JointServe treats weight shards and KV cache blocks as competing residency objects under a unified memory budget, enabling multi-model LLM agent serving on memory-constrained GPUs without serial swap stalls. It targets private…

#semiconductors #AIinfra #GPUs #LLMs
https://arxiv.org/abs/2610.19822

#GPUS Hyperscale Data Announces Preliminary Third Quarter 2026 Revenue of Approximately $40 Million, Up 64% Compared to Third Q…

https://www.stocktitan.net/news/GPUS/hyperscale-data-announces-preliminary-third-quarter-2026-revenue-of-u6f5ypk13t9t.html?utm_source=stocktitan-bluesky&utm_medium=social

VibeServe explores using coding agents to automatically generate bespoke LLM serving stacks, matching vLLM and SGLang on Llama-3.1-8B and reaching 2.3x the goodput of tuned SGLang on a 397B MoE model across four…

#AIInfrastructure #LLMServing #GPUs #Semiconductors
https://arxiv.org/abs/2605.06068

New arXiv paper proposes Token Window Pacing (TWP), a closed-loop scheduler that uses runtime ITL, queue depth, and workload feedback to admit serverless LLM requests at the token level, balancing TTFT and inter-token…

#semiconductors #AIinfrastructure #GPUs #LLMs
https://arxiv.org/abs/2610.17689

Echo is a 4-bit all-reduce method for tensor parallelism in LLM training and inference that quantizes each value at most twice per call and feeds injected errors back into later calls to prevent accumulation.

#AI #LLMs #GPUs
https://arxiv.org/abs/2610.14277

#GPUS #GPUSPD Hyperscale Data Announces Plans for up to $33.5 Million in Tokenized Series D Preferred Stock Offering

https://www.stocktitan.net/news/GPUS/hyperscale-data-announces-plans-for-up-to-33-5-million-in-tokenized-lkwv1o02j4un.html?utm_source=stocktitan-bluesky&utm_medium=social

GraphTok moves tokenization from CPU to GPU, running pre-tokenization through BPE as a single replayable CUDA graph that matches reference token IDs across major open models and integrates with vLLM and NVIDIA Dynamo. The design…

#GPUs #AIInfrastructure #LLMs #vLLM
https://arxiv.org/abs/2610.13325

A new technique called Adaptive DHiSparse supports speculative decoding over a hierarchical sparse KV cache and adapts how much history stays in GPU memory based on load, helping long-context serving avoid running out of KV memory before hardware is…

#AI #LLM #GPUs
https://arxiv.org/abs/2610.13708

This arXiv paper proposes learning indexical policies with a choose rule to enable heuristic search that runs in polynomial space, trading space for time exponential only in choice depth. This could matter for AI infrastructure by…

#AI #Semiconductors #GPUs
https://arxiv.org/abs/2610.10954

Semiconductor and AI firms are preparing contingency plans for a potential "Day…

#Semiconductors #AIInfrastructure #GPUs #TechNews
https://news.google.com/rss/articles/CBMiW0FVX3lxTE5iOHh2Mmp2ejlOZDVoQzdVMU55dXVoSDBVbEFZbFRtTlpIYlBGTUExcmhMdXE5VjFhMlZzWnZxamhzdmtZZXk2VkNDSUVHMzlDc0JreDRTdDczWEk

A new project demonstrates building a fully open reranker for agentic search from scratch using only a handful of GPUs, avoiding third-party open-weight backbones and relying on common datasets with pretraining plus supervised…

#OpenSource #AI #SemanticSearch #GPUs
https://arxiv.org/abs/2610.11922

MiMo-V2.6 is an omni-modal model family scaling reinforcement learning compute across batch size, environment diversity, and grader compute to advance self-improvement in foundation models. The setup uses async training handling…

#semiconductors #AI #GPUs #MiMo
https://arxiv.org/abs/2610.11959

New arXiv paper introduces ResourceHijackBench, a benchmark with 300 scenarios and 900 prompts for studying LLM agent resource hijacking across GPUs, servers, and APIs.

#Semiconductors #GPUs #AIInfrastructure
https://arxiv.org/abs/2608.15108

New research introduces STATIC, a method that flattens prefix trees into a compressed sparse row matrix to enable efficient constrained decoding for LLM-based generative retrieval on TPUs and GPUs. The approach…

#AIInfrastructure #GPUs #LLMs #RecommenderSystems
https://arxiv.org/abs/2602.22647

New arXiv paper CoMoE tackles the cost barrier of MoE model deployment by enabling efficient inference on consumer GPUs over PCIe, using host-centric routing to bypass the need for expensive NVLink interconnects. If it holds up, this…

#AIInfrastructure #GPUs #MoE
https://arxiv.org/abs/2610.09424

South Korea's 2027 budget sets aside 4.7 trillion won for a single frontier AI model, including 10,000 Nvidia Vera Rubin GPUs, with a new contest opening after budget approval.

#AI #FrontierModel #Gpus #SouthKorea #SovereignAi

#GPUS Hyperscale Data Announces Live Access to Decentralized Exchange on the Ault Blockchain through Ault Markets

https://www.stocktitan.net/news/GPUS/hyperscale-data-announces-live-access-to-decentralized-exchange-on-c3c3k7k784kn.html?utm_source=stocktitan-bluesky&utm_medium=social