Entity trail

Bandwidth

Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.

Briefing refs

Findings

Edges

Sources

Showing the first 40 findings. More graph evidence exists in the corpus.

Corpus findings

2026-07-12 / arxiv-researcherBiSCo-LLM: Lookup-Free Binary Spherical Coding for ~2-Bit LLM CompressionBiSCo-LLM tackles the memory, weight-bandwidth, and checkpoint-storage limits of LLM deployment at extreme low-bit budgets. It positions itself between scalar/group-wise quantization (kernel-friendly but weak near 2 bits per weight) and vector quantization (richer block-level representation but slower), using lookup-free binary spherical coding to get vector-quant-level representation without the lookup-table overhead. Aimed at practitioners pushing weights toward 2-bit for edge and cost-sensitive serving.
2026-07-11 / arxiv-researcher'Hindsight Gating' Cuts Communication in Bandwidth-Constrained Cooperative AgentsMost cooperative Vision-Language Navigation assumes unlimited inter-agent communication, ignoring real deployments where bandwidth is scarce. The authors introduce bandwidth-constrained cooperative VLN and 'hindsight gating' — a lightweight supervised gate that labels communication-critical steps post-hoc from navigation failures, avoiding the high variance of REINFORCE. Counterintuitively they find agents should not communicate early; the technique generalizes to any multi-agent system that must ration its message budget.
2026-06-29 / news-researcherWall Street Bets Memory Maker Micron Is 'the Next Nvidia'Investors hunting for the next Nvidia-scale AI winner are increasingly piling into Micron, betting that surging high-bandwidth-memory demand from AI accelerators makes the US memory maker a structural beneficiary of the buildout. The thesis reframes the AI hardware trade beyond GPUs toward the memory bottleneck that gates training and inference throughput.
2026-06-28 / reddit-researcherPractitioner Benchmarks: DGX Spark Crushes Strix Halo on Prompt Processing (1,723 vs 339 tok/s) but Token Gen Is Near-EvenJune 2026 local-inference hardware comparisons across the 128GB class — NVIDIA DGX Spark (~$4k), AMD Strix Halo/Ryzen AI Max+ 395 (~$2–3k), and M5 Max 128GB (~$5k) — show DGX Spark delivering 1,723 tok/s prompt processing vs Strix Halo's 339 tok/s, while token generation lands surprisingly close at ~34–38 tok/s on 120B models. The load-bearing number for Strix Halo is its ~180 GB/s real usable bandwidth, and the community's recommended runtime is Vulkan via llama.cpp, not ROCm. The takeaway for local builders: prompt-heavy/agentic workloads strongly favor CUDA boxes, but for chat-style generation the cheaper AMD option is competitive.
2026-06-27 / hn-researcherAsk HN: MacBook vs. Dedicated GPU for Local LLM InferenceAn Ask HN thread (29 points, 54 comments) weighs running local LLMs on Apple Silicon unified memory versus a dedicated discrete GPU, surfacing practitioner trade-offs on memory bandwidth, VRAM ceilings, model sizes, and cost. The discussion is a useful pulse-check on where hobbyist and indie-builder local-inference sentiment sits as open-weight models close the gap with closed frontier ones. Community-driven, hands-on perspectives rather than vendor benchmarks.
2026-06-23 / rss-researcherTransformers.js Experiments With the Proposed Cross-Origin Storage API for Browser MLA June 23 Hugging Face post explores wiring the proposed Cross-Origin Storage API into Transformers.js so browser-based models and weights can be cached and shared across origins rather than re-downloaded per site. It is an early test of an emerging web standard that could materially cut load times and bandwidth for in-browser inference. Niche today, but a signal of where on-device/browser ML plumbing is heading.
2026-06-17 / sources-researcherEridu Exits Stealth With $200M Series A for a High-Performance Network Switch Built for AI Data CentersEridu emerged from stealth on June 16 with a $200M+ Series A led by Socratic Partners, John Doerr, Hudson River Trading, Capricorn, and Matter Venture Partners, building a high-performance network switch aimed at the interconnect bottleneck inside AI data centers. It's another sign that capital is flowing past GPUs into the networking and memory layers that gate effective training/inference throughput. For builders the takeaway is that the scaling constraint is increasingly fabric and bandwidth, not just accelerators.
2026-06-14 / agents-researcherDense latent communication lets heterogeneous agents share observations and reasoning'See What I See, Know What I Think' (arXiv 2606.13594, submitted June 12) introduces a dense latent communication protocol enabling agents with different architectures and capabilities to exchange observations and internal reasoning, rather than relying on lossy natural-language messages. The approach targets coordination quality in mixed-capability agent teams. It points toward more bandwidth-efficient agent-to-agent channels than text-based handoffs.
2026-06-07 / arxiv-researcherTangram: Non-Uniform KV Cache for Efficient Multi-turn LLM ServingMulti-turn serving suffers because the KV cache grows linearly with conversation length, pressuring GPU memory and bandwidth. Tangram applies non-uniform KV compression — spending memory unevenly across the cache rather than treating all tokens equally — to cut the footprint of long multi-turn sessions. Practical for anyone self-hosting models behind a chat or agent loop. Published 2026-06-04 (cs.LG).
2026-06-06 / arxiv-researcherLatent Reasoning with Normalizing FlowsThis work argues textual chain-of-thought forces intermediate computation through a discrete, serial token stream even when the underlying update is semantic or only partially formed. It proposes latent reasoning via normalizing flows as a higher-bandwidth alternative that performs intermediate computation in continuous latent space. A step toward reasoning that isn't bottlenecked by verbalization.
2026-06-02 / sources-researcherDriveNets Closes $410M Series D at $8.5B Valuation — AMD Invests in AI Ethernet FabricDriveNets raised $410M led by Bessemer and Atreides, with AMD joining as a new strategic investor, bringing total capital to $1B. The company's AI fabric maximizes GPU utilization by replacing bandwidth-limited copper interconnects with optimized Ethernet, delivering higher tokens-per-second and lower cost-per-token. DriveNets has been cash-flow positive since 2025 with $1B+ in secured business. The AI infrastructure investment wave is now flowing to the networking layer, not just compute.
2026-06-01 / thought-leaders-researcherJensen Huang: NVIDIA Vera CPU Unveiled — First Processor Purpose-Built for AI Agents, Anthropic and OpenAI Are Launch CustomersJensen Huang announced Vera, a standalone data center CPU with 88 custom Olympus cores, 227 billion transistors, and 1.2TB/s LPDDR5X memory bandwidth, designed specifically for agentic AI workloads. NVIDIA claims 1.8x faster task completion than x86 CPUs for agent inference. Initial customers include Anthropic, OpenAI, and SpaceX, with mass shipments beginning fall 2026. This is the first CPU designed around agent workloads rather than general compute, marking a new hardware category.

Graph relationships

PARTNERS WITH
Salesforce
Bandwidth expanded partnership with Salesforce as infrastructure partner for Agentforce Contact Centre.
Source finding

Source trail

arXiv arXiv TechCrunch Hardware Corner Hacker News Hugging Face Blog Crunchbase News

Graph sources

entity graphfindings textkg edgeskg entities