Artificial Intelligence - Page 27

Intel announced its new Gaudi 3 AI accelerator earlier this year, rocking up to 128GB of HBM2e memory with up to 3.7TB/sec of memory bandwidth... but how does it stack up against NVIDIA's dominant Hopper H100 80GB AI GPU?

Well, that comes down to cost -- Intel's new Gaudi 3 AI accelerator costs just $16,000 per AI accelerator with 128GB of HBM2e memory, compared to $30,000+ for NVIDIA's Hopper H100 AI GPU with 80GB of faster HBM3 memory. Intel is charging around $65,000 per 8 x Gaudi 2 accelerators on a baseboard, and $128,000 per 8 x Gaudi 3 AI accelerators on a baseboard.

This means that a single Gaudi 2 AI accelerator costs around $8125, while Gaudi 3 costs $15,650, when purchased in bulk, with the baseboards, of course.

NVIDIA's new GB200 Superchip is getting another supplier according to analyst Ming-Chi Kuo, reporting that Alpha and Omega Semiconductor Ltd (AOSL) is to join the GB200 AI server train.

In a new post on Kuo's own Medium blog and a post on X, the analyst said: "Alpha and Omega Semiconductor Ltd (AOSL) is poised to become a new GB200 supplier and one of the leading beneficiaries of NVIDIA's certification process change".

NVIDIA CEO Jensen Huang said recently that the company will be releasing new versions of its flagship AI chips annually, and while this will absolutely enhance NVIDIA's competitiveness, it introduces other issues. Execution, acceleration of supplier certification processes, as well as accelerating system development.

TSMC and contract chip design partner Global Unichip have won the bulk orders for base dies used with SK hynix's next-generation HBM4 memory chips.

In a new report from UDN, we're learning that after being the exclusive OEM for AI chips for technology giants like NVIDIA and AMD, TSMC is seizing another AI business opportunity. UDN reports that TSMC will collaborate with its special application IC (ASIC) design service factory Creative, now successfully developing the HBM key peripheral components required for AI servers, jointly inking a "large order" for base dies required for next-gen HBM4 memory.

UDN's report says that TSMC and Creative have "never commented on order developments," but the legal source of the site said that demand for AI is strong, and not just HPC-related chips, but HBM demand is "growing rapidly" and is "becoming a new business opportunity in the market" which has three major memory chip companies including SK hynix, Samsung, and Micron, investing as much as they can into their respective HBM supply chains.

NVIDIA's new Blackwell GB200 and B-series AI GPUs, including the B100 and B200, have received a "large number of customers," and that "demand exceeds supply," reports UDN.

After NVIDIA aggressively increased the production volume of TSMC's advanced process, the "order-chasing effect" spread to the back-end packaging and testing plants. ASE Investment Holdings and KYEC operations have "exploded," reports UDN, with related order volume doubling in the fourth quarter.

UDN reports from its sources that KYEC is "full of new orders" from NVIDIA, that it is moving internally to welcome those large volumes of orders, and that it needs to be relocated for this purpose alone. More production capacity can then meet NVIDIA's growing (unstoppable) needs.

Because of US sanctions, NVIDIA has been restricted from supplying multiple countries across the Middle East, but more importantly, China has been restricted from shipping its leading AI GPUs.

It appears that might be changing with a new report from Reuters teasing that NVIDIA believes it can get an export license to sell its AI GPUs to a telecom giant that operates in multiple Middle East countries. NVIDIA has reportedly inked a deal with Ooredoo to "introduce its AI technology" in datacenters across five Middle Eastern countries.

This would mark NVIDIA's first major expansion that has US export regulations stamped on it, somewhere that the company isn't allowed to provide its AI GPUs to: the Middle East (it's not just China). Ooredoo is a telecom giant that operates in the Middle East, and will soon be powered by NVIDIA AI and HPC GPUs across its datacenters located in Qatar, Algeria, Tunisia, Oman, Kuwait, and the Maldives.

Apple has announced it will be pausing the rollout of its Apple Intelligence across Europe as the company has to adhere to the European Union competition regulations.

Apple Intelligence was announced at WWDC a few weeks ago and is expected to be released alongside iOS 18 and the iPhone 16. Moreover, Apple Intelligence is Apple's foray into artificial intelligence, as the new update to Apple's operating system will integrate a slew of AI-powered features, and completely overhaul Siri.

However, Apple has said in a statement on Friday that specific iOS 18 features won't be rolled out in the EU, and those are: Phone Mirroring, SharePlay Screen Sharing enhancements, and Apple Intelligence.

HBM is the key to AI chips of today and more especially, the future, with HBM production capacity at its limits as AI chip makers like NVIDIA, have been scooping it all up.

SK hynix, Samsung, and Micron lead the HBM memory chip side, with HBM3 and HBM3E orders from all three major DRAM manufacturers completely out of stock until the end of 2025. We've been reporting on that side of the industry heavily, but now all three HBM manufacturers -- SK hynix, Samsung, and Micron -- are expanding HBM production capacity to keep up with the continuously growing demand of the AI market.

This means that the demand for application materials outsourcing equipment and components from the three leading DRAM factories has been "set aside to next year" reports UDN, with their legal person pointing out that due to multiple new customer orders from major semiconductor equipment manufacturers has pushed the order visibility into Q2 2025, meaning that performance of the markets is expected to be strong throughout this year, and into 2025 easily.

NVIDIA has been pushing new technologies into gaming through ray tracing, DLSS, AI-powered technologies, and so much more... but we all know that AI is the future of gaming.

NVIDIA CEO Jensen Huang said one of those AI-powered technologies could be similar to how RTX Remux acts, but would do it in real-time. Jensen was asked recently "AI has been used in games for a while now, I'm thinking DLSS and now ACE. Do you think it's possible to apply multimodality AIs to generate frames?"

Jensen replied: "AI for gaming - we already use it for neural graphics, and we can generate pixels based off of few input pixels. We also generate frames between frames - not interpolation, but generation. In the future we'll even generate textures and objects, and the objects can be of lower quality and we can make them look better. We'll also generate characters in the games - think of a group of six people, two may be real, and the others may be long-term use AIs".

Elon Musk's startup xAI has made a massive order with various hardware companies to create a powerful supercomputer that will be used to power Grok.

The CEO of Dell Technologies, Michael Dell, took to his personal X account to announce that Dell has partnered with xAI to construct a "Dell AI factory". Dell isn't the only company onboard with the supercomputer project, as NVIDIA will provide the horsepower with its powerful AI GPUs, and Supermicro will handle one-half of the total number of racks. Elon Musk clarified on his personal X account, "Dell is assembling half of the racks that are going into the supercomputer that xAI is building."

As for the horsepower, Musk stated earlier this month that xAI was on the verge of bringing 100,000 liquid-cooled NVIDIA H100 GPUs online, but Musk isn't totally thrilled with their efficiency, especially with NVIDIA's next-generation Blackwell GPUs on the horizon. Musk stated earlier this month, "Given the pace of technology improvement, it's not worth sinking one gigawatt of power into H100s. Next big step would probably be ~300k B200s with CX8 networking next summer."

SK hynix is expediting its HBM roadmap that includes HBM4 and HBM4E memory, which were originally planned for mass production in 2026 and 2027, respectively.

Now, the timelines have been moved up, and HBM4 is set for mass production in 2025, while HBM4E will enter mass production in 2026, according to sources from Business Korea. The adjustments "align" with NVIDIA's accelerated AI accelerator release cycle, which has shortened from two years to one year.

NVIDIA's current-gen Hopper GPU architecture with the H100 and beefed-up H200 are already dominating, with the new Blackwell GPU architecture with the B100 and B200 AI GPUs sporting faster HBM3E memory. But, then NVIDIA teased its next-gen Rubin R100 AI GPU, which will feature HBM4 memory and drop in Q4 2025.