Artificial Intelligence - Page 9

NVIDIA CEO Jensen Huang has flown to Taiwan to meet with key TSMC executives, while also expressing deep gratitude to the US government for approving export licenses for its H20 AI GPUs to China.

Jensen noted that while China has recently raised concerns about potential security backdoors inside of NVIDIA chips, there are no such vulnerabilities in H20. The NVIDIA CEO added that the company has provided enough explanations to address any concerns that Beijing may have.

During his conversation with President Trump, Jensen underlined the importance of AI, linking it to US chip tariffs, stressing that AI will advance regardless of US participation, but supporting the American technology ecosystem is extremely important. Jensen described TSMC as "one of the greatest companies in human history" and called the Taiwanese contract semiconductor manufacturer a powerful investment destination.

Apple is long overdue to release its next-generation version of Siri, which the company used as marketing to sell its latest generation of iPhone. Unfortunately for consumers and Apple, that release never happened, and instead, consumers got only a few AI-powered upgrades in the form of image editors, custom emoji makers, and several other smaller quality-of-life features.

The big selling points of the latest generation of iPhone were an AI-powered upgrade was for Siri, which would enable deep Siri integration into all of Apple's homebrewed applications and settings. Essentially, Siri would become smart.

However, due to Siri's struggles to meet Apple's extremely high standard for public release, the assistant's level-up has been delayed indefinitely. We are now hearing that the assistant may be powered by a third-party vendor, such as OpenAI, Anthropic, or even Google.

It may sound unlikely, but it's possible that Microsoft could bring functionality to Edge so that users can let Copilot take control of their browser - a feature already available more broadly in the enterprise world.

Leopeva64, a regular leaker of browser-related bits and pieces, discovered a new flag in the Canary version of Edge which allows for enabling a WebUI overlay for a banner that reads: 'Copilot is in control right now'.

This appears to be the very first hint in test builds of Edge that the browser could get a feature whereby Copilot interacts with the web page currently displayed via simulated clicks or key presses.

China is reportedly pushing for the ban of foreign chips made by any company -- namely NVIDIA -- for use in inferencing, which accounts for the most AI demand, according to new reports.

In an article from the well-trusted outlet the Financial Times, it's being reported that according to someone who was "recently summoned" to meet with Beijing policymakers in China are pushing for a ban on foreign-made chips used for AI inferencing, but it won't happen any time soon.

There is a major shortage of domestic AI chip supplies in China, but Beijing hopes to "significantly improve" that in 2026 with multiple advanced production lines coming online. China's foreign ministry said: "As a matter of principle, science, technology, and economic and trade issues should not be politicized, instrumentalized, or weaponized. Containment and suppression will not hold back China's development".

Samsung has reportedly offered its HBM3E memory for NVIDIA's tweaked H20 AI GPU for China, at prices 20-30% lower than South Korean memory rival SK hynix according to new reports.

In a new report from South Korean outlet SEdaily picked up by @Jukanrosleve on X, it's reported that: "at the end of this month, Samsung's 12-layer HBM3E product is also expected to pass NVIDIA's quality test and begin delivery. Industry observers believe that ongoing deadlock in supply and price negotiations between NVIDIA and SK hynix is due to Samsung's imminent delivery".

The article continues: "recently, NVIDIA received approval to export its low-spec H20, developed for China, under the condition of paying 15% of sales to the US government. Samsung is reported to have proposed an HBM3E supply price for the H20 that is 20-30% lower than SK hynix. Accordingly, NVIDIA is insisting on confirming the quality of Samsung's HBM3E and HBM4 before finalizing price agreements with SK hynix".

Samsung Electronics has reportedly had its next-gen HBM4 memory samples have passed NVIDIA's qualification tests to be used inside of its AI GPUs.

In a new report from SEdaily picked up by insider @Jukanrosleve on X, we're hearing that the semiconductor industry said that Samsung's new HBM4 memory samples provided to NVIDIA in their initial prototype and quality tests, and will now enter the "Pre-Production (PP)" stage at the end of this month (August 2025).

A key industry insider said: "It received positive evaluations regarding quality, including yield, and has entered the PP stage for mass-production testing", adding "If it passes PP, mass production will be possible in November-December". This is the final stage before large-scale semiconductor production, with the HBM4 prototype that Samsung provided last month being an "engineering sample" for simple operational tests.

NVIDIA is reportedly cooking up the Blackwell B30A chip that will replace Hopper H20 in China, where the new B30A will beat the H20, but it will offer just half the performance of the Blackwell Ultra B300 AI GPU.

In a new report from Reuters, NVIDIA has reportedly finalized the new B30A chip design for China that is compliant with US regulations, with B30A being faster than H20, but much slower than higher-end AI chip offerings like B200 and the upcoming B300.

NVIDIA's purported B30A will feature a single-chip design, compared to the dual chiplets found on B200 and B300. NVIDIA's upcoming Blackwell Ultra B300 features 50% more performance than B200, with 50% more HBM3E memory, and twice the interconnect speeds. The highest-end B300 AI GPU will feature two reticle-sized GPUs with a total of 15 PetaFLOPS of FP4 compute performance, and up to 288GB of HBM3E memory.

AI is quickly integrating into the day-to-day productivity of many people, from programmers to researchers, to students, engineers, and more. According to a new study from Google (here's the PDF link for the findings), the games industry is "currently in the midst of profound changes," and 97% of game developers agree that generative AI is already reshaping the industry. However, there's a big difference between asking GPT to summarize a document versus using AI Agents as dynamic NPCs that react to play actions in an open-world RPG.

One of the biggest surprises from Google's study, which engaged with 615 game developers across the United States, South Korea, Norway, Finland, and Sweden, is that 87% of game developers are already using AI Agents in their work. AI agents can be viewed as autonomous little digital helpers that can reason and plan, with memory.

In addition to AI Agents being NPCs in a game, they can also be used to optimize difficulty settings, act as virtual game testers, work together as a group of enemies to execute complex battle strategies, and react to players.

Samsung Electronics is coming out swinging in the semiconductor business in 2026, with its next-gen 1c DRAM ready to rock and roll in the HBM4 business.

Samsung Electronics Chairman Lee Jae-yong recently returned from a visit to big tech companies through San Francisco and Silicon Valley over the last two weeks, with the chairman saying: "we have prepared for next year's business". Analysts have interpreted that as the company chief personally stepping in and reinforcing its customer base for a full revival, as its semiconductor division in both memory and foundry, both suffering against rivals like SK hynix and TSMC.

We can expect some big changes in Samsung's major customers between the first half of 2025, and the first half of 2026, as it begins mass supplying its next-gen HBM4 memory to customers like NVIDIA and AMD. Samsung has been slowly but surely completing its next-generation technologies one at a time, after a few years of rather big struggles and missteps, but it's preparing for a huge year in 2026.

NVIDIA has announced its first major update to Project G-Assist, the experimental locally run AI assistant that enables RTX GPU owners to tune and alter their PC through voice and text commands.

For those who don't know, G-Assist can be accessed through the NVIDIA app, and on August 19, the latest update for the AI will become available. RTX owners will be able to download the feature and press Alt+G to access the chat window, where they will then be able to instruct G-Assist to carry out certain functions through text or voice commands. For example, G-Assist can run diagnostics on the PC to improve in-game performance. Display charts, frame rates, latency, and GPU temperatures.

NVIDIA intends to expand G-Assist's capabilities and has partnered with mod.io to launch the G-Assist Plug-In Hub, which enables users to easily access G-Assist plug-ins that were created by the community. An example of a community-created plug-in would be OpenRGB, a plug-in designed to enable a user to control all of the RGB lights on their PC.

DeepSeek rattled the AI industry when it unveiled the R1 model earlier this year, proving that a sophisticated AI model doesn't need an egregious amount of GPU horsepower to compete with already existing AI models.

But where is the company's next-generation AI model, R2? According to a report from the Financial Times, the upcoming AI model was blundered when the company was pressured by the Chinese government to use Huawei chips to develop it. Three unnamed sources spoke to the publication and said the upcoming model is struggling in development due to unstable Huawei chips, immature software, and other factors. Notably, these unnamed individuals said the new model wasn't even able to complete a single training run.

DeepSeek R2 was initially planned for release in May 2025, but was postponed with no reason provided. It was noted that the R2 model's performance was falling below the standards set by the company, specifically DeepSeek's CEO Liang Wenfeng. There were also trade restrictions on NVIDIA's AI GPUs, which are now coming down as a new deal has been struck between AI GPU makers and the US government.

There's a new fastest AI when it comes to Pokémon Red speed runs and it's GPT-5, beating out the previous record by a long way.

OpenAI's most recent model took just 6,470 steps to complete Pokémon Red, compared to ChatGPT-o3 which beat the game in 18,184 steps.

As TechRadar, which spotted this feat, points out, 6,500 steps translates to about a full week of gaming (non-stop) - well, just under a week - so it's not exactly fast. Well, not compared to a human speed runner (apparently something like 2,000 steps is possible, and the record time is for completion is 1 hour and 44 minutes).

Apple is rumored to be working on an AI tabletop robot for a 2027 release, which would revolutionize personal interactions with AI. The device sports a Pixar Lamp (Lumo) style design, with a 7-inch iPad-like display mounted on a moveable arm designed to act like a super-interactive version of Siri.

In a new article from Bloomberg, it's being reported that Apple is calling its device the "Pixar Lamp" internally, envisioned to act like a person being in the room with users, capable of back-and-forth conversations and discussions like planning trips, or where to have dinner that night.

The interactivity with the Siri-powered "Pixar Lamp" is said to be similar to how OpenAI's voice mode works, enhancing the user experience with dynamic and context-aware responses. We know that Apple is working on a next-generation version of Siri specifically for this robot, enabling the software to remember information and have more natural conversations.

NVIDIA's next-gen Rubin AI GPU architecture hasn't been delayed, after rumors flew in the last few hours that Rubin was being delayed to better fight AMD's next-gen Instinct MI450 AI accelerators.

But in a response from an NVIDIA spokesperson who said recent reports of Rubin being delayed are "incorrect" and that "Rubin is on track". This is after reports surfaced claiming NVIDIA has delayed Rubin, redesigning the chip from 1800W power to 2000W of power to better compete with AMD's next-gen Instinct MI450 AI accelerator.

The earlier report has been debunked, with official comment from NVIDIA reaffirming Rubin is on track, which last we heard there will be 5.7 million Rubin AI GPUs shipped in 2026, each with next-generation HBM4 memory and up to 1800W of power per R100 AI chip.

NVIDIA has introduced new members of the popular RTX PRO Blackwell workstation GPU family of cards, with new SFF-compatible variants now launched.

The company has launched its new RTX PRO 4000 SFF and RTX PRO 2000 workstation GPUs at SIGGRAPH 2025 this week, offering top-tier performance that slots into a SFF system. The new RTX PRO 4000 SFF replaces the RTX A4000 SFF with up to 2.5x more AI performance, 1.7x more ray tracing performance, and 1.5x more bandwidth, all within the same 70W power envelope.

NVIDIA's new RTX PRO 4000 SFF packs 24GB of GDDR7 memory with ECC, 4 x mini DP 2.1 display connections, all in a compact form factor under SFF certification, meaning the workstation GPU can be installed and use inside of smaller machines.

US authorities have been secretly placing location tracking devices inside of shipments of advanced AI chips and AI servers that are at high risk of illegal diversion into China.

In a new report from Reuters, it's being reported that the new measures have been enacted to detect AI chips being diverted through countries that are under US export restrictions, and apply only to particular shipments that are under investigation. This move shows how far the US is willing to go to enforce its chip export restrictions to China, and that's even with the Trump administration relaxing some of the US export restrictions to give China access to some advanced US-designed semiconductors.

The trackers are a decades-old investigative tool used by US law enforcement agencies, so that they can track products that are under export restrictions, including airplane parts. Trackers have been used to combat the illegal diversion of semiconductors in the past years, according to one of Reuters' sources.

Samsung Electronics has confirmed its latest HBM3E 12-Hi memory has been picked up by NVIDIA, with the company to supply 30K to 50K units of the AI memory to NVIDIA in the "near future".

In a new report from Korean media outlet Alpha Economy picked up by @Jukanrosleve on X, NVIDIA and Samsung Electronics reportedly agreed on the supply of HBM3 12-Hi memory. It's being reported that all of the Samsung HBM3E 12-Hi memory supplied will be used exclusively in liquid-cooled AI servers.

Samsung said in regards to its HBM3E 12-Hi memory supply being used exclusively in AI servers: "we cannot confirm this".

SK hynix is reportedly using the EUV process in the development of its next-gen DRAM, with its 1c DRAM (6th-generation 10nm-class) for which transitional investment will begin later this year according to new reports.

In a new report from ZDnet Korea, the outlet reports from industry sources that SK hynix plans to use 5 or more EUV layers on its new 1c DRAM. SK hynix first applied EUV to one layer of its 1c DRAM (4th-generation 10nm-class) which was later expanded to four layers for its 1b DRAM.

In the future, 1c DRAM will increase the number of EUV layers to five or more, with SK hynix the first to develop 16Gb (gigabit) DDR5 DRAM based on the new 1c DRAM process, with plans to convert investments into 1c DRAM in the second half of this year.

NVIDIA has just opened up pre-orders for its new Jetson AGX Thor Developer Kit, a new mini AI supercomputer packing industry-leading performance ready for humanoid robots, for $3499. Check it out:

Inside, the new NVIDIA Jetson AGX Thor Dev Kit features an NVIDIA Jetson T5000 with Blackwell GPU, featuring 2070 FP4 TFLOPS of AI compute power, with a 14-core Arm Neoverse-V3AE 64-bit CPU, and 128GB of LPDDR5 memory on a 256-bit memory bus with up to 273GB/sec of memory bandwidth.

The overall design resembles NVIDIA's in-house GeForce RTX series Founders Edition graphics cards, with a heatsink and fan keeping the Jetson T5000 with Blackwell GPU cool, a 1TB NVMe SSD + Wi-Fi 6e module, and rich I/O that includes an HDMI 2.1 connector, DisplayPort connector, super-fast 5GbE ethernet port, 2 x USB-C and 2 x USB-A ports.

SK hynix president says the HBM memory market will grow at 30% per year through to 2030, with leading AI GPUs from NVIDIA and AMD using the best HBM they can get, and the AI market isn't slowing down any time soon.

The upbeat projection for global growth of HBM used in AI chips brushes off any concerns in rising price pressures in a sector that, for decades, has been treated like commodities such as oil or coal, reports Reuters. US cloud companies like Amazon, Microsoft, and Google are spending hundreds of billions of dollars on AI servers, which is very healthy for one of the key ingredients: HBM.

SK hynix's Choi Joon-yong, the head of HBM business planning at SK hynix, said: "AI demand from the end user is pretty much, very firm and strong".