Artificial Intelligence - Page 29

NVIDIA's new Blackwell AI servers reportedly ran into issues last year with overheating and an architectural flaw, and it seems these issues haven't gone away, leaving big customers (paying big bucks) stranded, moving back to Hopper AI servers.

In a new report from The Information, we're learning that the first significant shipments of NVIDIA's new GB200 AI servers have big customers experiencing overheating, and glitching issues, with the big problem being the "way chips connect". Big customers like Amazon, Google, Meta, and Microsoft have reportedly cut down their orders because of the issues.

Back in October 2024, NVIDIA CEO Jensen Huang said "we had a design flaw in Blackwell" noting that it was "100% NVIDIA's fault" and not anything to do with the rumored issues with TSMC's new CoWoS advanced packaging. A few months later in December 2024, we reported that NVIDIA GB200 AI server mass production and its peak shipments could be delayed until Q2 or even Q3 2025... and here we are with more issues.

When pondering the future of tech: few images have been more symbolic in popular culture than the flying car. Whether it's the flying DeLorean in Back to the Future II or Howard Stark's hovering Cadillac in the halls of the Stark Expo. 

While not a science-fiction film, the future is here at CES 2025. International brand RICTOR, and Chinese aerospace company Xpeng Aero HT, both showcased unique, inventive vehicle offerings to take you from the ground to the skies.  

Rictor's Skyrider X1 is a 'moped-cum-quadcapter', meaning: half-moped, half-quadcopter. The vehicle's propulsion very much resembles a drone, with propellers attached to a cockpit somewhat reminiscent of a modern bullet train. In terms of functionality, that comparison remains somewhat true. The Skyrider X1 is capable of achieving 100km/h flight speeds, with a maximum of 40 minutes of flight time up to 200 meters above the ground. For comparison, the average multi-rotor drone is capable of 20-30 minutes of flight. With flight speeds anywhere from 70-100 km/hr. 

The journey of OpenAI will be a beautiful mess of a movie when it eventually gets made, with the sister of CEO Sam Altman filing a lawsuit claiming he sexually assaulted her, raped her, and more heinous acts when she was a minor. The weird thing is, Sam posted the entire thing to X...

Annie Altman has sued her brother accusing the OpenAI boss of sexually assaulting her when she was a minor, filing a complaint with the Missouri federal court claiming that Sam committed "numerous acts of rape, sexual assault, sexual abuse, molestation, sodomy, and battery" that spanned from 1997 to 2006, with the abuse starting when she was just 3 years old.

OpenAI CEO Sam Altman posted on X with a statement from his mom, brothers, and himself, where he said that all of his sister's claims were "utterly untrue" and that the Altman family have tried to help Annie in "many ways" over the years, including direct financial help.

KIOXIA AiSAQ (All-in-Storage ANNS with Product Quantization) is a research project created to transition from a "mostly in-storage" solution to an all-in-storage solution for AI. It's a groundbreaking technology because it offloads compressed vectors in large datasets from DRAM to storage. With memory limitations often being the key factor in AI workload speed and efficiency, AiSAQ exponentially improves performance.

AI is all about data, lots of it. There is so much data that it's often hard to wrap your head around the complexity of it all. AI models and datasets cover a range of things, including documents, images, music, and more. Retrieving specific data involves Approximate Nearest Neighbor Search (ANNS), vectors, and indices. These are usually stored on fast memory, like DRAM - KIOXIA AiSAQ moves this to the SSD.

KIOXIA's initial research paper on the technology from 2024 calls it a "novel method of index data placement." However, with only around 10MB of DRAM being used in an AiSAQ system regardless of the scale of datasets, with millisecond-order latency, it sounds more like a game changer.

SK hynix plans to show off prototypes of its new AI memory chips at CES 2025 this week, with the tease of 16-Hi HBM3E memory that stacks 16 x DDR5 DRAM chips on top of one another, for up to 48GB HBM3E capacity.

SK hynix will indeed showcase its new 16-Hi HBM3E memory chips at CES 2025, with the highest capacity of up to 48GB. The new 16-Hi HBM3E memory chips started making waves in November 2024 when SK hynix said they were in development, with confidence that the yields would be as good as 12-layer HBM3E which is the most advanced model in mass production.

SK hynix's new 16-Hi HBM3E memory chips will be manufactured using an advanced mass reflow-molded underfill process, enabling the 16-layer stack while effectively controlling chip warpage and maximizing thermal performance. SK hynix has also said that in order to hit 20-layer stacks with the process without hybrid bonding, a next-generation technology for bonding stacked chips.

Microsoft is celebrating the New Year, President Trump returning to the White House, and a new "golden opportunity" for American technology and economic competitiveness with the announcement of $80 billion of funds at the ready to create new AI data centers.

The company has marked $80 billion to build AI data centers ready for AI workloads of the future, in order to "train AI models and deploy AI and cloud-based applications around the world".

Microsoft Vice Chair and President Brad Smith explained in a blog post: "As we look into the future, it's clear that artificial intelligence is poised to become a world-changing GPT. AI promises to drive innovation and boost productivity in every sector of the economy. The United States is poised to stand at the forefront of this new technology wave, especially if it doubles down on its strengths and effectively partners internationally".

Samsung Electronics has announced a new multi-year partnership with Instacart, which will find its way inside of the new wave of Samsung Bespoke refrigerators, using AI to let you order new groceries directly in front of your fridge.

Samsung will use its in-house food recognition technology -- AI Vision Inside -- together with Instacart's advanced product-matching API, to enable simple, convenient, and tech-focused grocery shopping experiences for Samsung customers. The new 2025 Bespoke refrigerator models will have different types of screens, including the 32-inch AI Family Hub and a new 9-inch AI Home, with AI Vision inside.

With this combo, it'll enable consumers to replenish their favorite grocery items through Instacart by using Vision AI food recognition technology. Samsung says that with AI Vision Inside, consumers can not only manage their food inventory, but in the future, the experience will be amplified by enabling replenishment through Instacart, directly from your new refrigerator.

SK hynix has teased that it will be showcasing its innovative AI memory technologies at CES 2025, with a bunch of C-level executives including the CEO will be in attendance.

SK hynix Chief Marketing Officer, Justin Kim, said: "We will broadly introduce solutions optimized for on-device AI and next-generation AI memories, as well as representative AI memory products such as HBM and eSSD at this CES. Through this, we will publicize our technological competitiveness to prepare for the future as a Full Stack AI Memory Provider".

Ahn Hyun, Chief Development Officer at SK hynix, said: "As SK hynix succeeded in developing QLC (Quadruple Level Cell)-based 61 TB products in December, we expect to maximize synergy based on a balanced portfolio between the two companies in the high-capacity eSSD market. The company will also showcase on-device AI products such as LPCAMM2 and"ZUFS 4.04, which improve data processing speed and power efficiency to implement AI in edge devices like PCs and smartphones. The company will also present CXL and PIM (Processing in Memory) technologies, along with modularized versions, CMM (CXL Memory Module)-Ax and AiMX, designed to be core infrastructures for next-generation data centers".

SK Group Chairman Chey Tae-won is rumored to be meeting with NVIDIA CEO Jensen Huang at CES 2025 next week to discuss all things HBM4 memory for future-gen AI GPUs.

In a new post from The Korea Times, we're learning that the two executives would be meeting at CES 2025 to discuss HBM4, which the SK hynix boss was asked to expedite HBM4 memory by 6 months by Jensen for NVIDIA's next-gen Rubin R100 AI GPUs coming in 2025.

At SK Group's AI conference on November 4, Chey said that Jensen has been consistently requesting more HBM memory chips with each new GPU release. At the conference, the NVIDIA CEO sent in a video message where Jensen talked about the important role that HBM memory has in the breakthroughs from NVIDIA.

SK hynix has begun preparations for the mass production of its new 16-Hi HBM3E memory, weeks after unveiling the world's first 48GB 16-Hi HBM3E memory at the SK AI Summit in Seoul.

In a new report from ETNews, we're learning that SK hynix is now integrating new equipment for 16-Hi HBM3E memory production, and optimizing its existing facilities. Production tests have reportedly already begun, with supply of the new 16-Hi HBM3E memory expected to flow in 1H 2025.

An industry insider told ETNews: "It is the initial line preparation process for 16-layer HBM3E mass production. I understand that the results of the major process tests are also coming out well".

SK hynix is delivering a huge HBM memory order to Broadcom in 2025, which it will use on its new AI chips that it is making for cloud service providers (CSPs) including Google, Meta, and TikTok parent company ByteDance.

SK hynix already has NVIDIA as a major client of its industry-leading HBM memory chips, but sources of TheElec report that the South Korean memory giant will deliver new HBM to Broadcom in 2H 2025, with the US chipmaker to use the AI memory chips in new AI chips for Google, Meta, and ByteDance.

The deal will see SK hynix push its HBM production capacity to 160,000 - 170,000 wafers in 2025, up from the previous target of 140,000 - 150,000. SK hynix is providing most of the HBM memory chips it makes to NVIDIA for use in H200, GB200, B200, GB200, and other AI chips and AI servers, but now Broadcom wants some of that delicious HBM memory pie for its CSP clients.

NVIDIA is expected to unveil its next-generation GB300 AI server platform at GTC 2025 which takes place between March 17-20, with some huge surprises from the new AI GPUs.

In new leaks from UDN, we're learning that NVIDIA is preparing its new B300 AI GPUs will scream along with up to 1400W of power, with 1.5x the performance in FP4 performance on a single AI GPU over the B200 AI GPU. We can expect the HBM capacity of each GPU to increase from the max of 192GB HBM3E on the B200, to a much bigger 288GB of HBM3E on B300.

NVIDIA's current GB200 AI uses 192GB of HBM3E in an 8-Hi stack configuration, but the next-gen GB300 will use 288GB of HBM3E in a larger 12-Hi stack, which SK hynix has been working on for a while now.

Micron has updated the industry on its next-generation HBM4 and HBM4E memory, with mass production expected to start in 2026 ready for next-gen AI GPUs like the NVIDIA Rubin R100.

The AI memory market only has a few players with SK hynix leading the HBM market, Samsung and Micron have started spooling up production of their HBM3, HBM3E, and future-gen HBM4 and HBM4E memory for the insatiable requirements of the AI industry... it's mostly NVIDIA vacuuming up all HBM memory chips made.

Sanjay Mehrotra, President and Chief Executive Officer of Micron said: "Leveraging the strong foundation and continued investments in proven 1β process technology, we expect Micron's HBM4 will maintain time to market and power efficiency leadership while boosting performance by over 50% over HBM3E. We expect HBM4 to ramp in high volume for the industry in calendar 2026".

Apple has been caught in hot water after its new generative AI software produced a factually incorrect headline that was discovered by the BBC.

The BBC recently published a report stating it officially complained to Apple after an Apple Intelligence-generated headline of a BBC news story that was pushed out to iPhones last week contained misinformation. The headline stated Luigi Mangione, the man who was arrested for the murder of UnitedHealthcare CEO Brian Thomson, shot himself. The AI-powered summary was discovered by the BBC, which contacted Apple to raise the concern and "fix the problem".

Notably, Apple's AI that generated the summary was otherwise correct in the other information it provided, such as South Korean President Yoon Suk Yeol having his office raided by police, and the overthrow of Bashar al-Assad's regime in Syria. However, the BBC isn't the only publication to have suffered from Apple's AI spitting out headlines that are factually incorrect, as three articles on different topics from the New York Times were distributed in one notification, and one of those headlines stated the Israeli prime minister, Benjamin Netanyahu was "arrested".

NVIDIA's new GB200 rack-mounted AI servers are still experiencing issues, with a new report suggesting that the supply chain requires more time for optimization and adjustment, and it could see mass production and peak shipments delayed until Q2 or even Q3 2025.

In a new report from TrendForce, their latest report says that the supply chain needs more time for GB200 rack servers, and that it's mostly because of the higher design specifications of the GB200 rack, including its requirement for high-speed interconnect interfaces, and thermal design power (TDP), which "significantly exceed market norms".

TrendForce is now projecting that mass production and peak shipments of NVIDIA GB200 rack servers are "unlikely to occur until between Q2 and Q3 of 2025". The NVIDIA GB rack series includes the GB200 and new GB300 models, which feature even more complex tech and higher production costs.

NVIDIA has taken to its website to unveil its latest compact generative AI system that's designed to provide a development platform from commercial AI developers to mere hobbyists looking to explore the capabilities of generative AI.

NVIDIA, the company powering the mass push into AI-infused software, has announced the new NVIDIA Jetson Orin Nano Super Developer Kit, which provides as much as a 1.7x leap in generative AI interference performance, along with a 70% increase in performance to 67 INT8 TOPS. Additionally, NVIDIA has increased the Jetson Orin Nano Super's memory bandwidth by as much as 50% compared to its predecessor, bringing its total memory bandwidth to 102GB/s.

Moreover, if you are upset because you already own a Jetson Orin Nano and now want the performance of the Super variant, NVIDIA can get you partially the way there as the company explains in its announcement the software updates to the Super variant will benefit performance on the Jetson Orin Nano. NVIDIA writes the developer kit consists of a Jetson Orin Nano 8GB system-on-module (SoM), along with a reference carrier board designed for prototyping AI applications. More specifically, the SoM features an NVIDIA Ampere-based architecture GPU, a 6-core Arm CPU, and support for up to four cameras.

NVIDIA's next-gen B300 and GB300 AI GPUs are currently being tested, with some news that AOS's DrMOS were experiencing "severe thermal issues".

In a new post on X and a more detailed post on Medium, insider Ming-Chi Kuo says that the DrMOS chips are currently being tested for NVIDIA's new GB300/B300 system, and that they're AOS's 5x5 and MPS's 5x6, with Kuo's supply chain survey indicating that AOS's DrMOS were experiencing severe thermal issues.

Kuo said: "With GB300/B300 still some time away from mass production, This event doesn't mean that AOS can't eventually benefit from GB300/B300. However, to avoid repeating delays seen with GB200/B200, NVIDIA's efforts to seek alternative 5x5 DrMOS suppliers or to accelerate testing of MPS's 5x6 DrMOS could impact the market's high expectations for AOS's orders in the near term".

Google DeepMind has formally introduced Genie 2, a "foundation world model" that can generate playable 3D environments based on a single prompt image. According to Google, Genie 2 represents a massive improvement over Genie 1, which could generate 2D worlds. Simulating a 3D world also includes actions such as jumping and swimming.

With complex character interaction, physics, lighting, and more, it's an impressive demonstration that Google says could be used for creative workflows and creative workflows for prototyping interactive experiences. A game developer could create detailed concept art or even a pre-rendered image and then use Genie 2's model to see how it might look as a playable experience.

Even more interesting is the diverse trajectories; the simulation and what you see will change depending on the input or movement. For example, taking a different path at a fork in the road in a first-person fantasy Genie 2 creation will lead to a different environment.

Samsung has still yet to pass HBM3E memory qualification from NVIDIA, and now it seems that won't be happening this year as it is "realistically impossible" for Samsung to supply its HBM3E to NVIDIA this year.

In a new report from Korean media outlet Daily Korea, we're learning that Samsung's supply of HBM3E memory to NVIDIA this year looks "impossible" in both 8-layer and 12-layer HBM3E form. The reason for the delay is that Samsung is reportedly still unable to meet NVIDIA's requirements for chip performance.

Daily Korea's source said: "It is realistically impossible for Samsung Electronics to supply 8-layer and 12-layer HBM3E to NVIDIA this year. The reason for the delay in supply is that we were unable to meet NVIDIA's requirements for chip performance".

Google's CEO Sundar Pichai has weighed in on the recent conversation surrounding AI and its development - Are leading AI companies struggling to scale AI models due to a lack of high-quality training data?

Reports indicate that top AI development companies such as Anthropic, OpenAI, and Google will soon hit a wall that will hinder the explosive performance jumps of AI models such as OpenAI's GPT, or Google's Gemini. However, according to former Google CEO Eric Schmidt, these assumptions about AI model scaling are unfounded, saying, "There's no evidence that the scaling laws have begun to stop. They will eventually stop, but we're not there yet."

These assertions that AI models will still continue to develop rapidly in the immediate future were subtly reiterated by OpenAI CEO Sam Altman, who wrote on X, "There's no wall". Current Google CEO Sundar Pichai joined in on the conversation at The New York Times' Dealbook summit, where he said, "I think the progress is going to get harder when I look at '25. The low-hanging fruit is gone. The hill is steeper." Pichai recognized there will be challenges ahead that simply can't be solved by throwing more compute at the AI model.