Searcharxiv⌕ Search

arXiv · 2611.11036

When Is High-Bandwidth Flash Cost-Effective for LLM Inference?

Abstract

Can high-bandwidth flash (HBF) make LLM inference cheaper than adding GPUs? By storing weights closer to the GPU than SSD, HBF may reduce the number of GPUs needed for model capacity. However, this integration is only justified if it meets service targets and its cost is recovered. We simulate this integration using public component data across nine dense and mixture-of-experts (MoE) models, five workloads, two GPU platforms, and 1--100 GPUs. At fixed demand, HBF reduces the minimum GPU count to meet a service target in 98 of 468 comparisons; majority of them are paused and short sessions, 63 remain cost-effective at half a GPU price per HBF stack. For single short-chat users, HBF enables large MoEs, e.g. DeepSeek V3 and Kimi K2, to meet a relaxed service target on one RTX~4090 GPU, where conventional serving fails. At fixed GPU count, HBF improves throughput in 48\% of matched comparisons below 16 GPUs and 13\% at larger counts. Even at 100 RTX~4090-scale GPUs, cascaded HBF raises offline throughput by 16.4--26.5$\times$ for three large MoEs. We found savings and throughput gains arise in different scenarios.

Explore related subjects

Keep this discovery

Explore connections, maps & timelines

BibTeXRIS

Zhiyu An, Yiping Li, Wan Du. 2026-11-08. When Is High-Bandwidth Flash Cost-Effective for LLM Inference?. https://searcharxiv.com/abs/2611.11036

Citations retain the original authors and DOI and include this site’s record URL. If this discovery service contributed to your research, you can also cite Searcharxiv by Indiciae: BibTeX · RIS.

KEEP EXPLORING

Related papers

SPLASH: A Virtualized HBM-HBF Hierarchy with Flash-Aware Sparse Attention for Long-Context Inference

As context lengths, request lifetimes and concurrency grow, the key-value (KV) cache dominates memory usage in large language model (LLM) inference. High-bandwidth memory (HBM), the memory on contemporary GPUs, achieves far higher bandwidth than off-package memory, yet its capacity, bounded by DRAM density, has not kept pace with this demand. High-Bandwidth Flash (HBF), an emerging memory technology, bridges this gap, combining the advanced packaging of HBM with the density of flash to deliver terabyte-scale capacity at HBM-class read bandwidth. Exploiting this capacity for long-context KV caches, however, requires hiding the vastly different physical characteristics of flash from the GPU while keeping decode interactive. We therefore introduce Splash, an architecture that virtualizes HBM and HBF for the KV cache. We architect custom HBM and HBF base dies with novel memory and flash controllers that handle the access semantics and device management of flash. These controllers keep the GPU agnostic of the underlying memory, allowing the HBM-HBF hierarchy to behave as one logical memory for the KV cache. The added capacity of HBF enables long contexts, but dense attention, which reads the entire KV cache for every generated token, cannot keep decode interactive at these lengths. Splash therefore makes sparse attention flash-aware through query-driven address translation, in which the HBF base die resolves each query into the flash addresses of the relevant KV pages and schedules their reads across flash planes. Since the HBF base die holds the location of every page, this translation hides flash from the GPU and balances the selected reads across flash planes, coupling abstraction with performance. Across five models and up to one million tokens, Splash improves per-GPU decode throughput by up to 4.5x over the strongest baseline under a 100 ms latency budget.

cs.AR↗

Accelerating 3D Gaussian Splatting Training through Selective Freezing and Computation Caching

3D Gaussian Splatting (3DGS) becomes a key primitive for immersive 3D experiences. However, its applicability is limited by its long training overhead. Although prior work has shown that gradient accumulation is the dominant bottleneck in training, existing works either have limited speedup or degrade reconstruction quality. To address this issue, we present ElsaGS, an algorithm-hardware co-designed framework that accelerates 3DGS training with negligible accuracy degradation. Our key observation is that most Gaussians converge early, while later iterations only refine a small subset of Gaussians. Thus, we propose a technique, progressive Gaussian freezing, which dynamically freezes well-converged Gaussians and caches their cumulative contributions, thereby reducing computations in both forward rasterization and backward gradient accumulation. To make caching efficient, we further propose an architecture, ElsaGS core, which integrates a pixel-parallel rasterization array with a set-associative frozen Gaussian cache, FGcache. Across three 3DGS datasets, ElsaGS freezes up to 90% of Gaussians with negligible quality loss and achieves up to 20.6x end-to-end speedup and 99.4% energy reduction over GPUs.

cs.AR↗

PVTSizing 2.0: A Fast and Reliable Analog Circuit Sizing Operator for Agentic EDA Workflows

Analog circuit design automation requires both coordinated tool execution and efficient numerical search. Large language models (LLMs) offer new ways to organize tool execution, but finding one sizing that satisfies specifications across process, voltage, and temperature (PVT) corners remains costly. PVTSizing 2.0 addresses these needs through a human-agent workflow built around an independently usable numerical optimizer. An agent carries the designer's approved task through execution and result review. The optimizer reduces unnecessary training through loss-guided updates and uses critic-guided local refinement to explore additional candidates while accounting for cross-corner conflicts. An optional LLM adviser supplies parameter relationships to focus numerical search on circuit-informed directions, while the optimizer selects adjustment amplitudes and verifies their effects through simulation. Across four circuit benchmarks, the standalone optimizer completes all 40 runs and achieves a 7.3x speedup in overall mean runtime over the original PVTSizing, with a comparable mean sample count. In a separate study, the best tested knowledge-injection configuration further reduces mean runtime by 39.9% and simulation samples by 35.4% relative to the standalone optimizer.

cs.AR↗