LesenForschungRadarAnlageframework
Anmelden / Registrieren
AlleTechnikProdukteInvestierenBörseSonstiges

Lesen

2026-11-1419 Beiträge

Aktualisieren

Onsemi Showcases 20kW AI Data Center Power Prototype Using Vertical GaN

Thema · 安森美垂直GaN电源技术Kurzfassung
2026-11-13 18:30 GMT+8

At the electronica 2026 trade show in Munich, Germany, semiconductor company Onsemi showcased a prototype 20kW-class intermediate bus converter (IBC). The device uses Onsemi's proprietary 1200V-rated vertical gallium nitride (GaN) power devices to directly convert an 800V DC input to a 6V output, targeting the shift toward high-voltage DC power distribution in AI data centers.

As computing power increases in AI servers, traditional 48V or 52V distribution schemes require higher currents that increase transmission losses and cable size. Moving to 800V DC distribution reduces these losses but requires stepping down the voltage near the processors. By converting directly from 800V to 6V, the system eliminates intermediate conversion stages like 48V or 12V buses.

Onsemi's approach relies on its vertical GaN architecture, where current flows through the thickness of the chip rather than across its surface. This structure allows for higher voltage ratings and current density compared to conventional lateral GaN devices. The company claims this design can reduce chip size by approximately one-third compared to commercial lateral GaN products. The prototype utilizes an LLC resonant conversion topology with MHz-band switching frequencies to shrink magnetic components like transformers. According to the display panel at the event, the prototype achieved a conversion efficiency of 96.5%.

Onsemi representatives noted that the 1200V rating avoids the need to series-connect two 650V lateral GaN devices, simplifying the circuit. They also highlighted that GaN's lack of reverse recovery charge benefits electromagnetic interference (EMI) performance. However, they cautioned that GaN devices are sensitive to gate-source voltage control and require careful layout and thermal management due to their small chip size. This demonstration positions Onsemi against competitors such as Renesas Electronics, Texas Instruments, Navitas Semiconductor, STMicroelectronics, and Infineon Technologies, who have also announced developments in 800V-to-6V power systems for AI infrastructure.

Quellen:EE Times Japan

Forschung

SemiAnalysis's Dylan Patel: Blocking Lithography Equipment Slows China's AI More Than Banning Chip Exports

Thema · 中美AI出口管制Wesentlich
2026-11-14 07:13 GMT+8

On November 13, 2026, three members of the semiconductor research firm SemiAnalysis—Jordan Schneider, Dylan Patel, and Jordan Nanos—recorded an emergency podcast regarding Jensen Huang's resignation as CEO of Nvidia. During their discussion on policy directions, Dylan Patel proposed a strategic framework that contradicted the prevailing mainstream thinking in Washington at the time: rather than banning Chinese access to remote US computing power, the focus should be on blocking the equipment required to manufacture chips.

Patel's core logic is to 'flip the table.' He uses a jigsaw puzzle analogy to explain: if you remove one piece from the top of the supply chain (such as banning a specific chip export), China will specifically target and overcome that single piece; but if you flip over the entire puzzle board, the difficulty level changes completely. Specifically, instead of banning the chips themselves or restricting model access, it is more effective to prevent companies like ASML from selling DUV lithography equipment to China. Without this equipment, China cannot mass-produce advanced chips, regardless of how it designs architectures or optimizes algorithms.

The second pillar of this strategy is to permit, rather than prohibit, remote computing access. Patel's game-theory reasoning suggests that if companies such as Alibaba, ByteDance, and Moonshot must rent computing power from US cloud service providers, the United States gains three distinct advantages. First, Chinese data may be processed within the US, allowing intelligence agencies access. Second, if tensions escalate, the US can cut off access at any moment—a 'kill switch.' Third, every megawatt of computing deployed outside China generates massive profits for the US ecosystem: Nvidia earns approximately $7 million per deployment, with cloud service providers adding another $7 million to $10 million.

Patel argues that cutting off remote access would instead incentivize smuggling, accelerate the development of China's domestic chip industry (which was going to happen anyway), and eliminate China's economic incentive to maintain the status quo of manufacturing electronics for American companies. He cites Nvidia's largest PCB supplier as a Chinese company as an example. Maintaining the current situation gives China a financial reason to continue cooperation.

However, this argument faces pushback from within the same podcast. Jordan Schneider points out that parts of the Chinese decision-making apparatus (potentially including Xi Jinping himself) are deeply embarrassed by how desperately domestic enterprises want to import GPUs. Even when Trump offered to sell them, the Chinese government hesitated to allow these companies to make purchases. This implies that the RASA Act might provide China with a political excuse to 'force self-reliance,' rather than triggering the comprehensive economic war Patel fears. Schneider also questions whether China has the capacity or willingness to launch a full-scale retaliation—given its slowing economic growth, unemployment rates reaching 20%, and the fact that Trump continues to push forward despite the high costs of the war in Iran, suggesting the US government's pain tolerance may be underestimated.

Quellen:chinatalk.media

Forschung

Aviation Journalist Jeff Wise: The 'More Efficient' the AI Air Traffic Control System, the Thinner the Safety Redundancy May Become

Thema · FAA AI空管裁员Strukturell
2026-11-14 19:00 GMT+8

Air traffic control is undergoing a paradox: using AI to address staffing shortages may create new risks that require even more personnel to mitigate. Aviation journalist Jeff Wise traces this logical chain in a long-form investigation for WIRED. In May 2026, the US Federal Aviation Administration (FAA) lowered its target for active controllers from 14,633—set in 2024—to 12,563, citing "modern staffing models and scheduling tools." In September of the same year, the FAA deployed the SMART system, developed by Boston-based company Air Space Intelligence—a cloud platform that uses AI to assess flight schedules, weather, airport capacity, and airspace conditions to predict traffic flows and identify conflicts in advance. The contract totals $875 million over 12 years.

Wise points out that this logic is particularly dangerous in aviation. The industry has long been technologically conservative, not because it is backward, but because when human lives are at stake, time-tested reliability is preferable to cutting-edge unpredictability. Mary Schiavo, former Inspector General of the US Department of Transportation and an aviation lawyer, states plainly: certification standards and airworthiness requirements prove and demonstrate safety, which usually means incremental rather than radical change. Yet AI is penetrating the sector at an unprecedented pace. In September 2026, Joby Aviation completed a 3,200-mile cross-state autonomous flight—from Concord, California, to Dare County, North Carolina—navigating busy airspace and dodging storms without any human touching the controls. A safety pilot remained on board throughout, but never needed to intervene.

The technical risks are more subtle than they appear. Current research focuses on deterministic neural networks—models with frozen weights after training, which do not produce "hallucinations" like large language models. Wise cites a warning from Darren Cofer, a researcher at Collins Aerospace, in a 2021 paper: when such systems encounter situations not present in their training data, they generate "unexpected (and potentially dangerous) outputs." This is not a hallucination problem, but an out-of-distribution generalization issue—the model does not admit ignorance in scenarios it has "never seen," but instead provides uncalibrated outputs.

Political pressure is accelerating this process. A recently retired FAA employee, speaking anonymously to Wise, said: "We're introducing all this new technology to improve the system, and it's going to be a disaster. It's literally a Pandora's box we will come to regret." He attributes this to the "Trump/DOGE/Musk effect"—pushing through new technologies without fully understanding the complexity of the national airspace system. On October 13, nearly two dozen Democratic members of Congress wrote to the US Government Accountability Office (GAO), requesting a review of SMART due to concerns that it might "hallucinate safety-critical information."

On November 13 (Friday), the FAA issued a nationwide ground stop order, instantly throwing global airline schedules into chaos. The US government quickly stated there was no evidence of a cyberattack and no indication that SMART was linked to the communication failure. However, when asked if he could rule out an AI failure, US Secretary of Homeland Security Markwayne Mullin replied only: "There’s nothing certain right now."

Quellen:WIRED

Forschung

CrowdStrike Explains Centralized Control Principles for AI Traffic Gateways: What It Inspects and What It Cannot Block

Kurzfassung
2026-11-13 14:00 GMT+8

As AI agents shift from answering questions to executing actions, security teams face not just a visibility problem, but a control problem. In an explanatory article accompanying the general availability of its Falcon AI Gateway, CrowdStrike proposes a specific principle: make the approved path the only path. This means that teams use existing network controls to block applications and workloads from connecting directly to model providers, allowing only traffic routed through the gateway to reach the model endpoints.

The architecture operates in a straightforward manner: the gateway authenticates each request, routes it to the correct provider and model, applies usage limits, and enforces Falcon Guardian policies both before the request reaches the provider and when the response returns. Administrators can switch between Anthropic, Amazon Bedrock, Microsoft Foundry, and Google Vertex AI without reconfiguring every workload. Applications simply change their base URL to point to the gateway and use a gateway key, allowing them to continue using their original API formats.

The most valuable part of the article is the clear delineation of the gateway's capability boundaries. Falcon AI Gateway and the Falcon sensor share the same Falcon Guardian policy model, but enforce it at different layers: the sensor acts on the endpoint, including local processes, files, and commands run by the agent; the gateway acts on traffic routed through it and cannot stop local processes. This means that if an AI agent performs dangerous operations locally, the gateway will not see it; conversely, if the agent sends prompts containing sensitive data to the model, the sensor might not see it, but the gateway can intercept it.

Because traffic flows through a single point, the AI Kill Switch in Falcon Guardian can cut off agents' access to the model via traffic routed through Falcon AI Gateway. However, this capability is limited to traffic routed through the gateway—if an agent bypasses the gateway to call the model directly, the Kill Switch is ineffective. CrowdStrike states that this architecture stems from its experience running an internal AI Hub gateway, which has served tens of thousands of employees globally for years. A key lesson learned during this process was to keep the request path as short as possible, because every additional hop adds another failure point and latency.

For teams evaluating AI agent security architectures, the practical contribution of this article is providing a clear layered model: the traffic layer gateway handles prompt inspection, response filtering, usage control, and the Kill Switch; the endpoint layer sensor handles local processes, file operations, and command execution. The two cover different attack surfaces and are indispensable to each other. However, it must be remembered that this is entirely a vendor's description of its own product design, and its actual interception effectiveness and policy coverage have not been independently verified.

Quellen:CrowdStrike.com

Forschung

Analysis Claims Chinese AI Models Kimi K3 and DeepSeek V4 Trained on Banned Nvidia Chips

Thema · 中美AI出口管制Wesentlich
2026-11-14 17:54 GMT+8

According to a report by German tech outlet t3n, Will DePue, head of the non-profit research group Unemployment Labs, claims to have found "convincing evidence" that Moonshot AI's Kimi K3 and DeepSeek's V4 series models were trained on Nvidia Blackwell GPUs. The analysis relies on examining connection strengths within model weights, which are said to reflect the configuration of the underlying training hardware. These high-performance chips are subject to strict US export controls prohibiting direct sales to China.

DePue, a former OpenAI employee who worked on post-training for o3 and 4o models, stated that large language model weights contain signatures of the compute infrastructure used during training. This finding offers new technical clues regarding how Chinese AI laboratories may be circumventing restrictions on advanced compute access. Previously, White House technology advisor Michael Kratsios cited intelligence sources alleging illegal use of Nvidia chips by Moonshot AI, while Epoch AI estimated that approximately two million advanced chips flowed into China through indirect channels between 2024 and 2025.

TechCrunch notes that this analysis does not prove the models relied exclusively on Nvidia hardware nor necessarily indicate legal violations, as the chips could have been located in offshore data centers. Regulatory consequences from these findings remain unclear.

Quellen:t3n.de

Forschung

Speech-to-Text Company AssemblyAI: Distributing Across GPU Instance Families with Queue Retries Reduces Spot Inference Costs by 70%

Thema · AssemblyAIWesentlich
2026-11-14 03:35 GMT+8

Speech-to-text inference is a GPU-intensive workload. Each audio file triggers a deep learning pipeline comprising acoustic models, language models, and post-processing stages, running on NVIDIA GPUs to meet latency requirements. At scale, this means thousands of GPU instances are running continuously. AssemblyAI needed a way to run these workloads at scale while maintaining availability and latency commitments; using On-Demand Instances exclusively was not cost-effective.

Spot Instances offer significant discounts, but Amazon Elastic Compute Cloud (Amazon EC2) can reclaim instances with a two-minute notice when capacity is needed elsewhere. The challenge was: how to absorb Spot interruptions gracefully while maintaining service level agreement (SLA) commitments? AssemblyAI’s answer involves four pillars working in concert: ECS capacity providers mixing Spot and On-Demand instances, diversification of Spot across multiple GPU instance families, a queue-based retry architecture, and multi-region deployment to satisfy data residency requirements.

The most transferable insight is instance family diversification. AssemblyAI’s infrastructure spans three EC2 GPU instance families: g4dn, equipped with NVIDIA T4 GPUs, provides a cost-effective baseline for standard speech recognition models; g5, equipped with NVIDIA A10G GPUs, offers higher throughput for more demanding model architectures; and g6, featuring NVIDIA L4 GPUs, delivers improved performance-per-dollar for larger models, including large language model inference. By spreading demand across instance families and sizes, AssemblyAI draws from a significantly larger pool of available Spot capacity. These three families tap into independent capacity pools, so when one family experiences capacity pressure, the others remain available. The article explicitly states that operating at this scale, diversifying across instance families and sizes is the most effective lever for minimizing Spot interruption rates.

The recovery path during interruptions is equally critical. When a Spot instance receives a two-minute interruption notice, tasks that can complete within the window finish normally; tasks requiring more time are returned to the SQS queue, becoming immediately available to other workers. Another healthy GPU instance may pick up the message and complete the inference on a different instance type or in a different Availability Zone. To shorten recovery time, AssemblyAI creates custom AMIs containing model artifacts and container images, which take approximately two minutes to build, allowing new instances to start inference directly without pulling images or downloading models. For end customers, the impact of a Spot interruption is a slight increase in processing time, but the transcription ultimately completes successfully and is delivered as expected.

At the capacity scheduling level, AssemblyAI configures multiple capacity providers within a single capacity provider strategy in each ECS cluster. An On-Demand capacity provider serves as an always-on baseline, while a Spot capacity provider spans different GPU instance families and Availability Zones to handle demand spikes. Each capacity provider is associated with an EC2 Auto Scaling group, and ECS managed scaling automatically adjusts the number of GPU instances based on capacity provider weights and pending inference tasks. When the task queue grows, the Spot capacity provider requests additional instances from the diversified pool; when demand drops, instances scale in gracefully.

Quellen:Amazon Web Services

Forschung

German Host Mittwald Shuts Down Management Services Again Following Cyberattacks

Thema · MittwaldWesentlich
2026-11-12 22:08 GMT+8

German web hosting provider Mittwald shut down its customer center, administration interface mStudio, and all SSH connections again on November 12, 2026. This marks the second such outage within a week. The company had previously suffered a breach where attackers accessed customer projects through its Model Context Protocol (MCP) server interface by stealing user tokens.

According to reporting by TechGefrickel, Mittwald documented on its status page that attackers resumed their attacks on its infrastructure after last week's incidents, exploiting a vulnerability to compromise the company website. While Mittwald stated it has not yet found evidence of access to internal systems, the precautionary shutdown was implemented because the attackers appear to be professional actors using automated AI systems.

The disruption primarily affects agencies and freelancers who rely on Mittwald for hosting. Although customer websites and email services remain accessible, the lack of mStudio and SSH access prevents technical staff from performing deployments, domain changes, or database maintenance. Mittwald has not provided a timeline for restoring services, stating that further details will be released once countermeasures are complete.

Quellen:TechGefrickel

Forschung

NVIDIA Details Nova Driver vGPU Upstreaming Progress at KVM Forum, Targeting Linux 7.5/7.6

Kurzfassung
2026-11-13 19:09 GMT+8

NVIDIA engineer Simon Song presented the latest upstreaming progress for virtual GPU (vGPU) support in "Nova," the company's Rust-based open-source Linux kernel driver, at the 2026 KVM Forum in Vancouver, Canada. vGPU technology allows a single physical GPU to be split and shared among multiple virtual machines or containers, a critical component for cloud and data center deployments.

According to Phoronix, the effort is ongoing. Some NVIDIA vGPU enablement code for Nova is expected to land in the Linux 7.5 kernel version, but other remaining pieces may require waiting until Linux 7.6 or later to achieve full alignment of the kernel portion of vGPU handling. User-space tools, specifically vgpu-tools, are also being developed concurrently.

More advanced functionalities such as vGPU live migration and heterogeneous vGPU support remain in earlier stages of construction. This progress marks NVIDIA's gradual migration of core virtualization capabilities from its proprietary driver stack to the mainline kernel, potentially reducing reliance on closed-source out-of-tree drivers.

Quellen:phoronix.com

Forschung

NVIDIA xBeam Moves Radio Interferometry Correlation to GPUs, But Does Not Perform Imaging

Thema · NVIDIA xBeamKurzfassung
2026-11-14 05:17 GMT+8

The Event Horizon Telescope (EHT) links 13 radio observatory stations located in Hawaii, Chile, Spain, Mexico, and Antarctica via very-long-baseline interferometry (VLBI). Each telescope records high-speed voltage streams in two polarizations at rates up to 64 Gbps, generating petabytes of data. Correlation processing—multiplying voltage signals from any two stations channel-by-channel and accumulating them—is one of the most time-consuming steps in the entire data pipeline. In a technical blog post released in November 2026, NVIDIA introduced how cuPhoton.xBeam moves this step onto GPUs.

xBeam adopts an FX correlator architecture, using the same mathematical framework as the open-source CPU correlator DiFX. Its pipeline consists of five steps: reading VDIF frames and building a .vdix index to support random access; performing integer alignment and fringe rotation on each station according to the delay model in the job file; executing an FFT on the voltage stream to split it into frequency channels; calculating the product of the complex spectrum of each station pair's channel with its conjugate and accumulating over the integration time; and finally writing visibilities in DiFX SWIN or xbeam-v1 format. Among these, the FFT and cross-multiplication are computationally intensive and independent across channels, making them naturally suitable for the GPU's parallel execution model.

Multi-GPU scaling is achieved by sharding observation time. On a single node, multiple GPUs are launched simultaneously using the --devices parameter, with each GPU processing a segment of integration time. Across nodes, each process specifies its own time shard using --shard-count and --shard-index. After completion, the merge-shards command verifies that all shards originate from the same job, cover the complete time range, and have consistent SWIN headers before merging. NVIDIA reported that processing 10 TB of data on a Vera Rubin NVL72 system using 64 GPUs and GPUDirect Storage took 70 seconds, approximately 460 times faster than the x86 CPU baseline of 31,992 seconds (about 8.89 hours); read bandwidth increased from 20 GB/s on 4 GPUs to 170 GB/s on 64 GPUs. These figures are benchmark results self-reported by NVIDIA, and the article did not mention independent reproduction.

The article also delineates what xBeam does not do. It does not calculate delays based on station coordinates but instead directly uses the delay polynomials provided in the job file; it does not trace light paths around black holes; it does not perform calibration or image reconstruction. Downstream processes are still completed by tools such as HOPS, rPICARD, and eht-imaging. The Gaussian splatting method mentioned in the article—modeling emitting gas as compact glowing clumps moving along orbits to fit visibilities—is explicitly labeled as an independent component not currently included in cuPhoton. The paragraph describing AI agents orchestrating end-to-end pipelines outlines potential application directions rather than implemented features.

Quellen:NVIDIA Technical Blog

Forschung

GeekPark's Zhang Yongyi: The Window for AI Recording Hardware May Be Only Six Months Long, as Smartphone Makers Integrate All-Day Memory into Systems

Thema · AI录音硬件窗口期Wesentlich
2026-11-13 12:57 GMT+8

On November 4, 2026, Apple introduced two new AI features to the Apple Watch in the watchOS 27.3 developer beta: Live Rewind, which allows users to review conversations from the past 15 seconds at any time, and Siri Recap, which records conversations throughout the day and compiles them into summaries. These features require no additional hardware and have virtually no learning curve; Apple states that audio is neither recorded nor stored, and speaker identification is not performed. GeekPark author Zhang Yongyi draws a conclusion from this: the core value of AI recording hardware is being eroded by smartphone manufacturers, as recording and transcription no longer form a competitive moat. What is truly scarce is filtered context and the Agent entry point. Before smartphone makers fully integrate all-day memory into their systems, hardware companies may have a window of only about six months.

This judgment rests on the premise that recording hardware originally filled gaps that smartphones could not hear. Smartphones spend most of their time in pockets, and the current Android architecture does not allow apps to independently activate the microphone from the background. Carl Pei, founder of Nothing, once illustrated this with an example: while having coffee, someone says something worth noting, but by the time you take out your phone, the moment has passed. Chen Yusen, CEO of Qwen Office, also noted that today's AI speech-to-text technology is accurate enough that recording is no longer necessary. In other words, the remaining task for hardware is primarily to capture sound and convert it into text. However, this gap is currently being bridged by system-level solutions: Apple has placed audio capture on the watch, while OPPO shares a memory engine between its Xinliqiu (Heart Sphere) device and the smartphone system. Since smartphones accompany users all day—housing accounts, messages, schedules, and payments—the final output from Agents must ultimately return to the smartphone screen for user review.

A more critical change is that the entity reading these records is shifting from humans to Agents. Zhang Yongyi shared his own practical usage data: after connecting recordings from the OPPO Xinliqiu to Feishu (Lark) and importing them into tools such as Codex and DeepSeek Harness, he filtered thousands of dialogue segments down to approximately 200. Of these, about 30% were background noise, 40% were pleasantries like 'I'm sharing my screen,' and 20% were context useful only at that specific moment, meaning effective information accounted for less than 10%. This implies that filtering may be scarcer than simply remembering more. Which Agent reads these records depends on who holds the true entry point to intelligence: system Agents from smartphone makers do not need to wait for others to open interfaces, whereas companies without phones can only read from the outside. For hardware companies, becoming 'hardware peripherals for big tech's Personal Agents' is becoming a more pragmatic position.

Based on this, Zhang Yongyi provides a timeline judgment: by mid-2027, at least one domestic smartphone manufacturer should have integrated all-day recording as a built-in system capability, eliminating the need for users to purchase separate recording hardware. He also notes boundaries: the domestic gap may be slightly longer than in the United States because Recap is not yet available in China, and products like Bee and Muse have not entered the market; the variable lies in whether Huawei, OPPO, vivo, or Xiaomi will first integrate all-day memory into their systems. For those considering buying hardware, professional recording cards remain effective for meeting recordings; however, for those considering adding a new device solely for all-day recording, it may be wise to wait, as smartphone makers' watches and earbuds may fill this gap faster. It took smartphones over a decade to absorb cameras, maps, wallets, and contact lists into a single device; all-day recording will likely follow the same path, appearing first as standalone hardware before being integrated into smartphone systems.

Quellen:GeekPark

Forschung

AgentCore Runtime V2 Snapshot Mechanism: Recommendations to Reverse Cold-Start Optimization

Wesentlich
2026-11-14 06:02 GMT+8

Traditional cold-start advice is to delay imports and minimize initialization so the first request triggers actual work as quickly as possible. In an engineering blog post published on November 13, 2026, Amazon Web Services states that this advice needs to be reversed for AgentCore runtime V2: initialization runs only once and is frozen into the snapshot, while deferred static work in the handler incurs a cost with every new session.

The V2 mechanism works as follows: when creating or updating a runtime, the platform starts the container, waits for the health check to pass (within 120 seconds), clears the operating system file cache, freezes the microVM, and saves memory and disk state as a snapshot. Subsequently, each new session no longer pulls the image or runs initialization code; instead, it restores the snapshot into a new microVM, allowing the process to resume from the frozen point, with imported libraries, created objects, and clients already present in memory.

This brings three direct implications. First, complete memory restoration: any data held during the initialization phase, including content on RAM-backed file systems, becomes available at memory speed from the first access. Second, file reads do not carry over: because the file cache is cleared before freezing, files read during initialization must still be fetched from the image in restored sessions. Third, image files are loaded on demand: the first time a specific data block is read, it is pulled from snapshot storage and then cached closer to the host.

Therefore, the correct refactoring direction is to move static work into initialization. Create objects rather than merely importing modules, because object creation loads their dependencies and keeps them in active memory. Create boto3 clients during initialization; endpoint resolution and credential resolution are restored with the snapshot, and sockets automatically reconnect on the first call. Read static data such as prompts, configurations, and small indexes into memory-held objects rather than reading from the image with every request. For files that cannot be held within the process—such as independently running CLI toolchains or browser binaries—they can be copied to RAM disks like /dev/shm, or pre-warmed using build hooks during snapshot construction.

However, there is a trade-off that must be measured: data retained in memory during initialization increases snapshot size, which in turn increases restore time, and you pay for that memory in every session. The article explicitly recommends measuring end-to-end session start times before and after changes to confirm that latency is indeed reduced rather than merely shifted from the handler to the restore phase.

At the same time, certain values must remain in the handler. Random numbers, timestamps, credential refreshes, dynamic tool catalogs, and feature flags—if created during initialization, they would be frozen into the snapshot, causing every session to carry identical random states or expired time bases. The article draws an analogy to known pitfalls in AWS Lambda SnapStart and offers a module-level caching pattern with expiration times as a compromise for dynamic data.

Quellen:Amazon Web Services

Forschung

AMD Engineers Port 50 Models to Client GPUs in Three Days Using HRX Loom and AI Agents

Thema · AMD AI代理模型移植Kurzfassung
2026-11-13 23:53 GMT+8

Engineers Jeremy Fowers and Ben Vanik at U.S. semiconductor company AMD led an internal experiment leveraging the newly released ROCm HRX Loom v0.1 software stack and an AI agent workflow to port 50 popular Hugging Face models across 18 architectures to AMD client GPUs in approximately three to four days.

The experiment aimed to demonstrate the potential of accelerating the adaptation of new and niche models to AMD hardware through HRX Loom and agentic AI. According to AMD's report, the process achieved a medium speed-up of 1.5 to 2.1 times for 32K token prompt processing compared to Llama.cpp's ROCm backend, and a 1.9 times speed-up compared to its Vulkan backend.

HRX Loom remains an experimental project supporting CPU, GPU, and NPU execution. These results highlight AMD's effort to narrow the compatibility gap between its software ecosystem and mainstream open-source models via automated toolchains, though the specific performance figures have not been independently verified.

Quellen:phoronix.com

Forschung
Weiter →
Anmelden / Registrieren
LesenForschungRadarAnlageframework

Patel responds that every escalation-retaliation cycle in past US-China trade frictions ultimately led to compromise rather than total decoupling, because neither side could afford the cost of broken supply chains. He acknowledges he is arguing from the position of a 'devil's advocate' and concludes that one can be a hawk and still oppose the RASA Act, but it is difficult to oppose the MATCH Act.

From an industrial investment perspective, Patel's framework implies that if the MATCH Act passes, lithography equipment vendors like ASML will face structural risks to their revenue from China; meanwhile, Nvidia, cloud service providers, and compute rental platforms could continue to profit from permitted remote access. This judgment assumes that the US government indeed chooses the equipment blockade path over total decoupling, and that China does not retaliate with equivalent leverage using rare earths, market access, or other tools.

Two parallel systems will emerge in the future: one driven by human decision-making, the other by AI—in selected corridors near drone hubs and eVTOL vertiports. Michael McCormick, a professor of air traffic management at Embry-Riddle Aeronautical University, explains why traditional ATC cannot cope: "If you try to handle that volume and complexity of aircraft movements, the National Airspace System would be overwhelmed." But Wise’s conclusion is cautious: any computer system can be hacked; any AI system creative enough to devise novel solutions could also hallucinate crucial details. Without a human in the loop, seemingly obvious errors might go unnoticed until it is too late.

Wise documents a unfolding causal chain: the promise of AI efficiency is used to justify workforce reductions, which in turn weaken the oversight capacity needed to guard against AI failure modes. He quotes the retired FAA employee as the underlying tone of the entire piece—"We're introducing all this new technology to improve the system, and it's going to be a disaster"—and notes that in aviation, "efficiency" calculations can never look solely at labor costs: when a single error can endanger dozens of lives, redundancy is not waste, but safety itself.

Boundaries to emphasize: all performance data—including the 70% cost savings, approximately fourfold year-over-year growth in compute demand, and the approximately two-minute AMI creation time—are self-reported by AssemblyAI and AWS. The 70% figure compares against On-Demand pricing; the article notes that running entirely on On-Demand Instances would be cost-prohibitive but does not disclose the mixed configuration AssemblyAI previously operated. This pattern applies to short-lived, independently retriable GPU computing tasks; training jobs requiring long-term continuous state maintenance or inference scenarios maintaining context across requests do not fit the queue-retry assumption. Furthermore, Spot availability for g4dn, g5, and g6 in specific European regions is subject to AWS capacity fluctuations, meaning the effectiveness of the diversification strategy depends on the actual independence of these capacity pools.

For interactive scenarios, even a 2-second session startup is considered too slow. The article suggests initiating warm-up requests when the user opens a chat window or application, utilizing the time window before the user sends their first message to have the session ready. Running sessions respond in approximately 200 milliseconds, while new session startups take about 2 seconds—pre-warming hides the latter within user operation latency.

It should be noted that the aforementioned performance figures and mechanism descriptions come from Amazon Web Services' own explanations of its platform, representing vendor-reported data without independent third-party replication.