Artificial Intelligence - Page 3

It was only yesterday that we got our first look at the PCB of NVIDIA's new RTX PRO 6000 Max-Q Blackwell GPU (limited to 300W of power, but still rocks 96GB GDDR7) but now we have the PCB from the RTX PRO 6000 Blackwell GPU with its max 600W TDP.

The PCB design of NVIDIA's new RTX PRO 6000 Blackwell GPU was posted on Chiphell, and since the new workstation cards haven't been launched yet, these are some really early shots of the PCB. NVIDIA's new RTX PRO 6000 Blackwell features a compact PCB that is split into 3 parts: the motherboard for the GPU and GDDR7 memory, and the PCIe interface board, with the only missing part here being the display connector board.

NVIDIA has placed the GDDR7 memory on both sides of the PCB, with each side featuring 48GB GDDR7 for a full 96GB using 3GB GDDR7 memory modules.

NVIDIA is denying reports that it would be spinning off its Chinese operations so it acted as a standalone business, which would side-step strengthening US export controls.

There have been rumors and reports of this for the last few days, but NVIDIA has officially said: "There is no basis whatsoever for any of these claims. It is irresponsible to publish baseless claims and speculation as fact".

The original post was on X by leaker @Jukanlosreve and later reported by the likes of DigiTimes, that NVIDIA CEO Jensen Huang was considering establishing a joint venture in mainland China to maintain the operations of the company's leading CUDA computing platform, as well as other business interests in the company.

We haven't stopped hearing about AI for a couple of years now, but recently Microsoft CEO Satya Nadella admitted that as much as 30% of the company's new code is written by AI.

Microsoft CEO Satya Nadella revealed the news during the recent LlamaCon, which is Meta's conference focusing on generative AI tools, where when sitting across from Facebook founder and CEO Mark Zuckerberg when he said: "code reviews are very high. In fact the agents we have for reviewing code, that usage has increased, and so I would say maybe 20, 30 percent of the code that is inside of our repos today and in some of our projects are probably all written by software".

Nadella also asked Zuckerberg how much of Meta's code was being written by AI, to which Zuckerberg said he didn't know the exact figure off the top of his head, but noted that Meta is building an AI model that can in-turn build future versions of Meta's in-house Llama family of AI models.

Huawei has started delivering its new CloudMatrix 384 AI clusters to customers in China, powered by its Ascend 910C AI chips.

In a new report from the Financial Times, we're learning that 10 different clients have now adopted Huawei's new CloudMatrix 384 AI servers into their data center portfolios. We don't know which Chinese companies are using Huawei's new AI servers, but they are reportedly primary customers of Huawei's product offerings.

Huawei's new CloudMatrix 384 "CM384" AI cluster is powered by 384 Huawei Ascend 910C AI chips connected in an "all-to-all topology" configuration. Huawei is outweighing the architectural flaws of its AI chips by using 5x more of them than NVIDIA uses with GB200 inside of NVL72 servers. This is why the company doesn't care about the costs, performance inefficiency, scalability ratios, and more.

When Microsoft unveiled its new AI-powered range of Copilot+ PCs, one feature that drew immediate criticism was a new AI tool for Windows called Recall. Of course, Windows and criticism go hand in hand, but when it came to Recall - a tool that took screenshots of your PC that could then be used to navigate through your PC's usage history - it was a little different.

Early versions of the technology captured sensitive information like banking details and passwords and then put everything in an indexed database that could be searched, raising immediate security and privacy concerns. That's the old Recall; the new and improved version has been rebuilt with security in mind and is now available for all Copilot+ PC users.

"Recall is an opt-in experience with a rich set of privacy controls to filter content and customize what gets saved for you to find later," Navjot Virk, Microsoft Corporate Vice President, Windows Experiences, writes. "We've implemented extensive security considerations, such as Windows Hello sign-in, data encryption, and isolation in Recall to help keep your data safe and secure."

NVIDIA's beefed-up B300 AI chip production has been reportedly pulled forward to May, and will be fabbed on TSMC's new 5nm (N4P) process node and will use CoWoS-L advanced semiconductor packaging.

In a new post by Ctee, we're hearing that NVIDIA's new B300 AI chips will use Bianca compute boards with 1 x CPU and 2 x GPUs, with B300 to enter mass production before the end of the year. Analysts estimate that NVIDIA's new B300 AI GPU will boost the related supply chains including TSMC, Machtech, Inventec, Chipset, and assemly plants Quanta, Wistron, and Foxconn.

NVIDIA's new B300 using TSMC's new 5nm family (N4P) process node was used to fill the production capacity gap left wide open by the now-banned H20 AI GPU in China. Ctee reports that the supply chain noted a shipment of AP8 from Nanya Advanced Packaging in early April, in order to follow the CoWoS-L advanced packaging to be used for B300.

Huawei is working on its next-gen Ascend 910D AI GPU, which is said to offer AI performance matching NVIDIA's previous-gen Hopper H100 AI GPU for the Chinese market.

In a new report from Bloomberg, we're learning that Huawei's new Ascend 910D AI chip will have its first batch of samples out in late May, with development "still at an early stage". Huawei hopes that its new Ascend 910D will be more powerful than NVIDIA's previous-gen Hopper H100, which was released in 2022, and has been succeeded by Blackwell B200, and soon B300 and B300 Ultra AI GPUs.

NVIDIA's custom made-for-China H20 AI GPU was recently banned from being sold in China with new US export restrictions, with the company taking a $5.5 billion hit on its Q1 2025 revenue because of it. China can't get its hands on even lower-end AI GPUs, so it is having to rely on homegrown solutions, with Huawei and its new Ascend 910D.

SK hynix showed off its next-gen HBM4 memory at TSMC's recent North American Technology Symposium, with up to 16-Hi stacks and 2TB/sec memory bandwidth per stack, ready for NVIDIA's next-gen Vera Rubin AI hardware.

SK hynix showed off both 12-Hi and 16-Hi stacks of HBM4 memory, which feature a capacity of up to 48GB, up to 2TB/sec memory bandwidth, and I/O speeds rated at 8Gbps, with the South Korean memory leader announcing mass production for 2H 2025, and into AI GPUs by the end of this year, flooding the market with HBM4-powered AI GPUs like NVIDIA's next-gen Vera Rubin in 2026.

We will see SK hynix's world-leading HBM4 memory chips inside of NVIDIA's upcoming GB300 "Blackwell Ultra" AI GPUs, with the company planning to shift fully into the arms of HBM4 memory starting with Vera Rubin later this year. SK hynix also pointed out that they've managed the high number of layers through using Advanced MR-MUF and TSV technologies.

Chinese AI firm DeepSeek is cooking up its next-gen R2 AI model, which is said to be 97% cheaper to train than GPT-4, and it has been fully trained on Huawei AI GPUs.

In a new post on X by @deedydas has the hype train for DeepSeek R2 rocking and rolling, claiming that the new R2 model is going to adopt a hybrid MoE (Mixture of Experts) architecture, which is meant to be an advanced version of the existing MoE implementation, which should provide advanced gating mechanisms, or a combination of MoE + dense layers to optimize high-end AAI workloads.

DeepSeek R2 is set to double the parameters of R1, with 1.2 trillion parameters at the ready, and it's reportedly a whopping 97.3% cheaper to train than GPT 4o with the unit cost per token lower than 97.3% compared to GPT-4 at $0.07/M input token and 0.27/M output token. This means DeepSeek R2 is going to be uber-cheap for enterprise use, as it'll be the most cost-efficient AI model on the market.

In what comes as absolutely no surprise, Intel has said that its ai PC processors aren't selling anywhere near the numbers they were expecting, creating a shortage of production capacity for older CPUs.

Intel says that customers are buying less expensive, previous-generation Raptor Lake CPUs instead of the new AI PC-ready Meteor Lake and Lunar Lake processors inside of new laptops. During its recent earnings call, Intel said that it is currently facing production capacity issues for its in-house Intel 7 process node, and that it expects this shorage to "persist for the forseeable future".

Intel's current-generation processors are fabbed using newer process nodes from TSMC, and not using the older-gen Intel 7 process node, so the unexpected surge in demand in the lower-end, non-AI PC-based processors (Raptor Lake) is a strange problem for Intel to have, considering the insane marketing push for AI PC over the last 12+ months.