eMamba: Edge Acceleration Framework

• eMamba is an edge acceleration framework that deploys Mamba models on FPGA and ASIC platforms through algorithmic approximation, quantization, and hardware-software co-design.
• It replaces computationally intensive operations with hardware-friendly alternatives such as RangeNorm and ReLU, enabling faster and more efficient processing.
• Empirical evaluations show that eMamba significantly reduces latency, area, and power consumption compared to traditional accelerators, proving its effectiveness in edge computing.

eMamba most precisely denotes the framework introduced in "eMamba: Efficient Acceleration Framework for Mamba Models in Edge Computing", where it is defined as a comprehensive end-to-end hardware acceleration framework explicitly designed for deploying Mamba models on edge platforms (Kim et al., 14 Aug 2025). In adjacent Mamba literature, however, the same string or closely related variants are also used informally for event-centric Mamba systems, electron microscopy–focused Mamba systems, or a clinical SSM expert named EMamba rather than for the edge-deployment framework itself. The term therefore has a narrow canonical meaning and a broader, overloaded usage across recent arXiv papers (Chen et al., 2024, Yang et al., 21 Jan 2025, Shi et al., 2024, Gao et al., 1 Mar 2026, Khan et al., 3 Jan 2026).

1. Definition and terminological scope

The exact title "eMamba" refers to the edge-computing framework of Luo et al., whose central objective is to make Mamba practical on resource-constrained hardware through algorithmic approximation, quantization, and hardware-software co-design (Kim et al., 14 Aug 2025). That work is concerned with deployment: hardware-aware normalization, approximate nonlinearities, quantized recurrence, pipeline scheduling, and FPGA/ASIC realization.

The same label is not used uniformly elsewhere. Some papers explicitly frame Mamba variants for event data as an “eMamba” line, while others describe EM-oriented Mamba systems as fulfilling EM-focused goals associated with “eMamba.” A separate naming variant, EMamba, appears in a clinical Mixture-of-Experts model as the long-sequence SSM expert. This naming overlap is substantive, because these systems solve different problems and are not interchangeable architectures (Chen et al., 2024, Yang et al., 21 Jan 2025, Shi et al., 2024, Gao et al., 1 Mar 2026, Khan et al., 3 Jan 2026).

Term in literature	Meaning	Representative source
eMamba	Edge acceleration framework for Mamba on edge platforms	(Kim et al., 14 Aug 2025)
“eMamba” / Event Mamba	Event-centric Mamba for event-based vision	(Chen et al., 2024, Yang et al., 21 Jan 2025)
“eMamba” in EM context	EM-focused Mamba for electron microscopy	(Shi et al., 2024, Gao et al., 1 Mar 2026)
EMamba	Mamba-130M SSM expert in clinical MoE routing	(Khan et al., 3 Jan 2026)

2. State-space foundation and algorithmic modifications

The edge-deployment eMamba inherits the standard SSM view of Mamba. The paper states the standard form as

$$x_t = A\,x_{t-1} + B\,u_t,\quad y_t = C\,x_t + D\,u_t,$$

and uses input-dependent discrete-time versions with $\bar{A}_t$, $\bar{B}_t$, and $C_t$ derived from the current token representation (Kim et al., 14 Aug 2025). The motivation is that Mamba already offers linear-time recurrence, but practical edge deployment still requires hardware-friendly normalization, nonlinearity, and discretization.

The most visible modification is the replacement of LayerNorm with RangeNorm. The paper contrasts the original

$$\mathrm{LN}(x_i) = \gamma \frac{x_i - \mu}{\sqrt{\sigma^2 + \epsilon}} + \beta$$

with

$$\mathrm{RangeNorm}(x_i) = \gamma \frac{x_i - \mu}{\mathrm{range}(x_i - \mu)} + \beta,\quad \mathrm{range}(x) = \max(x) - \min(x),$$

thereby replacing variance and square-root operations with max–min statistics that are more amenable to comparator-based hardware (Kim et al., 14 Aug 2025). The design choice is not merely cosmetic; it is the first stage of the accelerator pipeline and a major latency bottleneck unless parallelized.

The framework also replaces softplus with ReLU for the step size $\Delta$. In the paper’s formulation, softplus is smooth but expensive, whereas ReLU is monotone increasing and preserved accuracy in experiments while removing expensive nonlinearity (Kim et al., 14 Aug 2025). For the gating path, eMamba approximates

$$\mathrm{SiLU}(x) = x\,\sigma(x) = \frac{x}{1 + e^{-x}}$$

with piecewise linear segments tuned to the empirical data range. Inputs smaller than $-7$ map to $0$, inputs larger than $\bar{A}_t$0 map to the identity, and the interval $\bar{A}_t$1 is approximated with $\bar{A}_t$2 linear segments, with at most $\bar{A}_t$3 max relative error versus true SiLU (Kim et al., 14 Aug 2025).

A similar approximation is applied to exponentiation in the discretization path. Inputs larger than $\bar{A}_t$4 are treated as a constant, inputs smaller than $\bar{A}_t$5 are approximated as $\bar{A}_t$6, and the interval $\bar{A}_t$7 is approximated with $\bar{A}_t$8 linear segments (Kim et al., 14 Aug 2025). Collectively, these substitutions define eMamba as a deployment-oriented variant of Mamba rather than a new task model: the recurrence is retained, but its numerics are re-engineered for edge hardware.

3. Approximation-aware NAS and quantization

eMamba couples its approximations with an approximation-aware neural architecture search. The search space comprises $\bar{A}_t$9 (token dim), $\bar{B}_t$0 (expansion), $\bar{B}_t$1 (patch size), $\bar{B}_t$2 (state dim), and $\bar{B}_t$3 (number of Mamba blocks), which jointly determine internal width, sequence length, state size, and depth (Kim et al., 14 Aug 2025). During training, the substitutions LayerNorm $\bar{B}_t$4 RangeNorm and softplus $\bar{B}_t$5 ReLU are already applied, whereas the piecewise approximations for exp and SiLU are used only for inference so as to avoid gradient mismatch (Kim et al., 14 Aug 2025).

The quantization pipeline is also central. The framework uses symmetric uniform quantization with zero-point $\bar{B}_t$6 and power-of-two scales. The paper gives a standard form

$\bar{B}_t$7

or equivalently

$\bar{B}_t$8

with $\bar{B}_t$9 and $C_t$0 (Kim et al., 14 Aug 2025). All weights, biases, and activations use INT8 by default.

The recurrent hidden state requires special handling. Direct INT8 quantization of $C_t$1 causes scale explosions because the recurrence repeatedly multiplies scales across time. eMamba therefore uses a scale-aware higher-precision path for SSM hidden states: $C_t$2 is computed at higher precision, re-scaled before storage, used immediately for output computation, and only then reduced to low precision where appropriate (Kim et al., 14 Aug 2025). The MARS example in the paper is explicit: $C_t$3 has scale $C_t$4, the INT24 hidden state is right-shifted by $C_t$5, stored as INT17 for recurrence, and $C_t$6 is INT8 (Kim et al., 14 Aug 2025).

This quantization scheme is one of the main distinctions between eMamba and a naive integer implementation of Mamba. The paper’s claim is not only that quantization is possible, but that recurrence-aware scaling is necessary for preserving accuracy under repeated state updates (Kim et al., 14 Aug 2025).

4. Hardware realization and empirical characterization

The framework is implemented on an AMD ZCU102 FPGA using Vivado 2024.1 and on an ASIC in GF 22FDX, 22 nm technology (Kim et al., 14 Aug 2025). On FPGA, the clock is $C_t$7 MHz and the design uses a layer-wise pipeline, token-by-token processing, and ready/valid synchronization. The RangeNorm stage is parallelized with configurable compute units, each containing a divider and one DSP48E2. For MARS, $C_t$8 units were selected, yielding a $C_t$9 reduction in RangeNorm latency and a $$\mathrm{LN}(x_i) = \gamma \frac{x_i - \mu}{\sqrt{\sigma^2 + \epsilon}} + \beta$$0 improvement in frame latency at negligible resource cost (Kim et al., 14 Aug 2025).

The full FPGA resource profile is reported. Reconfigurable eMamba uses LUT $$\mathrm{LN}(x_i) = \gamma \frac{x_i - \mu}{\sqrt{\sigma^2 + \epsilon}} + \beta$$1, FF $$\mathrm{LN}(x_i) = \gamma \frac{x_i - \mu}{\sqrt{\sigma^2 + \epsilon}} + \beta$$2, DSP $$\mathrm{LN}(x_i) = \gamma \frac{x_i - \mu}{\sqrt{\sigma^2 + \epsilon}} + \beta$$3, and BRAM $$\mathrm{LN}(x_i) = \gamma \frac{x_i - \mu}{\sqrt{\sigma^2 + \epsilon}} + \beta$$4; the embedded eMamba variant uses LUT $$\mathrm{LN}(x_i) = \gamma \frac{x_i - \mu}{\sqrt{\sigma^2 + \epsilon}} + \beta$$5, FF $$\mathrm{LN}(x_i) = \gamma \frac{x_i - \mu}{\sqrt{\sigma^2 + \epsilon}} + \beta$$6, DSP $$\mathrm{LN}(x_i) = \gamma \frac{x_i - \mu}{\sqrt{\sigma^2 + \epsilon}} + \beta$$7, and BRAM $$\mathrm{LN}(x_i) = \gamma \frac{x_i - \mu}{\sqrt{\sigma^2 + \epsilon}} + \beta$$8 (Kim et al., 14 Aug 2025). A quantized ViT accelerator under the same constraints does not fit, with LUT usage reported as $$\mathrm{LN}(x_i) = \gamma \frac{x_i - \mu}{\sqrt{\sigma^2 + \epsilon}} + \beta$$9 (Kim et al., 14 Aug 2025).

On ASIC, the operating frequency is $$\mathrm{RangeNorm}(x_i) = \gamma \frac{x_i - \mu}{\mathrm{range}(x_i - \mu)} + \beta,\quad \mathrm{range}(x) = \max(x) - \min(x),$$0 MHz. eMamba occupies $$\mathrm{RangeNorm}(x_i) = \gamma \frac{x_i - \mu}{\mathrm{range}(x_i - \mu)} + \beta,\quad \mathrm{range}(x) = \max(x) - \min(x),$$1 NAND2-equivalent gates and $$\mathrm{RangeNorm}(x_i) = \gamma \frac{x_i - \mu}{\mathrm{range}(x_i - \mu)} + \beta,\quad \mathrm{range}(x) = \max(x) - \min(x),$$2, with core dimensions approximately $$\mathrm{RangeNorm}(x_i) = \gamma \frac{x_i - \mu}{\mathrm{range}(x_i - \mu)} + \beta,\quad \mathrm{range}(x) = \max(x) - \min(x),$$3 post-layout. The corresponding ViT baseline uses $$\mathrm{RangeNorm}(x_i) = \gamma \frac{x_i - \mu}{\mathrm{range}(x_i - \mu)} + \beta,\quad \mathrm{range}(x) = \max(x) - \min(x),$$4 gates and $$\mathrm{RangeNorm}(x_i) = \gamma \frac{x_i - \mu}{\mathrm{range}(x_i - \mu)} + \beta,\quad \mathrm{range}(x) = \max(x) - \min(x),$$5, so eMamba is reported as $$\mathrm{RangeNorm}(x_i) = \gamma \frac{x_i - \mu}{\mathrm{range}(x_i - \mu)} + \beta,\quad \mathrm{range}(x) = \max(x) - \min(x),$$6 smaller in area (Kim et al., 14 Aug 2025). Total power is $$\mathrm{RangeNorm}(x_i) = \gamma \frac{x_i - \mu}{\mathrm{range}(x_i - \mu)} + \beta,\quad \mathrm{range}(x) = \max(x) - \min(x),$$7 mW for eMamba versus approximately $$\mathrm{RangeNorm}(x_i) = \gamma \frac{x_i - \mu}{\mathrm{range}(x_i - \mu)} + \beta,\quad \mathrm{range}(x) = \max(x) - \min(x),$$8 mW for ViT, and energy per inference is $$\mathrm{RangeNorm}(x_i) = \gamma \frac{x_i - \mu}{\mathrm{range}(x_i - \mu)} + \beta,\quad \mathrm{range}(x) = \max(x) - \min(x),$$9 versus $\Delta$0, corresponding to $\Delta$1 lower power and $\Delta$2 lower energy (Kim et al., 14 Aug 2025).

The paper evaluates accuracy and size on Fashion-MNIST, CIFAR-10, MARS, and WikiText2. On Fashion-MNIST, eMamba FP32 reaches $\Delta$3 with $\Delta$4K bytes versus ViT FP32 at $\Delta$5 with $\Delta$6M bytes; under INT8, eMamba reaches $\Delta$7 with $\Delta$8K bytes versus ViT INT8 at $\Delta$9 with $$\mathrm{SiLU}(x) = x\,\sigma(x) = \frac{x}{1 + e^{-x}}$$0K bytes (Kim et al., 14 Aug 2025). On CIFAR-10, eMamba FP32 reaches $$\mathrm{SiLU}(x) = x\,\sigma(x) = \frac{x}{1 + e^{-x}}$$1 with $$\mathrm{SiLU}(x) = x\,\sigma(x) = \frac{x}{1 + e^{-x}}$$2K bytes versus ViT FP32 at $$\mathrm{SiLU}(x) = x\,\sigma(x) = \frac{x}{1 + e^{-x}}$$3 with $$\mathrm{SiLU}(x) = x\,\sigma(x) = \frac{x}{1 + e^{-x}}$$4M bytes; eMamba INT8 reaches $$\mathrm{SiLU}(x) = x\,\sigma(x) = \frac{x}{1 + e^{-x}}$$5 with $$\mathrm{SiLU}(x) = x\,\sigma(x) = \frac{x}{1 + e^{-x}}$$6K bytes versus ViT INT8 at $$\mathrm{SiLU}(x) = x\,\sigma(x) = \frac{x}{1 + e^{-x}}$$7 with $$\mathrm{SiLU}(x) = x\,\sigma(x) = \frac{x}{1 + e^{-x}}$$8K bytes (Kim et al., 14 Aug 2025). On MARS, eMamba FP32 reports RMSE $$\mathrm{SiLU}(x) = x\,\sigma(x) = \frac{x}{1 + e^{-x}}$$9 cm with $-7$0K bytes, and eMamba INT8 reports RMSE $-7$1 cm with $-7$2K bytes (Kim et al., 14 Aug 2025).

The end-to-end hardware comparison is strongest on MARS. At $-7$3 MHz on FPGA, eMamba completes inference in $-7$4 cycles, compared with $-7$5 for a CNN accelerator, $-7$6 for ViT, and $-7$7 for naive Mamba. Throughput is $-7$8 Mb/s, compared with $-7$9 Mb/s for CNN, $0$0 Mb/s for ViT, and $0$1 Mb/s for naive Mamba (Kim et al., 14 Aug 2025). The paper summarizes these results as $0$2–$0$3 lower latency and $0$4–$0$5 higher throughput than baseline solutions while maintaining competitive accuracy (Kim et al., 14 Aug 2025).

The language-modeling result is used to argue that the framework is not confined to small vision tasks. On WikiText2, eMamba FP32 perplexity remains stable from sequence length $0$6 to $0$7 tokens, changing from $0$8 to $0$9, while eMamba INT8 changes from $\bar{A}_t$00 to $\bar{A}_t$01 (Kim et al., 14 Aug 2025). This is presented as evidence that the approximated, quantized implementation preserves Mamba’s long-sequence behavior.

5. Broader uses of “eMamba” in the Mamba literature

Outside the edge-acceleration paper, “eMamba” often functions as a loose descriptor rather than a formal model name. In event-based vision, "SpikMamba" is explicitly presented for readers searching for “eMamba (Event Mamba),” combining SNNs with Mamba for event-based human action recognition and reporting gains of $\bar{A}_t$02, $\bar{A}_t$03, $\bar{A}_t$04, and $\bar{A}_t$05 over previous state of the art on PAF, HARDVS, DVS128, and E-FAction, respectively (Chen et al., 2024). "SMamba" makes the same association even more directly, stating that it can be regarded as “eMamba” in the sense of an event-focused Mamba architecture for event-based object detection (Yang et al., 21 Jan 2025).

In electron microscopy, the term is similarly informal. "ShapeMamba-EM" is described as an EM-focused Mamba approach that can be read as an “eMamba,” even though the paper itself does not introduce the specific name eMamba (Shi et al., 2024). "VEMamba" includes a dedicated discussion of “Connecting to ‘eMamba’,” where the label denotes EM-focused goals rather than edge acceleration, and the model is instead centered on isotropic reconstruction through an Axial-Lateral Chunking Selective Scan Module and a Dynamic Weights Aggregation Module (Gao et al., 1 Mar 2026).

A different but orthographically similar term appears in "MambaFormer." There, EMamba is a customized, frozen Mamba-130M SSM expert specialized for efficient modeling of long clinical and biomedical sequences inside a token-level routed Mixture-of-Experts system (Khan et al., 3 Jan 2026). It is not the edge-deployment eMamba framework, even though both rely on the linear-time recurrence of Mamba.

The naming ambiguity is sharpened by negative evidence. "MoEMba," despite being Mamba-based and focused on high-density EMG hand gesture recognition, explicitly does not discuss or mention “eMamba” and provides no direct comparison to it (Shabanpour et al., 9 Feb 2025). This indicates that “eMamba” is not a blanket label for every Mamba model whose application domain begins with “EM.”

6. Limitations, distinctions, and research trajectory

The edge-deployment eMamba paper is explicit about its assumptions. The piecewise approximations are tuned to observed data ranges, so tasks whose activations fall outside the profiled regimes may require re-profiling and segment retuning. RangeNorm still requires division hardware, even if it avoids square roots and variance computation. Quantization of recurrent SSMs remains sensitive to outliers and therefore depends on careful calibration of scales and clipping. The reported models are also modest in size and are positioned for small-to-moderate edge workloads rather than very large LLM-scale deployments (Kim et al., 14 Aug 2025).

Across the broader literature, the term’s evolution suggests two distinct trajectories. One is hardware realization: eMamba in the strict sense is a deployment framework that reworks Mamba’s numerics, quantization, and pipeline structure for FPGA and ASIC execution (Kim et al., 14 Aug 2025). The other is domain specialization: event-based recognition and detection, electron microscopy segmentation and reconstruction, and clinical question answering all adapt Mamba’s selective state-space machinery to highly specific data regimes and constraints (Chen et al., 2024, Yang et al., 21 Jan 2025, Shi et al., 2024, Gao et al., 1 Mar 2026, Khan et al., 3 Jan 2026).

A plausible implication is that future references to eMamba will remain context-dependent unless the literature converges on stricter naming conventions. At present, the most exact usage is the edge-computing acceleration framework, whereas the broader uses denote application-specific Mamba variants that are “event-centric,” “EM-focused,” or “EMamba” only by analogy or local convention.