Chunmeng Dou | Searcharxiv

arXiv2026

FlashAccel: Leveraging High-Bandwidth Flash (HBF) for High-Throughput LLM Inference

Large language model (LLM) inference is increasingly limited by the capacity of High-Bandwidth Memory (HBM) in GPUs, as model weights and KV cache grow rapidly. High-Bandwidth Flash (HBF) provides higher capacity than HBM while offering comparable bandwidth, making it a promising substrate for capacity-constrained LLM inference. However, its inherently high access latency, low bandwidth utilization, and lack of support for heterogeneous resource management make it difficult to integrate HBF into GPUs for LLM inference. We present FlashAccel, a co-designed system that enables efficient LLM inference using HBF. FlashAccel integrates HBF into HBM-based GPUs, providing architectural support to mitigate access latency. It improves bandwidth utilization through specialized data layouts for both model weights and KV cache, and introduces an HBF-aware storage management layer together with a programming model to organize persistent data in HBF and coordinate heterogeneous memory resources at the system level. Experimental results demonstrate that integrating six HBF stacks into the GPU enables FlashAccel to deliver an average improvement of 2.49$\times$ and 1.93$\times$ in throughput per GPU and energy efficiency over the HBM-only GPU under a 100ms latency constraint, respectively.

cs.AR↗

arXiv2023

Binary stochasticity enabled highly efficient neuromorphic deep learning achieves better-than-software accuracy

Deep learning needs high-precision handling of forwarding signals, backpropagating errors, and updating weights. This is inherently required by the learning algorithm since the gradient descent learning rule relies on the chain product of partial derivatives. However, it is challenging to implement deep learning in hardware systems that use noisy analog memristors as artificial synapses, as well as not being biologically plausible. Memristor-based implementations generally result in an excessive cost of neuronal circuits and stringent demands for idealized synaptic devices. Here, we demonstrate that the requirement for high precision is not necessary and that more efficient deep learning can be achieved when this requirement is lifted. We propose a binary stochastic learning algorithm that modifies all elementary neural network operations, by introducing (i) stochastic binarization of both the forwarding signals and the activation function derivatives, (ii) signed binarization of the backpropagating errors, and (iii) step-wised weight updates. Through an extensive hybrid approach of software simulation and hardware experiments, we find that binary stochastic deep learning systems can provide better performance than the software-based benchmarks using the high-precision learning algorithm. Also, the binary stochastic algorithm strongly simplifies the neural network operations in hardware, resulting in an improvement of the energy efficiency for the multiply-and-accumulate operations by more than three orders of magnitudes.

cs.NE↗