Artificial Intelligence - Page 31

AI servers of the future are expected to use over 1000 kW of power, with the rise of NVIDIA's next-gen Rubin Ultra AI GPU and its new HBM4 memory coming that will see a rise in power consumption of AI servers.

The current best AI servers and clusters with a rack density of 130 kW to 250 kW of power according to an analyst by data center firm Vertiv. But these numbers are going to skyrocket with Blackwell Ultra and Rubin AI servers using between 250 kW and 900 kW of power with up to 576 GPUs in a single rack expected in 2026-2027.

After that, NVIDIA will have its future-gen, beefed-up Rubin Ultra AI GPUs inside of new AI servers in 2028-2029+ that will use 1000 kW or more. Vertiv notes that "densification roadmaps point to extreme peak rack densities of 1mW+ resulting in estimated new build average rack density increasing across varying workloads applications and market segments".

ASUS has just unveiled its next-gen infrastructure solutions at SC24, with a product showcase that shows the company is working with NVIDIA and Ubitus/Ubilink to prove the "immense computational power of supercomputers, using AI-powered avatar and robot demonstrations that leverage the newly-inaugurated data center".

ASUS has revealed its new ASUS AI POD, a complete rack solution powered by the NVIDIA GB200 NVL72 AI server platform, integrating GPUs, CPUs and switches in seamless, high-speed direct communication, enhancing the training of trillion-parameter LLMs and enables real-time inference.

The key points from ASUS at SC24 include:

TensorWave is a high-end cloud service provider (CSP) that uses AMD AI hardware, and has just announced that it is working on making the world's largest GPU clusters built on AMD Instinct MI300X, MI325X, and MI350X AI accelerators.

In a statement, TensorWave CEO Darrick Horton said that the company is working towards building the world's largest AMD AI GPU cluster, in full he said: "with our 1 Gigawatt of capacity, we will massively scale deployments in 2025 and build the world's largest AMD GPU clusters, powered by MI300X, MI325X, and MI350, and beyond".

Horton continued: "these clusters will be the first to leverage Ultra Ethernet fabrics and will offer truly compelling performance with unmatched scale and efficiency".

NVIDIA has announced its new GB200 NVL4, a new module that is a bigger extension to the original GB200 Grace Blackwell Superchip AI solution, doubling the CPU and GPU power while also increasing memory.

The new NVIDIA GB200 NVL4 features 2 x Blackwell GB200 GPUs configured onto a larger board with 2 x Grace CPUs, with the new module designed as a single-server solution with 4-way NVLink domain controlling a larger 1.3TB pool of coherent memory.

NVIDIA promises performance improvements over the previous-gen Hopper GH200 NVL4 with the new GB200 NVL4 module of 2.2x in simulation, and a 1.8x increase in Training and Inference performance. We know that the GB200 Grace Blackwell Superchip uses around 2700W of power, with the larger GB200 NVL4 solution to use around 6000W of power.

Google's AI chatbot, Gemini, has seemingly threatened a school student in Michigan after they were using it to discuss the challenges and solutions faced by older adults.

The student is Vidhay Reddy, who spoke to CBS News about the strange interaction with Gemini. Reddy said he was incredibly shaken by the message and that it "definitely scared me, for more than a day, I would say." Reddy is a 29-year-old student who was using Gemini to learn more about what he was studying for homework, and when the AI seemingly snapped out of nowhere, Reddy and his sister, Sumedha Reddy, were shocked at the response.

For those interested in reading the full transcript of the conversation Reddy had to arrive at the threatening response from Gemini, check out this link here. As for Gemini, Google has stated that Gemini has safety features that should prevent the AI from engaging with users in a negative, provocative, or harmful manner. However, Google does state, along with every other big tech company behind a major AI platform, that these AI systems are prone to hallucinating, which is when the AI responds with something nonsensical.

A contractor of the Department of Defence has showcased the impressive defensive capabilities of a new machine gun that's designed to combat the rise of drones throughout warfare.

The new machine gun was developed by Allen Control System (ACS) and is officially called the "Bullfrog." The new weapon was tested out at the Technology Readiness Experimentation event earlier this year, which is an event that enables contractors to display their technology to the Pentagon for potential acquisition. The new machine gun uses an AI-powered system to target small drones in the air and shoot them down, with the entire system being designed to combat the meteoric rise in uncrewed aircraft throughout warfare.

ACS cofounder Steve Simoni spoke to Wired and said the Russian and Ukraine war demonstrated the increase in the use of drones, and after seeing the footage of the war and Ukrainians firing AK-47s at drones, they decided to investigate a potential robotics program capable of automatically targeting drones. ACS claims its Bullfrog is cheaper to use compared to other drone-centered solutions such as microwave or laser weapons systems.

Phone scammers are only becoming more frequent and sophisticated, especially with the rise of AI-powered tools that can manipulate a scammer's voice into sounding like a loved one. However, AI can also be used to thwart scammers, and the easiest way to get back at scammers is to keep them on the phone, as that costs them money.

UK telecommunications company Virgin Media O2 knows the next best thing besides shutting down scammers completely is keeping them on the phone, as it means that the scammer isn't on the phone with a potential victim who could lose their life savings. In what seems like a stroke of genius, Virgin Media has unveiled a new AI entity designed to take on the persona of a confused grandmother.

The AI sounds like a grandmother, responds like a grandmother, and is simply designed to waste as much of the scammer's time as possible with constant ramblings about topics that have nothing to do with what the scammer is trying to do.

NASA has been capturing valuable data on our planet for many years now, and according to a recent blog post from Microsoft, the US space agency has gathered more than 100 petabytes of data on various Earth-related scientific domains.

Microsoft has announced that it's partnering with NASA to create a new AI chatbot that will be trained on NASA's mountain of Earth data. The goal of the new AI chatbot is to make it easier to shift through the plethora of data NASA has gathered, making it easier to answer questions about the planet. The new AI chatbot will be called Earth Copilot, and its purpose is to "democratize" access to scientific data, while also developing a keen understanding of the information within NASA's databases.

Tyler Bryson, Microsoft's corporate vice president of health and public sector industries, explained in the announcement that those interested in acquiring Earth-related answers through scientific data have to know how to navigate technical interfaces, understand how to correctly interpret data, and have already acquired other various specialized skills. However, with the implementation of Earth Copilot, users will be able to engage with the chatbot through natural language and then be presented with answers cleanly and concisely. This would considerably speed up the time it takes to gain an insight.

NVIDIA has just published some juicy benchmarks of its new Blackwell AI GPUs in MLPerf v4.1 AI training workloads, where against Hopper the new Blackwell chips are up to 2.2x faster. Check it out:

The new Blackwell AI GPUs have set all 7 per-accelerator records using its Nyx AI supercomputer, which packs DGX B200 systems. The Nyx AI supercomputer is 2.2x faster in Llama 2 70B (Fine-Tuning) versus Hopper H100, 2x faster in GPT-3 175B (Pre-Training) versus Hopper H100, and it also demolished the entire set of workloads inside of the MLPerf Training 4.1 suite.

NVIDIA explains: "The first Blackwell training submission to the MLCommons Consortium - which creates standardized, unbiased and rigorously peer-reviewed testing for industry participants - highlights how the architecture is advancing generative AI training performance. For instance, the architecture includes new kernels that make more efficient use of Tensor Cores. Kernels are optimized, purpose-built math operations like matrix-multiplies that are at the heart of many deep learning algorithms".

Samsung has reportedly started HBM4 development with the company rumored to offer customized HBM4 memory solutions to tech giants Meta and Microsoft.

The company has started developing "custom HBM4" memory that will be "customized" for Microsoft and Meta according to Korean outlet MK. According to a semiconductor company official, Microsoft "has its own artificial intelligence (AI) chips called Maia 100 and Meta Artemis". Samsung Electronics is the best partner for these big tech companies because it has a memory division and an LSI division that can design computational chips directly".

Most of the big tech companies have their own AI data centers and AI supercomputers, so there is a huge need for the reduction of chips purchased (which saves lots of money). This is why these companies design and use AI accelerator chips separately, while buying AI chips made by the likes of NVIDIA and AMD.

Artificial intelligence companies such as OpenAI, Microsoft, and other big players such as Elon Musk's X-based AI Grok are all in a race to develop artificial general intelligence (AGI). But what if no one makes it to the finish line because it's not economically viable?

This suggestion has been made by several experts and is a theory held by Gary Marcus, a cognitive scientist and AI skeptic. Marcus says the current AI market is a bubble and that people haven't yet realized the valuations of companies such as OpenAI and Microsoft are predicated on the bet that the underlying technology powering AI systems, Large Language Models (LLMs), will grow in power with scale. However, that might not be the case.

The Information reported last week that OpenAI researchers discovered the company's latest AI model, codenamed Orion, was a noticeably less improvement in performance compared to the leaps made from GPT-3 to GPT-4. IIya Sutskever, co-founder and former chief science officer of OpenAI, previously told Reuters that improvements from scaling AI models bigger have plateaued, which harms the general theory that the bigger the AI model the more powerful.

Japan's SoftBank is the first to receive NVIDIA's new Blackwell GPU design for its AI supercomputer, with plans to also use Grace Blackwell chips to "further its supercomputer" says NVIDIA.

In an AI event held in Tokyo, Japan on Wednesday, both SoftBank Group CEO Masayoshi Son and NVIDIA CEO Jensen Huang were in attendance. The two CEOs held a "fireside chat" at the event, with Jensen telling a story that Son once offered to lend him money to buy NVIDIA, as its value wasn't understood in the market at the time.

Jensen said, laughing: "He wanted to lend me money to buy NVIDIA. All of it. Now I regret not taking it".

TSMC's 5nm and 3nm production lines are fully loaded until the first half of 2025, thanks to the continued strong demand of AI chips, and Qualcomm + MediaTek smartphone chips.

In a new report from Ctee, we're learning that AI chip orders will keep TSMC 5nm lines over 100% utilized in the first half of 2025 (which is normally a weaker time of the year for the semiconductor giant). TSMC is also expected to see CoWoS advanced packaging output at over 36,000 wafers per month by the end of 2024, and 90,000+ per month by the end of 2025.

Ctee reports that the "breaking point" will be concentrated in Q4 2024, which will alleviate the current pressure on advanced packaging production capacity. The supply chain suggest that NVIDIA's new B200 series GPUs are gradually increasing in volume and production of 200,000 B200 chips will be made by the end of 2024, with B300A launching in Q3 2025.

Supermicro is in even more trouble: financial strife, a large DOJ investigation, its Malaysian factory has been affected, and its expansion has been suspended. Now... Supermicro has stuffed a huge order of NVIDIA's new GB200 NVL72 AI servers, with a Taiwanese firm stepping up to help.

In a new report from UDN, we're learning that the first wave of substantive transfer orders were received by Taiwanese factories after the "US Supermicro storm". Wiwynn, a subsidiary of Winstron Group, was approached for assistance post-Supermicro issues, after shipping the highest-end NVL72 full container to NVIDIA filled with GB200 AI chips, and the order volume was "huge" adds UDN.

Wiwynn hasn't commented on the news just yet, with speculation that the transfer of orders from Supermicro would be led by NVIDIA, but this time the transfer of orders "comes from a large Supermicro customer".

Microsoft could be quietly planning a complete rebrand for its AI suite that is currently under the monkier "Copilot," according to reports from last week.

The report cites a template file for Microsoft's Group Policy Object Editor or AppPrivacy.adml, which makes reference to something called "Windows Intelligence". Notably, Apple decided to brand its AI features and services under the umbrella "Apple Intelligence," which was recently debuted across various Apple products in iOS 18.2 - and respective OS updates. As for Microsoft, the company may be internally referring to its suite of AI features as "Windows Intelligence," and this is no indication of a looming rebranding for Copilot.

Additionally, Microsoft has only been using the "Copilot" branding for a little over a year, and switching to "Windows Intelligence" would be an apparent play-by-play copy of Apple's approach. Microsoft recently rolled out a fleet of new AI-focussed laptops under the "Copilot" branding. These Copilot laptops are available in several prominent brands, such as ASUS, Acer, Dell, and more. Abandoning the Copilot branding would harm the marketing for each of these products.

Quanta Computer is reportedly gearing up to boost production of its AI servers in the US, as the Taiwan AI server giant has its US factories "fully loaded" until the end of 2025.

NVIDIA's new entry with GB200 has been trial-produced in factories in both Taiwan and the United States, with volume of GB200 slowly increasing in Q1 2025 for Quanta Computer, noting that "the situation is very good". Yang Qiling, executive deputy general manager of Quanta attended the recent Taiwan Semiconductor Industry Association (TSIA) annual meeting and was interviewed, which is where the new information is coming from.

Qiling noted that Quanta has production capacity on both the east and west coasts othe United States, but will continue to expand its two factories in the US "in response to demand in the future". There was concern about US manufacturing issues that might arise in the new Trump administration, but "Quanta is more comfortable with it".

NVIDIA is reportedly already redirecting orders away from Super Micro Computer, (SCMI) after the company has found itself embroiled in controversy... its accounting firm resigning on the spot, and a DOJ investigation.

SCMI's biggest client is NVIDIA, and its second-biggest client is an entity backed by NVIDIA... pretty much forcing NVIDIA to start redirecting orders away from SMCI to other suppliers. In a new report from DigiTimes, we're learning that NVIDIA has started redirecting its orders placed with SCMI, to other suppliers.

SCMI is now looking at allegations of financial malfeasance, a preliminary DOJ investigation, and the sudden resignation of its accounting firm Ernst & Young which was in the process of auditing SCMI. NVIDIA shifting orders for its new B200 and GB200 AI servers away from Super Micro Computer, with Team Green being its biggest client... isn't looking good at all.

Researchers at "Which?," a UK organization that's a subsidiary of the Consumers' Association, discovered that some air fryers may be sharing their owners personal information and sending it back to China for marketing purposes.

The researchers at the organization looked at three air fryer brands and found that all of them requested permission to record audio through the user's phone, but didn't specify the reason why the microphone permission was necessary. More specifically, the air fryers were from brands such as Aigostar and Xiaomi, both of which were found to be sending their buyers personal data back to servers located in China.

Air fryers weren't the only device the team looked at as it also conducted an analysis on smartwatches, smart speakers and smart TVs. Starting with smart speakers, the researchers found that Amazon's Echo device was more consumer-friendly as it has the option to skip various requests to share data. However, Amazon and Google accounts are required to use the Echo Pop or Nest Mini, which contain trackers that users cannot opt out of.

Amazon's latest feather in the cap of its Prime Video service is a feature that recaps a series you're watching that's powered by AI.

It's called X-Ray Recaps, and the capability uses generative AI to create brief and easily digestible summaries of seasons of TV shows, Amazon explains.

We've all been there, and what with gaps of a year between seasons (or longer in some cases), it's easy to forget what was going on in the various subplots and threads of relationships in the show you were watching.

Meta's plan to construct a nuclear-powered data center to support its exponentially growing AI ventures was foiled when a species of bees was discovered.

According to a new report from The Financial Times, Meta was negotiating a deal with an owner of an already established nuclear power plant, but the deal fell apart due to environmental and regulatory problems. The deal was for the operator of the nuclear power plant to provide electricity to a presumably nearby soon-to-be-constructed data center. The publication didn't reveal the location of the proposed data center, or the nuclear power plant operator, but it did reveal a rare species of bees was discovered on the land the project was meant to be built.

The FT report stated Mark Zuckerberg discussed this point at a company meeting last week, where he said the discovery of the bee species complicated the project, and contributed to the ultimate cancellation. The unfortunate setback for Meta won't stop the company from finding a new building location, along with a clean energy supply. With the rise in AI data centers and the incredible amounts of power needed to fuel them, nuclear power has been thrown back into the conversation as a viable solution.