News

Utran’s PCIe 5.0 card holds 28 M.2 SSDs, reaching a total 224TB capacity Delivers 109GB/s read speed using Broadcom switch and advanced cooling Ideal for AI workloads, but overkill for most enterprise storage needsUtran Technology has introduced a new PCIe 5.0 add-in card which feels more like a GPU than a storage solution.Unveiled at Computex 2025, the device can host up to 28 NVMe Gen5 M.2 8TB SSDs in a single slot, delivering 109GB/s sequential read speed and a total storage capacity of 224TB.Two versions of the 28x M.2 Host Card: HM-5281A and HM-5282A will be available - both use the Broadcom AtlasII PEX89144 switch to handle internal bandwidth and connectivity. The HM-5281A uses a single PCIe Gen5 x16 upstream link, while the HM-5282A doubles that with two x16 links, bringing total bandwidth up to 1024 GT/s.Surprise hot plug supportCooling comes via a high-pressure fan and radiator combo. Although it has a dense footprint, the layout is built for rack-scale deployment. In theory, eight cards could deliver nearly 1.8PB of flash inside a single server.Both models run on an EPS 8-pin connector and support surprise hot plug, meaning the system can detect and manage the 28 M.2 drives even if they’re swapped in unexpectedly. This is particularly useful for testing or dynamic environments. You'd need to take care doing so in real-world deployments though, especially as the card itself isn't hot-swappable.28x M.2 Host Card also lacks card-level power loss protection, so you’d need to rely on SSDs that include their own safeguards.The card does, however, support USB terminal control for firmware updates and system monitoring.Supported operating systems include Windows, Windows Server, and Linux, making it relatively flexible at the software level.It's hard to argue with the raw numbers - 109GB/s read speed and sub-millisecond latency are unquestionably impressive - but outside of certain HPC or AI use cases, it's frankly difficult to see a wide audience. Even in dense environments, this level of performance might outpace most storage needs.Utran says it plans to begin shipping its 28x M.2 Host Card in summer 2025.Via Tom's HardwareYou might also likeThese are the fastest SSDs you can buy right nowAnd these are the largest SSDs and hard drives on the marketHere's why 100TB+ SSDs will play a huge role in ultra LLMs in the near future

Here's an AI-government collaboration of a less… unsettling variety than some. Google DeepMind is teaming up with the National Hurricane Center (NHC) for tropical cyclone season. The AI research lab claims it can predict hurricane paths and intensities with at least the same accuracy as traditional methods.

NHC forecasters have already begun using DeepMind's AI model. Google says they're designed to support, not replace, human NHC forecasters. (Although President Trump's National Weather Service cuts have already reduced its headcount.) The company is also careful to repeatedly describe its models as "experimental."

Google claims that its models offer fewer trade-offs than physics-based predictions. The more accurate those methods are at forecasting a hurricane's path, the worse they are at predicting its intensity. (And vice versa.) The company says its experimental system offers "state-of-the-art" accuracy for both.

DeepMind backs that up with data from real-life storms over the last two years. On average, its five-day hurricane track prediction gets 87 miles closer to the storm's actual path than ENS, a widely used traditional model. Google's was comparable to a 3.5-day prediction model. In other words, it's like gaining an extra 1.5 days of warning with the same level of confidence. The company says such an improvement typically takes over a decade to achieve.

Google

Alongside the NHC collab, Google is launching a new website that you can try. Now in a public preview, Weather Lab lets you see the AI storm predictions. It lets you view both live and historical predictions. You can even compare them to physics-based models to see how the AI version measures up.

It's important not to treat Weather Lab's experimental forecasts as official. But the website could come in handy if you live in Hurricane Alley. You can check it out now.This article originally appeared on Engadget at https://www.engadget.com/ai/google-deepmind-is-sharing-its-ai-forecasts-with-the-national-weather-service-173506456.html?src=rss