Artificial Intelligence - Page 17

The VRAM capacity debate is currently being discussed in the PC gaming space. The consensus is that in 2025, you will need more than 8GB of VRAM for high-end 1440p and 4K gaming. VRAM is also a key and crucial component for running local AI, and the demand for more memory is growing alongside the arrival of more complex models.

The powerful Stable Diffusion 3.5 large language model for creating images from text models uses 18GB of VRAM. This limits the use of the GeForce RTX 50 Series to the flagship GeForce RTX 5090. Well, not anymore, as NVIDIA has collaborated with Stability AI to quantify the model to FP8, reducing the VRAM requirement by 40% to 11GB.

Alongside optimizations with TensorRT to double performance, this now means that five GeForce RTX 50 Series GPUs (the RTX 5060 Ti 16GB, RTX 5070, RTX 5070 Ti, RTX 5080, and RTX 5090) can run the model locally.

AMD has confirmed that its newly-announced Instinct MI350 series AI accelerators are using Samsung's latest HBM3E 12-Hi memory.

Samsung Electronics has been almost famously tripping over trying to get HBM certification from NVIDIA for use in its AI GPUs, but now AMD's just-announced Instinct MI350 and MI355X AI accelerators are using Samsung and Micron's new 12-Hi HBM3E memory. Samsung has been supplying HBM to AMD for a while now, but this is the first time that AMD has confirmed it.

AMD's new Instinct MI350 series AI accelerators boast 185 billion transistors and up to 288GB of HBM3E memory, but the company also teased its next-gen Instinct MI400 series AI chips that will feature up to 432GB of next-gen HBM4 memory. AMD said that its new Helios AI server racks will feature 72 x Instinct MI400 series GPUs with 31TB of HBM4 per rack, featuring 10x the AI computing power over its newly-announced Instinct MI355X-based server rack.

AMD launched its new Instinct MI350 series AI accelerators today, rocking 185 billion transistors, up to 288GB of HBM3E memory, FB4 and FP6 support, and more.

The new Instinct MI350 series AI chips were launched during AMD's huge Advancing AI event, where it also unveiled its new Zen 6-based EPYC "Venice" CPUs, a tease of its next-next-gen Zen 7-based EPYC "Verano" CPUs, as well as a tease of its new Instinct MI400 and even next-next-gen Instinct MI500 series AI accelerators.

AMD's new Instinct MI350 series AI accelerators are based on the company's new CDNA 4 architecture, and fabbed on TSMC's 3nm process node. Inside, the chip itself features 185 billion transistors and comes in two different variants: the MI350X and MI355X, offered in both air-cooled, and liquid-cooled configurations.

AMD has just teased its next-gen Instinct MI400 AI accelerator, which will double the AI compute performance over the just-announced MI350 series, with 50% more memory, and close to 2.5x the memory bandwidth thanks to the use of next-gen HBM4 memory.

The company shared some fresh new details on its next-gen Instinct MI400 series AI accelerator, with 40 PFLOPs (FP4) and 20 PFLOPs (FP8) which is double the AI compute speeds of the new Instinct MI350 that was just launched today. AMD's new Instinct MI400 series AI chip will boast 50% more memory capacity over the MI350 which has 288GB of HBM3E, while the new MI400 has a huge 432GB of HBM4 memory.

AMD's use of the new HBM4 standard will bring the company up to full competitiveness against NVIDIA which will be using HBM4 on its upcoming Rubin R100 AI GPU, with AMD's new Instinct MI400 AI chip and its 432GB of HBM4 offering a huge 19.6TB/sec of memory bandwidth, up from just 8TB/sec on the MI350 series. The new AI GPU will also sport a 300GB/sec scale-out bandwidth per GPU, so we should expect big things from MI400 in 2026 as it battles Rubin R100 in the HBM4-powered AI fight.

A lawsuit filed in a California district court by Disney and NBCUniversal has claimed that the popular generative artificial intelligence program and service, Midjourney, has claimed the company behind the tool has continuously violated both companies' intellectual rights>

The lawsuit accuses Midjourney of ignoring its previous requests to stop violating its intellectual property rights, which Disney and NBCUniversal have traced back to its generative AI tools, such as image generation, More specifically, the lawsuit provided an example of how it believes Midjourney is violating intellectual property rights, such as the AI tool enabling users to generate an image of Disney-owned "Star Wars" character Darth Vadar in a variety of different settings and performing particular actions. The AI "obliges by generating and displaying a high-quality, downloadable image."

Disney and NBCUniversal are framing the lawsuit against Midjourney as a stance on protecting the "hard work of all the artists whose work entertains and inspires us," with Disney's chief legal compliance officer, Horacio Gutierrez saying in a statement, "Our world-class IP is built on decades of financial investment, creativity and innovation-investments only made possible by the incentives embodied in copyright law that give creators the exclusive right to profit from their works."

It's no secret that the advanced and powerful AI models used for image generation have been trained using copyrighted material. Case in point: the images in this article were created using Midjourney's AI image generator and simple prompts like "Mickey Mouse and Darth Vader in a courtroom."

And with news breaking that Disney and Universal are now suing Midjourney for copyright infringement, it's not all that surprising when you use the powerful AI tool to create an image of any Disney or Universal character with impressive accuracy in seconds.

The suit, filed in Los Angeles this week, claims that Midjourney has scraped the massive content libraries of Disney and Universal to train its AI image generation model. The suit doesn't hold back, either, calling Midjourney a "copyright free-rider" and a "bottomless pit of plagiarism."

AMD is gearing up to unleash its new CDNA 4 architecture inside of its new Instinct MI350 series AI accelerators, with its new flagship Instinct MI355X featuring 288GB of HBM3E memory and 1400W of peak board power.

AMD CTO Mark Papermaster unveiled the company's new Instinct MI350 series and new MI355X for AI and HPC at the ISC 2025 event recently, with 1400W of peak board power being close to double what the company's previous-gen Instinct AI accelerator consumed, according to new reports from ComputerBase.

We can expect the full unveiling of AMD's next-gen CDNA 4-based Instinct MI350 series AI accelerators during a livestream later this week on Thursday, with the new Instinct MI350 series AI chips based on the optimized CDNA 4 architecture. CDNA 4 is effective in supporting formats, something that was previously a downside for AMD.

PCI-SIG has announced the official release of its new PCIe 7.0 specifications, offering up to a blistering 128GT/s transfer speeds, while teasing that pathfinding next-gen PCIe 8.0 specifications are "in progress".

In a new press release issued today, PCI-SIG's new PCIe 7.0 specifications see a doubling of bandwidth of the not-even-out-yet PCIe 6.0 specification (64GT/s) with PCIe 7.0 bursting at up to 128GT/s raw bit rate, and up to 512GB/sec bi-directionally through a PCIe 7.0 x16 connection.

PCIe 7.0 will be using PAM4 (Pulse Amplitude Modulation with 4 levels) signaling and Flit Mode, with improved power efficiency, backwards compatibility with previous generations of PCIe standards, and supports AI/ML (artificial intelligence and machine learning), HPC (high-performance computing), cloud, data centers, and more.

Applied Digital Corporation is a designer, builder, and operator of next-gen digital infrastructure for HPC applications, announcing new 15-year lease agreements with AI Hyperscaler, CoreWeave.

This new agreement will see Applied Digital delivering 250 megawatts (MW) of critical IT load to hose CoreWeave's extensive AI and high-performance computing (HPC) infrastructure at its Ellendale, North Dakota data center campus. Across the expansive approximately 15-year term, Applied Digital anticipates generating around $7 billion in total revenue from the leases.

CoreWeave also retains the option to access an additional 150 MW of critical IT load at Ellendale, which has positioned the data center campus as a scalable hub for expanding AI and HPC workloads.

White House AI czar David Sacks has said that China isn't years behind the US when it comes to AI, but more like 3-6 months behind, and that the AI race is "very close".

Sacks was speaking at the AWS Summit in Washington when the White House AI and crypto czar said: "China is not years and years behind us in AI. Maybe they're three to six months. It's a very close race". He also downplayed the issues of American AI chips being smuggled by bad actors, and raised concerns that regulating US AI too tightly could hurt growth, and cede the critical AI market to China.

He said: "We talk about these chips like they could be smuggled in the back of a briefcase. That's not what they look like. These are server racks that are eight feet tall and weigh two tons. They don't walk out doors. It's very easy to basically verify that they're where they're supposed to be".

AMD will be launching its new Instinct MI350 series AI accelerators this week during the ISC25 keynote, where AMD CTO Mark Papermaster will unveil the company's next-gen Instinct MI350 series AI chips offering up to 35x improvements in AI inferencing capabilities, and more.

We know that the AMD Instinct MI350X AI accelerator will boast up to 288GB of HBM3E memory, and inside, be powered by AMD's next-gen "CDNA 4" architecture. The new Instinct MI350 series AI accelerators will better compete with NVIDIA's new Blackwell AI GPUs, bridging the gap of AI performance between AMD and NVIDIA later this year.

AMD's new Instinct MI350 series AI accelerators should find a position in the market between NVIDIA's now-on-the-market Blackwell B200 AI GPUs, as well as its beefed-up Blackwell Ultra B300 AI GPUs.

Being able to "beat the computer" in chess has been something PC, gaming, and tech enthusiasts have discussed for decades. In 1997, IBM's Deep Blue supercomputer made headlines worldwide for defeating Garry Kasparov in a chess game. Compared to the PC hardware we have in 2025, Deep Blue is now what you'd consider retro regarding its capabilities, even though it could brute force its way to evaluate 200 million chess moves per second.

Compared to Deep Blue, the Atari 2600 video game console is effectively from the Stone Age. It debuted in 1977, predating most modern gamers. With the rise of AI and chatbots like ChatGPT, you might think that it would have no problem beating Deep Blue in a round of chess, let alone going up against an Atari 2600 to play a game using the Atari Chess 'beginner' opponent difficulty setting.

You'd be wrong, as Citrix Architecture and Delivery specialist Robert Jr. Caruso decided to carry out this little experiment and posted his findings on LinkedIn.

NVIDIA CEO Jensen Huang is jet-setting once again, after his trips to Taiwan and the Middle East recently, where Jensen will be flying to Germany to announce the country's biggest AI factory this week.

NVIDIA has its eyes on the EU to expand its large-scale AI factory in Germany, where Jensen will meet with German Chancellor Friedrich Merz in Berlin to discuss the deal. In new reports, they suggest that the EU wants to catch up in the ever-evolving AI race with the rest of the world but there is a major hurdle: it has no other option but NVIDIA for advanced AI chips.

Jensen is expected to unveil a new mega datacenter in Germany, using 100,000 of its high-end AI GPUs that come at a cost of around $30,000 per chip, which will see the project costing a total of around $3 billion for AI chips alone. The new deal would mean this is Germany's largest AI venture, and for NVIDIA it is another notch in its ever-growing AI dominance worldwide.

Micron has announced that it has started shipping its new HBM4 36GB 12-Hi memory to "multiple key customers" with the first to likely be NVIDIA and its next-gen Rubin R100 AI GPU.

In a press release, the US-based memory maker said it's now shipping HBM4 36GB 12-Hi memory to select customers, built on its well-established 1b DRAM process node, 12-high advanced packaging technology, and highly capable memory built-in self-test (MBIST) feature. Micron says its new HBM4 provides seamless integration for customers and partners developing next-generation AI platforms.

Micron's new HBM4 has a 2048-bit memory interface that pushes more than 2.0TB/sec per memory stack, and over 60% more performance over its previous-gen HBM3 memory. These expansions in speed are perfect for AI, which is where HBM4 will lie: with AI GPUs inside of servers.

NVIDIA commissioned Samsung Electronics, SK hynix, and Micron to develop SoCAMM: a new memory module standard conceived by NVIDIA, with Micron reportedly the first to receive approval for mass production... putting the US-based memory manufacturer ahead of its South Korean rivals.

In new South Korean media reports picked up by insider @Jukenlosreve, we're hearing that the new SoCAMM standard is a memory module conceived by NVIDIA, consisting of 16-stacked LPDDR5X chips bundled in groups of four.

HBM is a DRAM that supports GPUs in AI accelerators, but SoCAMM is a new DRAM that attaches to the CPU, and while it connects to the CPU, the primary role of SoCAMM is to provide auxiliary support to make sure AI accelerates reach their peak performance. We should expect to see SoCAMM make an appearance in NVIDIA's next-gen AI GPUs -- Rubin -- which will be dropping in 2026.

SK hynix has teased its DRAM memory chip roadmap covering the next 30 years, which will include 4F2VG (vertical gate) technology that will improve DRAM chips, overcoming manufacturing challenges, with 3D DRAM as the "main driver" of its future growth.

SK hynix presented a new DRAM technology roadmap for the next 30 years and the direction for a sustainable innovation at the IEEE VLSI symposium 2025 held in Kyoto, Japan. At the event, Cha Seon Yong, the Chief Technology Officer (CTO) of SK hynix, delivered a prenary session on "Driving Innovation in DRAM Technology: Towards a Sustainable Future".

In this speech, CTO Cha said that it is getting more and more difficult to improve performance and capacity with scaling through current technology platforms. He said: "in order to overcome such limitations, SK hynix will apply the 4F² VG (Vertical Gate) platform and 3D DRAM technology to technologies of 10nm level or below with innovation in structure, materials, and components".

Huawei's new Ascend AI chips are reportedly experiencing major issues, so bad that big Chinese tech companies like Alibaba, Tencent, and TikTok parent company ByteDance are no longer adopting Huawei's new AI accelerators at scale.

In a new report from The Information, the new Huawei Ascend AI chips are experiencing overheating issues as well as risks associated with US sanctions, forcing the AI chips that are mainly sold to state-owned enterprises and local governments in China.

The report says that the ecosystem and technical issues surrounding Huawei's Ascend AI chips are causing major headaches, as companies are having to switch from CUDA at a "high cost".

NVIDIA's new B40 AI GPU destined for the Chinese market is expected to enter mass production in June, with ZJK Industrial Co., Ltd (ZJK) announcing it is preparing to expand its production capacity to meet the growing demand for NVIDIA's new "B40" AI GPU project.

In a new press release issued by ZJK spotted by insider @Jukanlosreve on X, we're hearing that NVIDIA is advancing the development of its new custom AI accelerator chip -- the new B40 -- which has been tailor-made for the Chinese market. NVIDIA's new B40 AI chip is based on the Blackwell GPU architecture, and will be aimed at mid-to-high-end markets, with mass production expected in June 2025.

ZJK CEO Ning Ding said: "The B40 project has many notable market advantages. It complements NVIDIA's current RTX Pro 6000 workstation GPU, which is sold globally. According to supply chain forecasts, shipments of the B40 chip are expected to exceed 1 million units by the end of 2025. As a supplier to NVIDIA, we anticipate a significant year-over-year revenue increase from this project".

NVIDIA's next-generation Rubin GPU and Vera CPU chips will reportedly finalize tape-out at TSMC in June, and will enter trial production with sample chips provided in September at the earliest according to the latest reports.

In a new post on X by financial analyst Dan Nystedt on X, we're hearing that TSMC will be making NVIDIA's next-gen Rubin AI GPUs on its N3P process node and will be using CoWoS-L advanced packaging. NVIDIA's new Rubin will be the company's first chiplet design for a GPU, after using traditional monolithic designs up until now (Hopper, Blackwell, etc are all monolithic large GPUs).

Rubin will require even more advanced packaging capacity from TSMC with a 4x reticle size, which is a chunky piece of silicon compared to Blackwell and its 3.3x reticle size. The I/O die on NVIDIA's next-gen Rubin GPUs will also be in the package, fabbed on TSMC's N5B process node.

NVIDIA is preparing a new AI GPU for China with new reports suggesting the new chip will be called the B30, meeting US export restrictions for the Chinese market, and is already getting major attention from Chinese cloud giants including TikTok parent company ByteDance, as well as Alibaba and Tencent.

The new NVIDIA B30 AI GPU is expected to shift away from HBM memory to GDDR7 memory, and will see NVIDIA counter-attacking the Chinese market that has been shifting to using in-house AI chip designs, with the B30 chip being fabbed by TSMC, it will also give Taiwanese supply chain partners to work directly with Chinese cloud companies.

In a new report from The Information, the new NVIDIA B30 AI GPU will allow customers to connect multiple B30 chips together to act as a single high-performance AI computing cluster. This means that even if the B30 is slower than homegrown AI chips in China, the fact that you can connect multiple one of them to act as one single high-perf computing cluster is a big, big deal.