C.W.K.
Stream
Lesson 04 of 05 · published

Why SSMs Became Exciting Again

~12 min · hardware, production, ibm, nvidia

Level 0Observer
0 XP0/50 lessons0/14 achievements
0/100 XP to next level100 XP to go0% complete

The hardware finally caught up

SSMs as a deep-learning idea aren't new — they were tried in the 2010s and didn't go anywhere. What changed is the hardware. Modern GPUs have enormous matrix-multiply throughput and large fast memory hierarchies (HBM + SRAM). The dual training/inference trick — FFT convolution at training, recurrence at inference — turned out to map beautifully onto that hardware once the kernels were written. S4 was the moment SSMs became practical to scale.

The convolution view works because the linear recurrence h_t = Ā·h_{t-1} + B̄·x_t can be expanded into a single global convolution kernel K = (CB, CAB, CA²B, ..., CA^(n-1)B). Applying K to the input via FFT gives O(n log n) training, fully parallel. The recurrence view is what you use at inference: cheap, sequential, constant-memory. Same parameters, two execution patterns.

Production deployments are the proof

The 2024–2026 wave of production deployments is what should change your mental model from "SSMs are research" to "SSMs ship". IBM Granite 4.0 uses a 9:1 Mamba-to-attention ratio in its production-grade open-weights models, achieving 70%+ RAM reduction over comparable pure Transformers. NVIDIA Nemotron-H ships 8B and 56B models with ~92% of attention layers replaced by Mamba-2. AI21 Jamba is on AWS Bedrock, Azure, and GCP with effective 256K context.

None of these are pure SSMs. They're hybrids. But the SSM layers are doing the heavy lifting — the attention layers are minority residents whose job is to recover the recall capability that pure SSMs can't deliver. The productization story validates the SSM bet: the architecture works at scale, in production, with real customers paying real money.

External links

Exercise

Pull the IBM Granite 4.0 model card from Hugging Face and skim the architecture section. Note the exact ratio of Mamba layers to attention layers and the layer count. Then look up Jamba's ratio and Nemotron-H's ratio. You'll find they cluster in the 8:1 to 9:1 range. That convergence isn't a coincidence; it's the empirical answer the field landed on for the recall-vs-efficiency tradeoff at production scale.

Progress

Progress is local-only — sign in to sync across devices.
Spotted a bug or have feedback on this page?Report an Issue

Comments 0

🔔 Reply notifications (sign in)
Sign inPlease sign in to comment.

No comments yet — be the first.