cs.ARNov 16, 2026

Integrated GPU Core for Mamba-3

Authors: Shiva Raja, William Desueza, Ajay Joshi

Organizations: ECE Dept, Boston University

Abstract

Mamba-3, the latest Mamba state-space model (SSM), enables efficient long-context modeling, matching Mamba-2 perplexity with half the state size and increasing decoding arithmetic intensity. However, its multi-input multi-output (MIMO) computation, exponential-trapezoidal discretization, and rotary positional embeddings (RoPE) introduce dense, element-wise, and sequential operations. Despite kernel fusion, conventional GPUs execute these operations inefficiently. Our characterization of Mamba-3 hybrid models reveals numerous kernel launches, substantial data movement, and increased register pressure that reduces warp occupancy and memory-latency hiding. We propose RescuCore, a specialized GPU execution unit for Mamba-3 inference that retains other hybrid model layers on existing GPU resources. Integrated alongside CUDA Cores and Tensor Cores within each Streaming Multiprocessor (SM) sub-partition, RescuCore shares the register file, shared memory, and warp schedulers. Dense projections remain on conventional GPU resources, while RescuCore pipelines discretization, normalization, RoPE, MIMO computation, recurrent state updates, and output gating to reduce kernel fragmentation and intermediate data movement. We select BF16 after evaluating multiple numerical formats, preserving model accuracy with an estimated area overhead of 5.9% of an A100 die for state size 32 and MIMO rank 4. We evaluate RescuCore using cycle-accurate Accel-Sim simulation, RTL synthesis, and synthesis-informed AccelWattch power modeling. Compared with conventional GPU execution, RescuCore achieves 1.4-9.4x speedup and 1.4-5.0x energy reduction during prefill, and 1.3-2.8x speedup and 1.2-2.6x energy reduction during autoregressive decoding. These results demonstrate the potential of GPU-integrated specialized execution for SSM-based hybrid models.

Figures & tables

Explore similar work

Nov 2, 2026

Diagonal Equivalence in State Space Models: Characterization, Retrieval Rank, and Selective Extensions

We classify real and complex state-space models (SSMs) with linear transitions, distinguishing fixed-basis diagonal equivalence from input-output equivalence and tracking coefficient-conditioning windows under coordinate changes. We characterize diagonal equivalence through commuting transition families whose members are diagonalizable over the chosen field. With time-shared token-local matrix maps, we identify this class's state-mediated kernels with unnormalized scalar-mode linear attention with multiplicative masks. Mamba 3's multiple-input, multiple-output (MIMO) construction lifts one value per local interface; SSD-MIMO routes independent values for local retrieval rank above one. HybridMamba combines Mamba 1's per-coordinate decay, Mamba 2-style heads and Mamba 3's complex rotations and two-endpoint writes, mixing real and rotating state blocks with different decay laws. They recover Mamba 1-3 single-input, single-output (SISO) configurations as exact special cases. On five large H100 bfloat16 workloads, our fused GPU kernels improve geometric-mean training time over official Mamba 1/3 in matched SISO recoveries, with Mamba 2 near parity at SSM-core scope. Memory tradeoffs depend on the baseline and metric: saved-for-backward versus peak allocation. In small-model multi-token recall experiments, both designs outperform Mamba 1-3 on the hardest tier; higher routing ranks perform best in gated or phase-diverse configurations. In single-seed language models with about 500M parameters, both improve on Mamba 1/2 in test loss at 2B tokens, but the routing gains do not translate into lower language-model loss. Our best configurations approach Mamba 3 SISO through 10B tokens while remaining behind it and Mamba 3 MIMO in language-model loss.
Mar 16, 2026

Mamba-3: Improved Sequence Modeling using State Space Principles

Scaling inference-time compute has emerged as an important driver of LLM performance, making inference efficiency a central focus of model design alongside model quality. While the current Transformer-based models deliver strong model quality, their quadratic compute and linear memory make inference expensive. This has spurred the development of sub-quadratic models with reduced linear compute and constant memory requirements. However, many recent linear models trade off model quality and capability for algorithmic efficiency, failing on tasks such as state tracking. Moreover, their theoretically linear inference remains hardware-inefficient in practice. Guided by an inference-first perspective, we introduce three core methodological improvements inspired by the state space model (SSM) viewpoint of linear models. We combine: (1) a more expressive recurrence derived from SSM discretization, (2) a complex-valued state update rule that enables richer state tracking, and (3) a multi-input, multi-output (MIMO) formulation for better model performance without increasing decode latency. Together with architectural refinements, our Mamba-3 model achieves significant gains across retrieval, state-tracking, and downstream language modeling tasks. At the 1.5B scale, Mamba-3 improves average downstream accuracy by 0.6 percentage points compared to the next best model (Gated DeltaNet), with Mamba-3's MIMO variant further improving accuracy by another 1.2 points for a total 1.8 point gain. Across state-size experiments, Mamba-3 achieves comparable perplexity to Mamba-2 despite using half of its predecessor's state size. Our evaluations demonstrate Mamba-3's ability to advance the performance-efficiency Pareto frontier.
Feb 24, 2026

Scaling State-Space Models on Multiple GPUs with Tensor Parallelism

Selective state space models (SSMs) have rapidly become a compelling backbone for large language models, especially for long-context workloads. Yet in deployment, their inference performance is often bounded by the memory capacity, bandwidth, and latency limits of a single GPU, making multi-GPU execution increasingly necessary. Although tensor parallelism (TP) is widely used to scale Transformer inference, applying it to selective SSM blocks is non-trivial because the SSM mixer couples large projections with a sequence-wise recurrent state update and local mixing whose efficiency depends on preserving locality and avoiding synchronization in the critical path. This paper presents a communication-efficient TP design for selective SSM inference that addresses three practical engineering challenges: enabling TTFT improvements via an SSM state cache across prefill and decode, partitioning the mixer's packed parameter tensor so that recurrent updates remain local while minimizing communication, and reducing TP aggregation overhead with quantized AllReduce. We evaluate on three representative SSM-based LLMs spanning pure-SSM and hybrid architectures - Mamba, Falcon-Mamba, and Zamba - on NVIDIA A6000 and A100 clusters. Our experiments show substantial throughput gains from tensor-parallel SSM inference, improving batch-request throughput by ~1.6-2.1x on 2 GPUs and ~2.6-4.0x on 4 GPUs for Mamba, with the largest benefits at long context lengths, and achieving a further ~10-18% throughput improvement from quantized all-reduce by lowering synchronization bandwidth overhead.