Artificial Intelligence - Page 34

MSI has officially rolled out a new feature on its new range of X870 and Z890 motherboards designed for Intel and AMD's new generation of CPUs, and the new feature enables more performance to be squeezed out of onboard NPUs.

The new motherboards for AMD's Ryzen 9000 series and the Intel Core Ultra (Series 2) come with brand-new chipsets that have a range of hardware improvements. To accompany these hardware improvements MSI has overhauled its BIOS interface into what its calling Click BIOS X, and during a recent tour of MSI's motherboard factory in Shenzhen, China, we were able to spend some time with some setups featuring Intel's new Arrow Lake CPUs and the new BIOS interface.

All of MSI's new range of motherboards come with the new BIOS interface, and one of the built-in features is AI Boost. This was a particularly impressive feature as the setting within the BIOS enabled overclocking for the NPU, which MSI claimed to improve AI performance and efficiency by up to 5%. According to MSI, enabling the new feature will provide the user with "faster data processing, enhanced AI performance, improved efficiency in AI tasks, better multi-tasking capabilities, and maximized hardware utilization."

OpenAI has just received one of the first engineering builds of the NVIDIA DGX B200 AI server, posting a picture of their new delivery on X:

Inside, the NVIDIA DGX B200 is a unified AI platform for training, fine-tuning, and inference using NVIDIA's new Blackwell B200 AI GPUs. Each DGX B200 system has 8 x B200 AI GPUs with up to 1.4TB of HBM3 memory and up to 64TB/sec of memory bandwidth. NVIDIA's new DGX B200 AI server can pump out 72 petaFLOPS of training performance, and 144 petaFLOPS of inference performance.

OpenAI Sam Altman is well aware of the advancements of NVIDIA's new Blackwell GPU architecture, recently saying: "Blackwell offers massive performance leaps, and will accelerate our ability to deliver leading-edge models. We're excited to continue working with NVIDIA to enhance AI compute".

Artificial intelligence systems have now become so sophisticated they are being awarded Nobel prizes for their academic achievements, and now AI has gained its second Nobel prize, but this time for protein prediction.

Geoffrey Hinton, a computer scientist whose work on deep learning is the foundation of all AI models currently used today, was awarded a Nobel prize, along with Princeton University professor John Hopfield. Both researchers were awarded the Nobel Prize in physics for their contributions to deep learning technologies, which have become the underpinning technology we now broadly call AI.

Now, AI has done it again, with a Nobel Prize being given to Demis Hassabis, the cofounder and CEO of Google DeepMind, and John M. Jumper, a director at DeepMind, for the creation of an AI capable of accurately predicting the structures of protein. Half of the Nobel Prize is awarded to Hassabis and Jumper, and the other half is awarded to David Baker, a professor of biochemistry at the University of Washington, who was recognized for his work on computational protein design. Each of the prize winners shares a $1 million pot.

We knew it was coming, but now it's official: NVIDIA is teaming with Foxconn to build Taiwan's most powerful supercomputer powered by its new Blackwell AI GPU architecture.

NVIDIA and Foxconn announced the new Hon Hai Kaohsiung Super Computing Center at its recent Hon Hai Tech Day, which will be built around NVIDIA's groundbreaking new Blackwell GPU architecture. The new AI supercomputer will feature GB200 NVL72 AI servers, with a total of 64 racks and 4608 Tensor Core GPUs.

The company is expecting to see over 90 exaflops of AI performance, making the new Taiwan-based supercomputer the fastest on the island. Foxconn has plans to use the supercomputer once it's operational, to power breakthroughs in cancer research, large language model development, and smart city innovations, positioning Taiwan as a global leader in AI-driven industries.

AMD is reportedly set to make next-gen, high-performance HPC AI chips at TSMC's new fab in Arizona, joining as the second major company making next-gen chips... the other is Apple.

In a new post from insider Tim Culpan, who reports that AMD is "lined up to produce high-performance computing chips from TSMC Arizona, making the American fabless chip designer another client for the new US facility" according to his sources.

Culpan explains that production is already in the planning phase, with tape out and manufacturing of AMD's next-gen HPC chips expected to kick off at TSMC's 5nm process node in 2025. Apple is the first customer of TSMC's fresh new fab in Arizona, which will be producing some of the A16 processors that go inside of the new iPhone 16 family of handsets.

"We're not going to hit the climate goals anyway because we're not organized to do it." That's former Google CEO Eric Schmidt responding to a question about the rise in energy consumption due to the AI boom at SCSP's inaugural AI+Energy Summit.

AI is putting a strain on energy grids everywhere due to the sheer amounts of power required to run complex generative AI systems, so it's a definite issue.

Eric Schmidt's response is somewhat cynical but indicative of the debate surrounding how governments, corporations, and people everywhere should be dealing with climate change and its potentially devastating impacts. His response wasn't simply a shoulder shrug, as Schmidt confirmed that energy concerns surrounding AI "will be a problem."

Meta's smart glasses have been converted into a facial recognition device that enables the user to identify random people in real time.

The conversion of Meta's smart glasses into a facial recognition device came from two Harvard students who called the glasses I-XRAY. Here's how it works. The students took advantage of the smart glasses livestreaming directly to Instagram feature and combined that with an AI program that monitors the video live stream to identify any faces within the video. Images of the faces are then captured and fed into public databases, which result in phone numbers, names, addresses, and more personal information being fed back to the wearer of the glasses through a phone app.

AnhPhu Nguyen posted the above video detailing the process of creating the glasses, and within the video, you can see them being used to identify classmates and, perhaps more shockingly, strangers in public, which the students pretended to know based on the information the device was able to obtain on the individual. For those concerned about the potential impact of releasing such a product to the market, fear not, as the students behind the project said the glasses were created to raise awareness about potential privacy issues with smart glasses.

NVIDIA CEO Jensen Huang spoke with CNBC earlier today, commenting on its Blackwell AI GPUs and that they're in full production as planned, and that the "demand is insane, everyone wants to be first, everyone wants to have the most".

Jensen told CNBC: "the thing that we have done with Blackwell, and what we have announced, this new AI infrastructure generation every single year, and so we're going to update our platform every single year and the reason foe that is if we can increase the performance as we've done with Hopper to Blackwell by 2-3x each year we're effectively increasing the revenues or the throughput of our customers on these infrastructures by a couple (2-3x) each year, decrease, or how you can think about it decreasing cost every 2 or 3 years".

Jensen continued: "reducing energy efficiency every single year, and so at a time when the technology is moving so fast, it gives us an opportunity to triple down and to really drive the innovation cycle, so that we can increase capabilities, increase our throughput, decrease our cost, decrease our energy consumption, and so we're on a path to do that, and everything's on track".

Rambus has provided more details on its upcoming HBM4 memory controller, which offers some huge upgrades over current HBM3 and HBM3 memory controllers.

JEDEC is still finalizing the HBM4 memory specifications, with Rambus teasing its next-gen HBM4 memory controller that will be prepared for next-gen AI and data center markets, continuing to expand the capabilities of existing HBM DRAM designs.

Rambus' new HBM4 controller will pump over 6.4Gb/s speeds per pin, which is faster than the first-gen HBM3 and has more bandwidth than faster HBM3E memory using the same 16-Hi stack and 64GB max capacity design. HBM4 starting bandwidth is at 1638GB/sec (1.63TB/sec) which is 33% faster than HBM3E and 2x faster than HBM3.

NVIDIA's upcoming GB200 NVL72 AI server development has some big challenges ahead of it, which mostly stem from the insane 132kW TDP requirement, making it the highest-power-consuming server in HISTORY.

In a new post on Medium, analyst and insider Ming-Chi Kuo said that NVIDIA has halted the development of its GB200 NVL36x2 AI server (the dual-rack 72 GPU version) which you can read more about in the links below. Moving back to the 'biggest challenges' in NVL72 development from the 132kW thermal design point (TDP), with NVIDIA and its supply chain requiring more time to solve "unprecedented technology issues".

Kuo points out that it's important to note that the TDP "refers to average power consumption during continuous operation. If poor design leads to peak power consumption (electrical design point (TDP) as NVIDIA calls it) exceeding TDP, two or more sidecars may be required. This would not only increase cooling design complexity and production difficulties but also negate NVL72's data center space-saving advantage".

NVIDIA is reportedly halting development of its dual-rack 72-way GB200-based NVL36x2 AI server, shifting focus into the single-rack NVL36 and NVL72 AI servers, according to analyst Ming-Chi Kuo..

In a new post on Medium, Ming-Chi Kuo deep dives into why NVIDIA is nixing its NVL36x2 AI server, with the single-rack GB200 NVL36 will "maintain its original development and shipment plans". NVIDIA originally plans to have 3 different GB200 projects: NVL36, NVL72, and NVL36x2 under simultaneous development.

However, due to "uncertainties in NVL36 development, simultaneous development of two 72 GPU versions (NVL72 and NVL36x2) is even more challenging" writes Kuo. The analyst breaks it down on his Medium post, but I've included some of the highlights below.

We all know that Taiwan leads the planet when it comes to semiconductor manufacturing, with global semiconductor leader TSMC absolutely dominating... so much so, that China is at least 10 years behind Taiwan according to a new report.

Taiwan's chairman of the National Science Council, Wu Chengwen, said that he doubts recent statements that China's semiconductor technology is only 3 years behind Taiwan. The Education and Culture Committee of the Legislative Yuan invited Wu Chengwen to provide a special report on the "Sustainable Scientific Research and Policy Communication" and to prepare for some questions.

Democratic Progressive Party legislator Wu Peiyi made a point that according to recent Japanese media reports, that after a Japanese semiconductor survey company dismantled Huawei smartphones, that China has caught up and that its chip making technology strength is only 3 years behind Taiwan... Chengwen was asked if this was true, to which he replied: "I actually have doubts". The gap should be over 10 years, especially with TSMC knocking on the door of 2nm mass production.

TikTok parent company ByteDance will be training its next-generation AI model using Huawei AI chips, according to three people familiar with the matter, reports Reuters.

The outlet said that ByteDance is diversifying its domestic suppliers of chips used for AI and accelerated development of its own since the United States started restricting exports on advanced AI chips -- you know, the world's best are made by NVIDIA and export controls have stopped all high-end NVIDIA chips from entering China. So, ByteDance is side-stepping that, and keeping it all within China by tapping Huawei AI chips.

ByteDance will be tapping Huawei's new Ascend 910B AI chips to train its new large language model (LLM) according to Reuters' sources, who of course, declined to be identified as the plan is confidential (except leaking it to one of the largest news sources on the planet seems to be OK, even for Chinese companies).

The AI chip market, led by NVIDIA, is crucial to the push toward generative AI and the age of AI PCs and AI-powered smart devices. However, with US trade restrictions on AI hardware, China and its tech giants are finding it increasingly challenging to source NVIDIA chips like the H100. This includes companies like ByteDance, Alibaba, and Baidu.

With NVIDIA no longer able to ship its flagship gaming GPU, the GeForce RTX 4090, to the region, there's room for a company like Huawei to step up and develop new AI hardware specifically for the Chinese market. The tech giant has hit the ground running, and its latest AI chip - the Ascend 910C - is currently being tested.

According to a new South China Morning Post report, Huawei is seeding its new Ascend 910C chip to local AI giants currently using NVIDIA's H100 AI GPU. The Ascend 910C is apparently comparable to the H100 regarding raw performance, so it's positioned as an alternative chip for China's AI market.

South Korea is pushing all-in with AI, with South Korean companies pledging to invest a combined 65 trillion won (around $48.9 billion USD) into AI development by 2027 in order to better compete with the global leaders in AI, the presidential office said on Thursday.

The investment plan was announced during an inaugural meeting of the presidential committee on AI, which was launched to create an AI strategy and coordinate AI research and development efforts, said Park Sang-wook, the presidential secretary for science and technology.

In the meeting, he called for collaborate efforts between the South Korean government and private sector for the country to become one of the top three global leaders in AI by 2027. He said: "It is a time when AI determines national capabilities and economic growth, becoming a key factor in the economy and security. The world's major nations are staking everything on securing AI technology and leadership to win the digital supremacy competition".

With powerful GPUs built to run complex AI workloads and new mobile (and desktop) CPUs integrating NPUs into their architecture for generative AI tasks, the PC as we know it is changing. And that change could be here sooner than we think, with research and consulting firm Gartner forecasting that AI PCs will account for 43% of global PC shipments by 2025.

"The debate has moved from speculating which PCs might include AI functionality to the expectation that most PCs will eventually integrate AI NPU capabilities," said Ranjit Atwal, Senior Director Analyst at Gartner. "As a result, NPU will become a standard feature for PC vendors."

This will present a dramatic shift for all consumers buying laptops or PCs, including students, creators, and professionals. Not everyone will jump in and start using Windows' AI features with their new NPU-powered CPU, but all new CPUs will include an NPU and AI hardware as standard.

NVIDIA's issues with its Blackwell AI GPUs have been "addressed" according to tech journalist Tim Culpan on X, with actual GB200 AI server shipments going out to major cloud service providers (CSPs) "earlier than previously-feared delay".

Microsoft is expected to take delivery of one of the largest allocations of NVIDIA's new Blackwell AI GPUs, as well Oracle, AWS, Meta, and others. NVIDIA's new GB200 NVL36 AI servers were originally slated for delivery at the end of this month, while the higher-end (and more expensive) GB200 NVL72 AI servers were scheduled for early November.

Culpan took to his Substack to explain that delays pushed the timeline back to January, but now both NVL36 and NVL72 AI servers will "ship around the first week of December".

SK hynix has announced volume production of its new 12-layer HBM3E memory, with up to 36GB capacities and speeds of 9.6Gbps.

The South Korean memory leader announced it has started mass production of the world's first 12-layer HBM3E memory with 36GB, the largest capacity of existing HBM to date. SK hynix plans to supply mass-produced 12-layer HBM3E memory chips to companies (NVIDIA) within the next 12 months, and only 6 months after launching 8-layer HBM3E to customers for the first time in the industry in March 2024.

SK hynix is the key to the world of AI chips, with NVIDIA using its HBM3 and HBM3E memory inside of its Hopper H100 and H200 AI GPUs, with HBM3E also used in its new Blackwell AI GPUs. SK hynix has been leading the industry with HBM, with its new 12-layer HBM3E memory chips boosted up to 9.6Gbps of bandwidth, the highest memory speed on the market.

Stability AI has just announced that legendary filmmaker, technology innovator, and visual effects pioneer James Cameron has joined its Board of Directors.

Stability AI is the team behind the infamous Stable Diffusion AI model, with Cameron stepping up into the board of Stability AI as a driving force in cutting-edge technology with visionary storytelling, said the AI startup in its press release.

Cameron joining the Stability AI team represents a "represents a significant step forward in Stability AI's mission to transform visual media. Both Cameron and Stability AI operate at the intersection of emerging technology and creativity. Cameron's artist-centric perspective, paired with his business and technical acumen, will support Stability AI in continuing to unlock new opportunities to empower creators to tell stories in ways once unimaginable".

Intel has officially launched its Gaudi 3 AI accelerator, with the new AI chip coming in slower than NVIDIA's dominant H100 and its new HBM3E-fueled H200 AI GPUs, meaning Intel is aiming its Gaudi 3 by pushing that it's cheaper, and has a lower total cost of ownership (TCO).

Inside, the new Intel Gaudi 3 AI accelerator features two chiplets with 64 tensor processor cores (TPCs, 256x256 MAC structure with FP32 accumulators), eight matrix multiplication engines (MMEs, 256-bit wide vector processor), and 96MB of on-die SRAM cache with a 19.2TB/s bandwidth.

Gaudi 3 also features 24 x 200GbE networking interfaces and 14 media engines, with the media engines capable of handling H.265, H.264, and VP9 to support vision processing. Intel's new Gaudi 3 AI accelerator features 128GB of HBM2E memory with up to 3.67TB/sec of memory bandwidth.