According to Bloomberg (paywalled), the Pentagon has reportedly scrapped its plan to allocate $2.5 billion in grants to Intel, causing the firm's stock to slip in extended-hours trading. From a report: The decision now leaves the U.S. Commerce Department, which is responsible for doling out the funds from the U.S. CHIPs and Science Act, to make up the shortfall, the news outlet said. The Commerce Dept. was initially only supposed to cover $1B of the $3.5B that Intel is slated to receive for advanced defense and intelligence-related semiconductors. The deal is slated to position Intel as the dedicated supplier for processors used for military and intelligence applications and could result in a Secure Enclave inside Intel's chip factory, the news outlet said. With the Pentagon reportedly pulling out, it could alter how much Intel and other companies receive from the CHIPs Act, the news outlet said.

An anonymous reader shares a report: If you're confused about what makes a PC an "AI PC," you're not alone. But finally have something of an answer: if it packs a GPU, a processor that boasts a neural processing unit and can handle VNNI and Dp4a instructions, it qualifies -- at least according to Robert Hallock, Intel's senior director of technical marketing. As luck would have it, that combo is present in Intel's current-generation desktop processors -- 14th-gen Core, aka Core Ultra, aka "Meteor Lake." All models feature a GPU, NPU, and can handle Vector Neural Network Instructions (VNNI) that speed some -- surprise! -- neural networking tasks, and the DP4a instructions that help GPUs to process video.

Because AI PCs are therefore just PCs with current processors, Intel doesn't consider "AI PC" to be a brand that denotes conformity with a spec or a particular capability not present in other PCs. Intel used the "Centrino" brand to distinguish Wi-Fi-enabled PCs, and did likewise by giving home entertainment PCs the "Viiv" moniker. Chipzilla still uses the tactic with "vPro" -- a brand that denotes processors that include manageability and security for business users. But AI PCs are neither a brand nor a spec. "The reason we have not created a category for it like Centrino is we believe this is simply what a PC will be like in four or five years time," Hallock told The Register, adding that Intel's recipe for an AI PC doesn't include specific requirements for memory, storage, or I/O speeds. "There are cases where a very large LLM might require 32GB of RAM," he noted. "Everything else will fit comfortably in a 16GB system."

An anonymous reader shares a report: Nvidia CEO Jensen Huang recently took to the stage to claim that Nvidia's GPUs are "so good that even when the competitor's chips are free, it's not cheap enough." Huang further explained that Nvidia GPU pricing isn't really significant in terms of an AI data center's total cost of ownership (TCO). The impressive scale of Nvidia's achievements in powering the booming AI industry is hard to deny; the company recently became the world's third most valuable company thanks largely to its AI-accelerating GPUs, but Jensen's comments are sure to be controversial as he dismisses a whole constellation of competitors, such as AMD, Intel and a range of competitors with ASICs and other types of custom AI silicon.

Starting at 22:32 of the YouTube recording, John Shoven, Former Trione Director of SIEPR and the Charles R. Schwab Professor Emeritus of Economics, Stanford University, asks, "You make completely state-of-the-art chips. Is it possible that you'll face competition that claims to be good enough -- not as good as Nvidia -- but good enough and much cheaper? Is that a threat?" Jensen Huang begins his response by unpacking his tiny violin. "We have more competition than anyone on the planet," claimed the CEO. He told Shoven that even Nvidia's customers are its competitors, in some cases. Also, Huang pointed out that Nvidia actively helps customers who are designing alternative AI processors and goes as far as revealing to them what upcoming Nvidia chips are on the roadmap.

A new report warns that a boom in computer chip manufacturing in the US could fuel demand for dirty energy, despite companies' environmental claims. The solution for manufacturers, surprisingly, might be to act more like other big tech companies chasing climate goals. From a report: New semiconductor factories being built in the US by four of the biggest manufacturers -- Intel, TSMC, Samsung, and Micron -- could use more than twice as much electricity as the city of Seattle once they're operational. These companies claim to run on renewable energy, but according to an analysis by nonprofit Stand.earth, that's not entirely true. Semiconductors happen to make up a big chunk of a device's carbon footprint. And unless companies turn to clean energy, they could wind up driving up greenhouse gas emissions as domestic chip manufacturing makes a comeback.

The CHIPS and Science Act, which passed in 2022, set aside $52.7 billion in funding for domestic chip manufacturing. Now, the four companies scrutinized in the report have plans to build megafactories in Arizona, Ohio, Oregon, Idaho, Texas, and New York. Each of those megafactories alone could use as much electricity as a medium-sized town, according to the report. Cumulatively, nine facilities could eventually add 2.1 gigawatts in new electricity demand. "We're not slowing down on any of our sustainability commitments, even with our recently announced investments," Intel said in an email.

In 2005 Gabe Rivera was a compiler software engineer at Intel — before starting the tech-news aggregator Techmeme. And last year his Threads profile added the words "This is a little self-serving, but I want all social networks to be as open as possible."

Friday Threads engineer Jesse Chen posted that it was Rivera's post when Threads launched asking for an API that "convinced us to go for it." And Techmeme just made its first post using the API, according to Chen. The Verge reports : Threads plans to release its API by the end of June after testing it with a limited set of partners, including Hootsuite, Sprinklr, Sprout Social, Social News Desk, and Techmeme. The API will let developers build third-party apps for Threads and allow sites to publish directly to the platform.
More from TechCrunch: Engineer Jesse Chen posted that the company has been building the API for the past few months. The API currently allows users to authenticate, publish threads and fetch the content they post through these tools. "Over the past few months, we've been building the Threads API to enable creators, developers, and brands to manage their Threads presence at scale and easily share fresh, new ideas with their communities from their favorite third-party applications," he said...

The engineer added that Threads is looking to add more capabilities to APIs for moderation and insights gathering.

The desert planet Arrakis in Frank Herbert's science fiction novel Dune is plausible, says Alexander Farnsworth, a climate modeler at the University of Bristol in England. According to Science News, the world would be a harsh place for humans to live, and they probably wouldn't have to worry about getting eaten by extraterrestrial helminths. From the report: For their Arrakis climate simulation, which you can explore at the website Climate Archive, Farnsworth and colleagues started with the well-known physics that drive weather and climate on Earth. Using our planet as a starting point makes sense, Farnsworth says, partly because Herbert drew inspiration for Arrakis from "some sort of semi-science of looking at dune systems on the Earth itself." The team then added nuggets of information about the planet from details in Herbert's novels and in the Dune Encyclopedia. According to that intel, the fictional planet's atmosphere is similar to Earth's with a couple of notable differences. Arrakis has less carbon dioxide in the atmosphere than Earth -- about 350 parts per million on the desert planet compared with 417 parts per million on Earth. But Dune has far more ozone in its lower atmosphere: 0.5 percent of the gases in the atmosphere compared to Earth's 0.000001 percent.

All that extra ozone is crucial for understanding the planet. Ozone is a powerful greenhouse gas, about 65 times as potent at warming the atmosphere as carbon dioxide is, when measured over a 20-year period. "Arrakis would certainly have a much warmer atmosphere, even though it has less CO2 than Earth today," Farnsworth says. In addition to warming the planet, so much ozone in the lower atmosphere could be bad news. "For humans, that would be incredibly toxic, I think, almost fatal if you were to live under such conditions," Farnsworth says. People on Arrakis would probably have to rely on technology to scrub ozone from the air. Of course, ozone in the upper atmosphere could help shield Arrakis from harmful radiation from its star, Canopus. (Canopus is a real star also known as Alpha Carinae. It's visible in the Southern Hemisphere and is the second brightest star in the sky. Unfortunately for Dune fans, it isn't known to have planets.) If Arrakis were real, it would be located about as far from Canopus as Pluto is from the sun, Farnsworth says. But Canopus is a large white star calculated to be about 7,200 degrees Celsius. "That's significantly hotter than the sun," which runs about 2,000 degrees cooler, Farnsworth says. But "there's a lot of supposition and assumptions they made in here, and whether those are accurate numbers or not, I can't say."

The climate simulation revealed that Arrakis probably wouldn't be exactly as Herbert described it. For instance, in one throwaway line, the author described polar ice caps receding in the summer heat. But Farnsworth and colleagues say it would be far too hot at the poles, about 70Â C during the summer, for ice caps to exist at all. Plus, there would be too little precipitation to replenish the ice in the winter. High clouds and other processes would warm the atmosphere at the poles and keep it warmer than lower latitudes, especially in the summertime. Although Herbert's novels have people living in the midlatitudes and close to the poles, the extreme summer heat and bone-chilling -40C to -75C temperatures in the winters would make those regions nearly unlivable without technology, Farnsworth says. Temperatures in Arrakis' tropical latitudes would be relatively more pleasant at 45C in the warmest months and about 15C in colder months. On Earth, high humidity in the tropics makes it far warmer than at the poles. But on Arrakis, "most of the atmospheric moisture was essentially removed from the tropics," making even the scorching summers more tolerable. The poles are where clouds and the paltry amount of moisture gather and heat the atmosphere. But the tropics on Arrakis pose their own challenges. Hurricane force winds would regularly sandblast inhabitants and build dunes up to 250 meters tall, the researchers calculate. It doesn't mean people couldn't live on Arrakis, just that they'd need technology and lots of off-world support to bring in food and water, Farnsworth says. "I'd say it's a very livable world, just a very inhospitable world."

Microsoft has outlined a new Windows API designed to offer a seamless way for game developers to integrate super resolution AI-upscaling features from Nvidia, AMD, and Intel. From a report: In a new blog post, program manager Joshua Tucker describes Microsoft's new DirectSR API as the "missing link" between games and super resolution technologies, and says it should provide "a smoother, more efficient experience that scales across hardware."

"This API enables multi-vendor SR [super resolution] through a common set of inputs and outputs, allowing a single code path to activate a variety of solutions including Nvidia DLSS Super Resolution, AMD FidelityFX Super Resolution, and Intel XeSS," the post reads. The pitch seems to be that developers will be able to support this DirectSR API, rather than having to write code for each and every upscaling technology.

The blog post comes a couple of weeks after an "Automatic Super Resolution" feature was spotted in a test version of Windows 11, which promised to "use AI to make supported games play more smoothly with enhanced details." Now, it seems the feature will plug into existing super resolution technologies like DLSS, FSR, and XeSS rather than offering a Windows-level alternative.

Intel's previously-unannounced Intel 10A (analogous to 1nm) will enter production/development in late 2027, marking the arrival of the company's first 1nm node, and its 14A (1.4nm) node will enter production in 2026. The company is also working to create fully autonomous AI-powered fabs in the future. Tom's Hardware: Intel's Keyvan Esfarjani, the company's EVP and GM and Foundry Manufacturing and Supply, held a very insightful session that covered the company's latest developments and showed how the roadmap unfolds over the coming years. Here, we can see two charts, with the first outlining the company's K-WSPW (thousands of wafer starts per week) capacity for Intel's various process nodes. Notably, capacity typically indicates how many wafers can be started, but not the total output -- output varies based on yields. You'll notice there isn't a label for the Y-axis, which would give us a direct read on Intel's production volumes. However, this does give us a solid idea of the proportionality of Intel's planned node production over the next several years.

Intel did not specify the arrival date of its coming 14A node in its previous announcements, but here, the company indicates it will begin production of the Intel 14A node in 2026. Even more importantly, Intel will begin production/development of its as-yet-unannounced 10A node in late 2027, filling out its roster of nodes produced with EUV technology. Intel's 'A' suffix in its node naming convention represents Angstroms, and 10 Angstroms converts to 1nm, meaning this is the company's first 1nm-class node. Intel hasn't shared any details about the 10A/1nm node but has told us that it classifies a new node as at least having a double-digit power/performance improvement. Intel CEO Pat Gelsinger has told us the cutoff for a new node is around a 14% to 15% improvement, so we can expect that 10A will have at least that level of improvement over the 14A node. (For example, the difference between Intel 7 and Intel 4 was a 15% improvement.)

OpenAI, Microsoft, Nvidia, and Jeff Bezos are all part of a pack of investors in a business "developing human-like robots," reports Bloomberg, "according to people with knowledge of the situation..."

At the startup — which is named "Figure" — engineers "are working on a robot that looks and moves like a human. The company has said it hopes its machine, called Figure 01, will be able to perform dangerous jobs that are unsuitable for people and that its technology will help alleviate labor shortages." Figure is raising about $675 million in a funding round that carries a pre-money valuation of roughly $2 billion, said the people, who asked not to be identified because the matter is private. Through his firm Explore Investments LLC, Bezos has committed $100 million. Microsoft is investing $95 million, while Nvidia and an Amazon.com Inc.-affiliated fund are each providing $50 million... Other technology companies are involved as well. Intel Corp.'s venture capital arm is pouring in $25 million, and LG Innotek is providing $8.5 million. Samsung's investment group, meanwhile, committed $5 million. Backers also include venture firms Parkway Venture Capital, which is investing $100 million, and Align Ventures, which is providing $90 million...

The AI robotics industry has been busy lately. Earlier this year, OpenAI-backed Norwegian robotics startup 1X Technologies AS raised $100 million. Vancouver-based Sanctuary AI is developing a humanoid robot called Phoenix. And Tesla Inc. is working on a robot called Optimus, with Elon Musk calling it one of his most important projects. Agility Robotics, which Amazon backed in 2022, has bots in testing at one of the retailer's warehouses. Bloomberg calls the investments in Figure "part of a scramble to find new applications for artificial intelligence."

Intel has landed Microsoft as a customer for its made-to-order chip business, marking a key win for an ambitious turnaround effort under Chief Executive Officer Pat Gelsinger. From a report: Microsoft plans to use Intel's 18A manufacturing technology to make a forthcoming chip that the software maker designed in-house, the two companies said at an event Wednesday. They didn't identify the product, but Microsoft recently announced plans for two homegrown chips: a computer processor and an artificial intelligence accelerator.

Intel has been seeking to prove it can compete in the foundry market, where companies produce custom chips for clients. It's a major shift for the semiconductor pioneer, which once had the world's most advanced chipmaking facilities and kept them to itself. These days, Intel is racing to catch up with companies like Taiwan Semiconductor Manufacturing Co., which leads the foundry industry. Microsoft, meanwhile, is looking to secure a steady supply of semiconductors to power its data-center operations -- especially as demand for AI grows. Designing its own chips also lets Microsoft fine-tune the products to its specific needs. "We need a reliable supply of the most advanced, high-performance and high-quality semiconductors," Microsoft CEO Satya Nadella said in a statement. âoeThat's why we are so excited to work with Intel."

An anonymous reader shared this report from PCWorld: The Standard Performance Evaluation Corporation, better known as SPEC, has invalidated over 2600 of its own results testing Xeon processors in the 2022 and 2023 version of its popular industrial SPEC CPU 2017 test. After investigating, SPEC found that Intel had used compilers that were, quote, "performing a compilation that specifically improves the performance of the 523.xalancbmk_r / 623.xalancbmk_s benchmarks using a priori knowledge of the SPEC code and dataset to perform a transformation that has narrow applicability."

In layman's terms, SPEC is accusing Intel of optimizing the compiler specifically for its benchmark, which means the results weren't indicative of how end users could expect to see performance in the real world. Intel's custom compiler might have been inflating the relevant results of the SPEC test by up to 9%...

Slightly newer versions of the compilers used in the latest industrial Xeon processors, the 5th-gen Emerald Rapids series, do not use these allegedly performance-enhancing APIs. I'll point out that both the Xeon processors and the SPEC 2017 test are some high-level hardware meant for "big iron" industrial and educational applications, and aren't especially relevant for the consumer market we typically cover.
More info at ServeTheHome, Phoronix, and Tom's Hardware.

An anonymous reader quotes a report from The Verge: The US government has committed $42 million to further the development of the 5G Open RAN (O-RAN) standard that would allow wireless providers to mix and match cellular hardware and software, opening up a bigger market for third-party equipment that's cheaper and interoperable. The National Telecommunications and Information Administration (NTIA) grant would establish a Dallas O-RAN testing center to prove the standard's viability as a way to head off Huawei's steady cruise toward a global cellular network hardware monopoly.

Verizon global network and technology president Joe Russo promoted the funding as a way to achieve "faster innovation in an open environment." To achieve the standard's goals, AT&T vice president of RAN technology Robert Soni says that AT&T and Verizon have formed the Acceleration of Compatibility and Commercialization for Open RAN Deployments Consortium (ACCoRD), which includes a grab bag of wireless technology companies like Ericsson, Nokia, Samsung, Dell, Intel, Broadcom, and Rakuten. Japanese wireless carrier Rakuten formed as the first O-RAN network in 2020. The company's then CEO, Tareq Amin, told The Verge's Nilay Patel in 2022 that Open RAN would enable low-cost network build-outs using smaller equipment rather than massive towers -- which has long been part of the promise of 5G.

But O-RAN is about more than that; establishing interoperability means companies like Verizon and AT&T wouldn't be forced to buy all of their hardware from a single company to create a functional network. For the rest of us, that means faster build-outs and "more agile networks," according to Rakuten. In the US, Dish has been working on its own O-RAN network, under the name Project Genesis. The 5G network was creaky and unreliable when former Verge staffer Mitchell Clarke tried it out in Las Vegas in 2022, but the company said in June last year that it had made its goal of covering 70 percent of the US population. Dish has struggled to become the next big cell provider in the US, though -- leading satellite communications company EchoStar, which spun off from Dish in 2008, to purchase the company in January. The Washington Post writes that O-RAN "is Washington's anointed champion to try to unseat the Chinese tech giant Huawei Technologies" as the world's biggest supplier of cellular infrastructure gear.

According to the Post, Biden has emphasized the importance of O-RAN in conversations with international leaders over the past few years. Additionally, it notes that Congress along with the NTIA have dedicated approximately $2 billion to support the development of this standard.

Microsoft appears to be readying its own DLSS-like AI upscaling feature for PC games. From a report: X user PhantomOcean3 discovered the feature inside the latest test versions of Windows 11 over the weekend, with Microsoft describing its automatic super resolution as a way to "use AI to make supported games play more smoothly with enhanced details." That sounds a lot like Nvidia's Deep Learning Super Sampling (DLSS) technology, which uses AI to upscale games and improve frame rates and image quality. AMD and Intel also offer their own variants, with FSR and XeSS both growing in popularity in recent PC game releases.

IDC, in a press release: A new forecast from IDC shows shipments of artificial intelligence (AI) PCs -- personal computers with specific system-on-a-chip (SoC) capabilities designed to run generative AI tasks locally -- growing from nearly 50 million units in 2024 to more than 167 million in 2027. By the end of the forecast, IDC expects AI PCs will represent nearly 60% of all PC shipments worldwide. [...] Until recently, running an AI task locally on a PC was done on the central processing unit (CPU), the graphics processing unit (GPU), or a combination of the two. However, this can have a negative impact on the PC's performance and battery life because these chips are not optimized to run AI efficiently. PC silicon vendors have now introduced AI-specific silicon to their SoCs called neural processing units (NPUs) that run these tasks more efficiently.

To date, IDC has identified three types of NPU-enabled AI PCs:
1. Hardware-enabled AI PCs include an NPU that offers less than 40 tera operations per second (TOPS) performance and typically enables specific AI features within apps to run locally. Qualcomm, Apple, AMD, and Intel are all shipping chips in this category today.

2. Next-generation AI PCs include an NPU with 40 to 60 TOPS performance and an AI-first operating system (OS) that enables persistent and pervasive AI capabilities in the OS and apps. Qualcomm, AMD, and Intel have all announced future chips for this category, with delivery expected to begin in 2024. Microsoft is expected to roll out major updates (and updated system specifications) to Windows 11 to take advantage of these high-TOPS NPUs.

3. Advanced AI PCs are PCs that offer more than 60 TOPS of NPU performance. While no silicon vendors have announced such products, IDC expects them to appear in the coming years. This IDC forecast does not include advanced AI PCs, but they will be incorporated into future updates. Michael Dell, commenting on X: This is correct and might be underestimating it. AI PCs are coming fast and Dell is ready.

The Biden administration on Thursday said leading AI companies are among more than 200 entities joining a new U.S. consortium to support the safe development and deployment of generative AI. From a report: Commerce Secretary Gina Raimondo announced the U.S. AI Safety Institute Consortium (AISIC), which includes OpenAI, Alphabet's Google, Anthropic and Microsoft along with Facebook-parent Meta Platforms, Apple, Amazon, Nvidia, Palantir, Intel, JPMorgan Chase and Bank of America. "The U.S. government has a significant role to play in setting the standards and developing the tools we need to mitigate the risks and harness the immense potential of artificial intelligence," Raimondo said in a statement.

An anonymous reader quotes a report from Reuters: Intel is delaying the construction timeline for its $20 billion chipmaking project in Ohio amid market challenges and the slow rollout of U.S. grant money, the Wall Street Journal reported on Thursday. Its initial timeline had chip-making starting next year. Construction on the manufacturing facilities now is not expected to be finished until late 2026, the report said, citing people involved in the project. Shares of the chipmaker were last down 1.5% in extended trading.

"We are fully committed to completing the project, and construction is continuing. We have made a lot of progress in the last year," an Intel spokesperson said, adding that managing large-scale projects often involves changing timelines. Uncertain demand for its chips used in the traditional server and personal computer markets had led the company to forecast revenue for the first quarter below market estimates late last month. This came as a shift in spending to AI data servers, dominated by rivals Nvidia and aspiring AI competitor Advanced Micro Devices sapped demand for traditional server chips -- Intel's core data center offering.

It's "a free and open-source, software-defined storage platform," according to Wikipedia, providing object storage, block storage, and file storage "built on a common distributed cluster foundation". The charter advisory board for Ceph included people from Canonical, CERN, Cisco, Fujitsu, Intel, Red Hat, SanDisk, and SUSE.

And Nite_Hawk (Slashdot reader #1,304) is one of its core engineers — a former Red Hat principal software engineer named Mark Nelson. (He's now leading R&D for a small cloud systems company called Clyso that provides Ceph consulting.) And he's returned to Slashdot to share a blog post describing "a journey to 1 TiB/s". This gnarly tale-from-Production starts while assisting Clyso with "a fairly hip and cutting edge company that wanted to transition their HDD-backed Ceph cluster to a 10 petabyte NVMe deployment" using object-based storage devices [or OSDs]...) I can't believe they figured it out first. That was the thought going through my head back in mid-December after several weeks of 12-hour days debugging why this cluster was slow... Half-forgotten superstitions from the 90s about appeasing SCSI gods flitted through my consciousness...

Ultimately they decided to go with a Dell architecture we designed, which quoted at roughly 13% cheaper than the original configuration despite having several key advantages. The new configuration has less memory per OSD (still comfortably 12GiB each), but faster memory throughput. It also provides more aggregate CPU resources, significantly more aggregate network throughput, a simpler single-socket configuration, and utilizes the newest generation of AMD processors and DDR5 RAM. By employing smaller nodes, we halved the impact of a node failure on cluster recovery....

The initial single-OSD test looked fantastic for large reads and writes and showed nearly the same throughput we saw when running FIO tests directly against the drives. As soon as we ran the 8-OSD test, however, we observed a performance drop. Subsequent single-OSD tests continued to perform poorly until several hours later when they recovered. So long as a multi-OSD test was not introduced, performance remained high. Confusingly, we were unable to invoke the same behavior when running FIO tests directly against the drives. Just as confusing, we saw that during the 8 OSD test, a single OSD would use significantly more CPU than the others. A wallclock profile of the OSD under load showed significant time spent in io_submit, which is what we typically see when the kernel starts blocking because a drive's queue becomes full...

For over a week, we looked at everything from bios settings, NVMe multipath, low-level NVMe debugging, changing kernel/Ubuntu versions, and checking every single kernel, OS, and Ceph setting we could think of. None these things fully resolved the issue. We even performed blktrace and iowatcher analysis during "good" and "bad" single OSD tests, and could directly observe the slow IO completion behavior. At this point, we started getting the hardware vendors involved. Ultimately it turned out to be unnecessary. There was one minor, and two major fixes that got things back on track.
It's a long blog post, but here's where it ends up:

• Fix One: "Ceph is incredibly sensitive to latency introduced by CPU c-state transitions. A quick check of the bios on these nodes showed that they weren't running in maximum performance mode which disables c-states."

• Fix Two: [A very clever engineer working for the customer] "ran a perf profile during a bad run and made a very astute discovery: A huge amount of time is spent in the kernel contending on a spin lock while updating the IOMMU mappings. He disabled IOMMU in the kernel and immediately saw a huge increase in performance during the 8-node tests." In a comment below, Nelson adds that "We've never seen the IOMMU issue before with Ceph... I'm hoping we can work with the vendors to understand better what's going on and get it fixed without having to completely disable IOMMU."

• Fix Three: "We were not, in fact, building RocksDB with the correct compile flags... It turns out that Canonical fixed this for their own builds as did Gentoo after seeing the note I wrote in do_cmake.sh over 6 years ago... With the issue understood, we built custom 17.2.7 packages with a fix in place. Compaction time dropped by around 3X and 4K random write performance doubled."

The story has a happy ending, with performance testing eventually showing data being read at 635 GiB/s — and a colleague daring them to attempt 1 TiB/s. They built a new testing configuration targeting 63 nodes — achieving 950GiB/s — then tried some more performance optimizations...

An anonymous reader quotes a report from Wired: As more companies ramp up development of artificial intelligence systems, they are increasingly turning to graphics processing unit (GPU) chips for the computing power they need to run large language models (LLMs) and to crunch data quickly at massive scale. Between video game processing and AI, demand for GPUs has never been higher, and chipmakers are rushing to bolster supply. In new findings released today, though, researchers are highlighting a vulnerability in multiple brands and models of mainstream GPUs -- including Apple, Qualcomm, and AMD chips -- that could allow an attacker to steal large quantities of data from a GPU's memory. The silicon industry has spent years refining the security of central processing units, or CPUs, so they don't leak data in memory even when they are built to optimize for speed. However, since GPUs were designed for raw graphics processing power, they haven't been architected to the same degree with data privacy as a priority. As generative AI and other machine learning applications expand the uses of these chips, though, researchers from New York -- based security firm Trail of Bits say that vulnerabilities in GPUs are an increasingly urgent concern. "There is a broader security concern about these GPUs not being as secure as they should be and leaking a significant amount of data," Heidy Khlaaf, Trail of Bits' engineering director for AI and machine learning assurance, tells WIRED. "We're looking at anywhere from 5 megabytes to 180 megabytes. In the CPU world, even a bit is too much to reveal."

To exploit the vulnerability, which the researchers call LeftoverLocals, attackers would need to already have established some amount of operating system access on a target's device. Modern computers and servers are specifically designed to silo data so multiple users can share the same processing resources without being able to access each others' data. But a LeftoverLocals attack breaks down these walls. Exploiting the vulnerability would allow a hacker to exfiltrate data they shouldn't be able to access from the local memory of vulnerable GPUs, exposing whatever data happens to be there for the taking, which could include queries and responses generated by LLMs as well as the weights driving the response. In their proof of concept, as seen in the GIF below, the researchers demonstrate an attack where a target -- shown on the left -- asks the open source LLM Llama.cpp to provide details about WIRED magazine. Within seconds, the attacker's device -- shown on the right -- collects the majority of the response provided by the LLM by carrying out a LeftoverLocals attack on vulnerable GPU memory. The attack program the researchers created uses less than 10 lines of code. [...] Though exploiting the vulnerability would require some amount of existing access to targets' devices, the potential implications are significant given that it is common for highly motivated attackers to carry out hacks by chaining multiple vulnerabilities together. Furthermore, establishing "initial access" to a device is already necessary for many common types of digital attacks. The researchers did not find evidence that Nvidia, Intel, or Arm GPUs contain the LeftoverLocals vulnerability, but Apple, Qualcomm, and AMD all confirmed to WIRED that they are impacted. Here's what each of the affected companies had to say about the vulnerability, as reported by Wired:

Apple: An Apple spokesperson acknowledged LeftoverLocals and noted that the company shipped fixes with its latest M3 and A17 processors, which it unveiled at the end of 2023. This means that the vulnerability is seemingly still present in millions of existing iPhones, iPads, and MacBooks that depend on previous generations of Apple silicon. On January 10, the Trail of Bits researchers retested the vulnerability on a number of Apple devices. They found that Apple's M2 MacBook Air was still vulnerable, but the iPad Air 3rd generation A12 appeared to have been patched.
Qualcomm: A Qualcomm spokesperson told WIRED that the company is "in the process" of providing security updates to its customers, adding, "We encourage end users to apply security updates as they become available from their device makers." The Trail of Bits researchers say Qualcomm confirmed it has released firmware patches for the vulnerability.
AMD: AMD released a security advisory on Wednesday detailing its plans to offer fixes for LeftoverLocals. The protections will be "optional mitigations" released in March.
Google: For its part, Google says in a statement that it "is aware of this vulnerability impacting AMD, Apple, and Qualcomm GPUs. Google has released fixes for ChromeOS devices with impacted AMD and Qualcomm GPUs."

PC and microchip companies struggling to get consumers to replace pandemic-era laptops offered a new feature to crowds this week at CES: AI. From a report: PC and chipmakers including AMD and Intel are betting that the so-called "neural processing units" now found in the latest chip designs will encourage consumers to once again pay for higher-end laptops. Adding additional AI capabilities could help take market share from Apple. "The conversations I'm having with customers are about 'how do I get my PC ready for what I think is coming in AI and going to be able to deliver,'" said Sam Burd, Dell Technologies' president of its PC business. Chipmakers built the NPU blocks because they can achieve a high level of performance for AI functions with relatively modest power needs. Today there are few applications that might take full advantage of the new capabilities, but more are coming, said David McAfee, corporate vice president and general manager of the client channel business at AMD.

Among the few applications that can take advantage of such chips is the creative suite of software produced by Adobe. Intel hosted an "open house" where a handful of PC vendors showed off their latest laptops with demos designed to put the new capabilities on display. Machines from the likes of Dell and Lenovo were arrayed inside one of the cavernous ballrooms at the Venetian Convention Center on Las Vegas Boulevard.

Asus is back with another Zenbook Duo, the latest $2,161 device in its range of dual-screened laptops. But rather than including a small secondary display above this laptop's keyboard like previous Duos, the revamped version for 2024 has two equally sized 14-inch screens. The Verge has more: They're both OLED, with resolutions of up to 2880 x 1800, aspect ratios of 16:10, and a maximum refresh rate of 120Hz. Between them, they offer a total of 19.8 inches of usable screen real estate. It's a similar approach to the one Lenovo took with last year's dual-screen Yoga Book 9i, albeit with a couple of tweaks. Like Lenovo, Asus gives you a choice of typing on the lower touchscreen via a virtual keyboard or by using a detachable physical Bluetooth keyboard. But what's different here is that Asus' keyboard has a trackpad built in, so you don't have to use it in combination with an on-screen trackpad.

Asus envisages you using the new Zenbook Duo in a few different configurations. There's a standard laptop mode, where the bottom screen is entirely covered by a traditional keyboard and trackpad. Or you can rest the keyboard on your desk and have the two screens arranged vertically for "Dual Screen" mode or horizontally for "Desktop" mode. Finally, there's "Sharing" mode, which has you ditch the keyboard entirely and lay the laptop down on a flat surface with both its screens facing up and away from each other, presumably so you can share your work with a colleague sitting across the desk from you. Naturally, having launched a year later than its competitor, the Asus Zenbook Duo is also packed with more modern hardware. It can be specced with up to an Intel Core Ultra 9 185H processor and 32GB of RAM, up to 2TB of storage, and a 75Wh battery. Connectivity includes two Thunderbolt 4 ports, a USB-A port, HDMI out, and a 3.5mm jack, and the laptop can be used with Asus' stylus.