AI Today: NVIDIA 6G and AMD InferaSim
Authorship and approval
Published and updated
Knowledge state as of 2026-11-08. Source material gathered 2026-11-07. The set covers the expansion of NVIDIA's open tools for 5G and 6G networks, AMD's simulator for planning LLM serving, and the European Commission's preliminary findings regarding TikTok. It also includes new English varieties in ASR evaluation and the BlockLock method, studied for inference result stability under varying batch composition.
Highlights Today (TL;DR)
- NVIDIA: Open tools for 5G and 6G surpassed one million downloads.
- AMD: InferaSim used 9.1 GPU hours for 11 anchor measurements and covered projections for 226 configurations.
- TikTok: The European Commission preliminarily concluded that the "For You" feed breaches the DSA.
- Open ASR Leaderboard: Private data for Egyptian and Nigerian English varieties were added to the evaluation.
- BlockLock: In tests, BlockLock-Rows increased throughput by 115.2–122.2% compared to vLLM Batch-Invariant.
NVIDIA Releases Open Source Tools for 6G Network Research
NVIDIA reported on Friday the growing adoption of its open tools for 5G and 6G network development, including Aerial and Sionna. The company open-sourced Aerial a year ago; its toolset combines AI simulations and training with experiments on real wireless systems. According to NVIDIA, over the past year global usage of Aerial and Sionna grew 10-fold, open source tools for 5G/6G surpassed one million downloads, and more than 8,000 research and academic institutions use the platforms.
Use Case Examples
Researchers from King’s College London combined a robotic arm equipped with a camera, a Jetson AGX Orin computer, and Sionna. In fruit-packing and block-stacking experiments, the robot transmitted concise object descriptions instead of regular images. According to NVIDIA, this reduced task completion time by up to 52.6% and improved success rates compared to sending images. Testing also included intentionally introduced faults. SoftBank integrated a neural receiver with NVIDIA Aerial CUDA-Accelerated RAN (ACAR) and tested it with a commercial 5G smartphone. In controlled conditions with interference, it achieved higher uplink throughput than a conventional receiver across all 15 tested settings; the maximum increase reached approximately 60%. Meanwhile, ETH Zurich researchers found that fine-tuning receivers on local measurements reduced the block error rate by more than a third compared to selected pre-trained versions. The University of Leeds team developed a system predicting network demand 10 minutes ahead to steer base stations into sleep mode. In tests with a portable 5G network and a digital twin, they recorded a 45.6% drop in RAN energy consumption compared to full-power operation while maintaining device connectivity. Meanwhile, CTTC combined GPU-accelerated radio channel models with the 5G-LENA simulator. In preliminary evaluation, simulation time dropped from about three hours to 10–20 minutes, depending on integration. Aerial and Sionna tools are available on GitHub. For AI builders: A research team developing RAN networks can check the integration of GPU channel models with 5G-LENA and compare simulation times on their own scenarios. The material does not provide hardware requirements for the complete toolset or license details, which must be verified before use.
AMD Details ROCm InferaSim for Planning LLM Serving Configurations
AMD described the ROCm InferaSim tool on November 6, aimed at reducing costly configuration tests for serving LLM models on GPU nodes. Instead of measuring every variant on hardware, the tool performs measurements for sets of used kernels and then computes projections for configurations that utilize them. A single measurement, called an anchor, records prefill and decode operation times along with information about the executed configuration. Changes such as batch size, context length, tensor parallelism width, or replica count can be analytically ported from an existing measurement according to the description. Changing the data type, attention backend, collectives, graph mode, or speculative decoding requires an anchor for the new kernel set. Projection takes less than half a second on a single CPU core. In AMD tests on MI355X, performing 11 anchor measurements took 9.1 GPU hours and covered 226 configurations. Measuring these configurations separately required 204 GPU hours. The authors estimate that measuring all candidates in their search would consume around 25,000 GPU hours; hardware is only needed for anchors and final short-list confirmation, while intermediate calculations run on CPU. InferaSim can estimate HBM memory fit, latency for a given arrival request model, and the number of replicas needed to meet SLOs, among other metrics. The tool also creates a Pareto frontier showing the trade-off between GPU throughput and interactivity. In simulation mode, it does not require an anchor, but prices only kernels and serves to compare candidates within a single engine. The code is available on GitHub, and instructions describe installation on a ROCm host. For AI builders: An infrastructure team planning LLM serving on AMD GPUs can prepare anchors for used kernel sets, run configuration projections, and confirm selected variants on GPUs. The article does not state the tool's price or license terms—these must be checked before deployment.
European Commission Preliminarily Finds TikTok Recommendations Breach DSA
On November 6, the European Commission preliminarily found that TikTok’s personalized "For You" feed breaches the EU Digital Services Act (DSA). The Commission points to the "rabbit hole" effect: the feed displays a continuous stream of videos tailored to user interests, and interactions with a specific type of content lead to the recommendation of similar materials. Repeated exposure to content concerning sadness, isolation, or body image can harm mental health, particularly for minors and adults susceptible to such influences. The preliminary findings concern both risk assessment and risk mitigation measures. The Commission's proceedings are ongoing. For AI builders: A team designing a personalized content feed can check whether they assess the risk of repeated recommendations and their impact on mental health. The Commission's statement does not outline specific required remedial actions or a proceeding schedule.
Open ASR Leaderboard Adds Egyptian and Nigerian English
The Open ASR Leaderboard announced the addition of private evaluation data for speech recognition in Egyptian and Nigerian English, in collaboration with Liva AI. The update aims to increase the representation of diverse accents in evaluations and limit the ability to fine-tune models specifically for public benchmarks. The leaderboard will also feature a richer accent breakdown based on Kachru's Three Circles concept. The authors announce that as more accents are added, they will separate the Outer and Expanding Circles. This is another step in expanding test sets. Previously, 11 subsets covering five accents were added, prepared with Appen and DataOceanAI, followed by a private subset for Indian English in cooperation with Voice Arena. Liva AI gathers multimodal data through consumer products used by real people; the company describes them as natural interactions with rich context rather than utterances elicited in test tasks. For AI builders: Teams evaluating ASR systems can check leaderboard results separately for Egyptian and Nigerian English. The announcement does not disclose the size of the new datasets or whether data is available outside leaderboard evaluation—these questions must be resolved before planning custom tests.
BlockLock Stabilizes LLM Inference Results with Variable Batch Composition
In a paper published on November 6, the authors introduced BlockLock, a method designed to preserve identical outputs for a protected request despite changes in other requests within the same batch, under a specified deployment contract. BlockLock maintains a shared, batched prefill phase and fixes the computation geometry seen by the protected request. The BlockLock-Rows variant uses fixed-width rows and preserves batched decoding for all requests. BlockLock-Pack packs data with end padding, adds an isolated placeholder segment, and then extracts the KV state of the protected request for decoding exclusively for it. The authors tested both variants on five dense models: Qwen3-8B, Qwen3-14B, Llama-3.1-8B-Instruct, Gemma-3-4B-IT, and Gemma-3-12B-IT. Experiments covered single NVIDIA A6000, A100, and H100 cards, tinyBenchmarks and BFCL suites, and preliminary tensor parallelism tests on two A6000 cards. In the described tests, both variants preserved output sequences and Top20 logit bits. At the same workload size, BlockLock-Rows reduced latency by 53.5–55.0% and increased throughput by 115.2–122.2% compared to vLLM Batch-Invariant. For AI builders: An inference server maintenance team can test both variants on their own batches and establish the required deployment contract. The source does not provide code availability or licenses, nor does it describe tests for Polish language tasks.
Frequently Asked Questions (FAQ)
How does AMD InferaSim limit the number of LLM serving configuration tests?
AMD describes InferaSim as a tool that measures sets of used kernels and then projects configurations utilizing those sets. In a test, 11 anchors took 9.1 GPU hours and covered 226 configurations, instead of 204 GPU hours to measure them individually.
Why did the European Commission challenge TikTok's "For You" feed?
The Commission points out that interactions with a specific type of content can lead to further similar recommendations. Repeated exposure to material about sadness, isolation, or body image can harm mental health, particularly for minors and vulnerable individuals.
What results did BlockLock-Rows achieve compared to vLLM Batch-Invariant?
In the described tests, both variants preserved output sequences and Top20 logit bits. BlockLock-Rows, at the same workload size as vLLM Batch-Invariant, reduced latency by 53.5–55.0% and increased throughput by 115.2–122.2%.
Sources
- How NVIDIA Open Source Tools Are Helping Researchers Build Next-Generation 6G Networks — 2026-11-06
- ROCm InferaSim: Cut LLM Serving Trial Costs and Deploy the Best Configuration — 2026-11-06
- Commission preliminarily finds TikTok’s ‘For You’ feed in breach of Digital Services Act for harmful ‘rabbit holes’ effect — 2026-11-06
- Egyptian and Nigerian English join the Open ASR Leaderboard — 2026-11-06
- BlockLock: Reproducible LLM Inference with Fixed Batch Geometry — 2026-11-06 --- Text prepared automatically by the aipoc-lab (AI) laboratory based on publicly available sources listed above. Own summaries, not copies. Pre-human-editing version.
Provenance manifest of this post (JSON) — which signals it was built from, what decisions the loops made, and when a human approved it.
Related posts
AI Today: Nemotron and PDD Kit
Discover how NVIDIA promotes open Nemotron models and Microsoft Research uses PDD Kit for verifying agent-written code.
AI Today: OpenAI, Microsoft, and AMD Instinct
Discover key updates on OpenAI's robots.txt standard, Microsoft's synthetic data testing for Copilot, and MONAI optimization for AMD Instinct.
AI Today: Google Genie 4 and Gemini 4 Argon
Discover the latest AI releases including Google Genie 4, Gemini 4 Argon, AMD ROCm 10.0.0, Liquid AI LFM3, and Anima's healthcare automation.
AI Today: Kolibri and NVIDIA Kumo Relational
Discover Kolibri, a sovereign German mixture-of-experts model, and NVIDIA Kumo Relational for direct database prediction.