Scalable Bitwidth Search for Hardware-Aligned Mixed-Precision LLMs (NeurIPS 2026 · similarity 0.85)
- SnapMLA: Efficient Long-Context MLA Decoding via Hardware-Aware FP8 Quantized Pipelining (NeurIPS 2026 · similarity 0.85)
- Block-Biased Mamba for Long-Range Sequence Processing (NeurIPS 2025 · similarity 0.85)
- Transformers are SSMs: Generalized Models and Efficient Algorithms Through Structured State Space Duality (ICML 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
