Artificial Intelligence - Page 41

SK hynix unveiled a new technology that will be used to create a "dream memory chip" that is capable of storing data and performing calculations for AI.

The South Korean memory giant unveiled its new technology during the International Memory Workshop (IMW 2024) held from May 12-15 at the Walkerhill Hotel in the Gwangjin district of Seoul, South Korea. The new technology enhances the accuracy of Multiply Accumulate (MAC) operations in Analog Computing in Memory (A-CIM) semiconductors using oxygen diffusion barrier technology.

MAC operations are critical for the high-speed multiplication and accumulation processes required in artificial intelligence (AI) inference and learning. SK hynix's recent development is a significant step for the company in the competitive field of creating a "dream memory semiconductor" that can both store information and perform calculations, passing the traditional limitations of memory-only semiconductors.

OpenAI has unveiled a new AI model that is designed to analyze audio, visual and text, and provide answers based on what it "sees/hears".

The company behind the immensely popular AI tool ChatGPT announced its latest flagship model called GPT-4o (omni), which OpenAI describes as being a step towards a "much more natural human-computer interaction". The new AI model is expected to match the performance of GPT-4 Turbo at processing text and code input, while simultaneously being faster and 50% cheaper with its API, making it a more affordable choice for third-party app integration.

More specifically, users will be able to submit a query by voice about what the AI agent is able to "see" on the devices screen, and an example of this would be asking the AI what game two people can play. OpenAI demonstrated this with two people that verbally asked the AI "what game can we play". The AI used the smartphone camera to "see" the two people sitting in front of it and suggested playing rock, paper, scissors. The quick demonstration showed the AI model being able to fluently interact with the individuals and also be extremely responsive to interruptions and new commands.

NVIDIA's new Blackwell GPU architecture is going to make the company a lot of money, and while we know the B200 AI GPUs will cost $30,000 to $40,000 each -- CEO Jensen Huang said that just after the GPUs were announced -- but, the GB200 Superchip (CPU + GPU) will cost upwards of $70,000.

In a post on X by a senior writer at Barron's, the new NVIDIA GB200 Superchip will cost between $60,000 and $70,000. We already know that an NVIDIA DGX NVL72 AI server cabinet will cost $3 million per AI server, which gets filled with 72 x B200 GPUs and 36 x Grace CPUs.

The new NVIDIA DGX NVL72 is the AI server with the most computing power, and thus, the highest unit price. Inside, the DGX NVL72 features 72 built-in Blackwell-based B200 AI GPUs and 36 Grace CPUs (18 servers in total with dual Grace CPUs and 36 x B200 AI GPUs per server) with 9 switches. The entire cabinet is designed by NVIDIA in-house, and cannot be modified, it is 100% made, tested, and provided by NVIDIA.

Samsung has reportedly failed to pass specific stages of HBM3E memory verification standards from NVIDIA, which is surely going to cause a headache for the South Korean memory giant.

SK hynix has been pumping out HBM3 and now HBM3E while still preparing not just the next-generation HBM4 memory, but even HBM4E memory... while South Korean HBM rival, Samsung, can't get its act together with HBM3 memory for NVIDIA according to Korean news outlet AlphaBiz.

Samsung has reportedly failed qualification tests for its HBM3 8-layer memory, which is a serious situation to be in considering how bleeding-edge HBM is, and how Samsung has been acting in an emergency-style manner to get its HBM business flourishing... and now this gigantic roadblock, as NVIDIA prepares its beefed-up Hopper H200 AI GPU, and next-generation Blackwell B100, B200, and GB200 AI GPUs which all use HBM3E memory.

Arm is developing its own artificial intelligence (AI) chips, with the first AI chips made by the company expected to launch in 2025.

The UK-based company will spool up an AI chip division that will deliver a prototype AI chip by spring 2025 according to a report from Reuters. The mass production of Arm's new AI chip will be handled by contract manufacturers -- TSMC -- and is expected to start in autumn 2025.

Arm Holdings is a SoftBank Group subsidiary -- SoftBank owns a 90% share in Arm -- with SoftBank CEO Masayoshi Son preparing a huge $64 billion strategy to transform SoftBank into a powerhouse AI company. Negotiations are reportedly already happening with TSMC and others to secure production capacity.

SK hynix has announced it plans to complete the development of its next-gen HBM4E memory by as early as 2026, preparing for the next-gen AI GPUs of the future.

SK hynix's head of the HBM advanced technology team, Kim Gwi-wook, announced the news this week of the direction of next-generation HBM development at the International Memory Workshop (IMW 2024). HBM was developed by SK hynix in 2014, with HBM2 (2nd generation) in 2018, HBM2E (3rd generation) in 2020, HBM3 (4th generation) in 2022, and HBM3E (5th generation) was introduced this year.

There's a two-year cadence between HBM generations, with HBM3E unleashed this year, it means that HBM4 (6th generation) should drop in 2025, and HBM4E (7th generation) in 2026. That's a bit faster than two years for HBM4 and HBM4E, which is because SK hynix is predicting that HBM performance advancements would become faster than previous generations.

GigaIO and SourceCode have just unveiled Gryf, an ultra-portable AI supercomputer-class system that weighs less than 55 pounds, and fits inside of a TSA-friendly carry-on suitcase. Impressive.

Gryf can handle data collection and processing on a scale that would usually see the data sent off-site, this means that the suitcase-sized supercomputer handles super-fast processing and analysis, all in a suitcase. Gryf supports disaggregating and reaggregating its GPUs, with owners capable of customizing the system's hardware configuration in the field, and on-the-fly.

You can create the absolute optimal hardware configuration for one assigned workload, and then the next workload gets another optimized hardware configuration. Each Gryf has multiple slots filled with compute, storage, accelerator, and network sleds that are perfect for their respective workloads. There's 6 sled slots in total, where you can insert and remove the modules as required.

NVIDIA has published a new blog post providing some more details about the next level of performance offered by its new Blackwell GPU architecture.

The new blog post by NVIDIA shows the gigantic performance leap that Blackwell will deliver for the research industry including quantum computing, drug discovery, fusion energy, physics-based simulations, weather simulations, scientific computing, and more.

NVIDIA has another major goal with Blackwell -- other than industry-leading AI performance -- in that Blackwell can simulate weather patterns 200x cheaper than Hopper, and use 300x less energy while running digital twins simultaneously encompassing the globe with 65x less cost, and 58x less energy used. Absolutely astonishing numbers from Blackwell.

I still remember using and playing games on the Commodore 64, but I never thought I'd see the day when the old-school PC was running generative AI to generate creative retro sprites. Check it out:

Nick Bild is a developer and hobbyist who documented his journey of building a generative AI tool for the Commodore 64, that can be used to create 8 x 8 sprites that are displayed at the 64 x 64 resolution. The idea behind this is to use AI to help inspire game design concepts, but we're talking about the Commodore 64 here, so we're not going to get some AI-powered Crysis on the C64.

Training the generative AI model was done on a traditional PC, so while the AI model itself runs on the Commodore 64, you'll need a modern PC to get it up and running. It will take 20 minutes or so to run just 90 iterations for the final 64 x 64 image, so it's not going to blow NVIDIA's current-gen Hopper H100 AI GPU out of the water, or put AI companies out of business. Impressive for the Commodore 64, nonetheless.

According to a new report at Reuters, OpenAI (the massive AI firm behind ChatGPT, backed by billions from Microsoft) is planning to launch an AI-powered search engine on Monday. The engine will compete with Google and Perplexity, a competing AI search startup founded by a former OpenAI researcher.

Going up against the search giant that is Google with the aid of AI is not uncommon. Microsoft's long-running Bing search recently added OpenAI ChatGPT integration - but only for paid customers. AI and search engines are set to go hand-in-hand, as Google is also integrating generative AI into search and other products like Gmail.

The Reuters report cites 'two sources familiar with the matter,' so nothing is official. However, OpenAI's stealth launch of a new search engine (possibly in beta or limited form) next week would mark an exciting turn for the company. Bloomberg has reported on the company's search engine plans in the past, so it sounds like it's only a matter of time.

A new report from Bloomberg reveals Microsoft has created a new generative AI model that is designed specifically for US intelligence agencies.

The report states the main difference between this new AI model and others that power popular AI tools, such as ChatGPT, is that it's completely divorced from the internet, making it the first of its kind. Known AI models such as ChatGPT, DALL-E, and Microsoft's Copilot rely on cloud services to process prompts, train data, and reach conclusions. However, the AI model now handed over to US intelligence agencies doesn't require any cloud services, meaning it is completely devoid of any internet access and, therefore, secure.

Why do US intelligence agencies want an advanced AI model? According to the report, due to the security of the AI model, top-secret information can now be inputted and analyzed, which will help intelligence agencies understand and filter through large swaths of classified information.

Microsoft has snapped up the same location Foxconn acquired to build an LCD panel manufacturing plant, but instead of panels, Microsoft will be constructing a data center.

After years of delays Foxconn failed to materialize the LCD manufacturing project at Mount Pleasant, Wisconsin, and over the years Microsoft has snapped up more and more of the land that was originally set aside for Foxconn's project, eventually resulting in the Taiwanese company pulling out and Microsoft scooping up the rest of the site.

Microsoft's proposal for a data center features infrastructure and community improvements to the local area, with promises that it will up-skill 100,000 residents across the state to be component at generative AI technologies such as Microsoft Copilot, train and certify 3,000 local AI software developers, and 1,000 cloud datacenter technicians. Moreover, Microsoft President Brad Smith, backed the push for AI as it has the potential to revolutionize manufacturing plants, assist workers and create more jobs.

NVIDIA is still cooking its new Blackwell GPU architecture and B200 AI GPU, and while we've had teases of the next-gen Vera Rubin GPU, now we're hearing the next-gen R100 AI GPU will be in mass production in Q4 2025.

In a new post by industry insider Ming-Chi Kuo, NVIDIA's next-generation AI chip will enter mass production in Q4 2025 with the R-series and R100 AI GPU, with the system/rack solution to enter mass production in Q1 2026. NVIDIA's next-gen R100 will be made on TSMC's newer N3 process node, compared to B100 which uses TSMC N4P, with R100 using TSMC's newer CoWoS-L packaging (the same as B100).

NVIDIA's next-gen R100 AI GPU features around 4x reticle design, compared to the B100 with 3.3x reticle design, while the interposer size for R100 "has yet to be finalized," with Ming-Chi saying there are 2-3 options. R100 will feature 8 x HBM4 units, while GR200's new Grace CPU will use TSMC's N3 process (compared to TSMC's N5 for GH200 and GB200's Grace CPUs).

The US government is reportedly preparing to make another move against China to prevent the nation from gaining access to the US's best artificial intelligence capabilities.

The Biden administration has already taken measures to prevent China from gaining AI supremacy by banning the exportation of specific high-end NVIDIA graphics cards, which are used to train the AI models, and proposing a rule that requires all US cloud companies to inform the government when foreign customers are using their cloud systems to train AI models.

According to reports, more guardrails are being considered by the Commerce Department, which plans on targeting the exportation of proprietary or closed-source AI models. The idea behind these new purported regulations is to prevent US-based AI giants such as the Microsoft-funded OpenAI, the company behind popular AI tool ChatGPT, or Google DeepMind, creators of Gemini, from taking their world leading AI models to global market and selling them to the highest bidder.

Meta has purchased 500,000 more AI GPUs for a total of 1 million AI GPUs, which is valued at $30 billion.

We're hearing about the gargantuan AI GPU hardware investment from Meta AI boss Yann LeCun at the AI Summit, also saying that upcoming variations of its Llama 3 large language model are on the way. LeCun emphasized the computational limitations and GPU costs as things that are slowing the progression of AI.

OpenAI CEO Sam Altman plans to spend $50 billion a year on AGI development (artificial general intelligence) by using 720,000 NVIDIA H100 AI GPUs that cost a hefty $21.6 billion. Microsoft is aiming for 1.8 million AI GPUs by the end of 2024, while OpenAI wants to have 10 million AI GPUs before the end of the year.

In a new interview with MIT Technology Review, OpenAI CEO Sam Altman revealed what the future will look like with artificial intelligence-powered systems becoming more engrained in our lives.

The CEO of one of the companies leading the charge in AI development began the interview by saying AI tools will replace the smartphone as the most dependent piece of technology in our daily lives, and the capabilities of AI will be so great that it will feel like "this thing that is off helping you." Altman went on to describe the perfect app for AI would be a "super-competent colleague that knows absolutely everything about my whole life, every email, every conversation I've ever had, but doesn't feel like an extension."

Additionally, this level of AI could attempt things outside its known capabilities, fail, and then come back to the user with follow-up questions. The answers the user provides to those questions are then integrated into its second attempt at the task. So, what will power this crazy new AI? Altman believes there's a chance that we won't even need a specific piece of hardware to use this AI on the go, as the new app could simply access the cloud.

Apple is reportedly working on its own AI chip according to sources of The Wall Street Journal, which reports the Curpentino-based giant is building AI chips for its data centers to run new AI-based features that'll be announced at WWDC 2024 next month.

The World Wide Developers Conference (WWDC) is where APple will unveil its plans for the future of AI with its products and services, with the WSJ reporting: "Apple has been working on its own chip designed to run artificial intelligence software in data center servers, a move that has the potential to provide the company with a key advantage in the AI arms race".

The project is called ACDC which stands for Apple Chips in Data Center, but I can see some truly awesome marketing from Apple using the "ACDC" branding if they do it right. Come on, Apple... you know you're going to do it anyway.

Samsung has established a new dream team of engineers to secure HBM memory chip deals for AI GPUs from NVIDIA.

The news is coming from South Korean outlet KED Global, which reports Samsung's new task force features about 100 "excellent engineers" who have been working on improving manufacturing yields and quality with the first objective being passing NVIDIA's tests.

NVIDIA CEO Jensen Huang asked Samsung to raise the yields and quality of its 8-layer and 12-layer HBM3E memory chips for supply according to industry insiders on Monday. HBM3E memory is the cornerstone of NVIDIA's next-gen Blackwell B200 AI GPUs, as well as the new beefed-up Hopper H200 AI GPU, each with HBM3E memory mostly from Samsung's South Korean HBM rival: SK hynix.

TSMC is focused on getting its advanced packaging production capacity to new levels, with the company boosting advanced packaging capacity at its Zhonghe, Nanka, and Jiake fabs, which are "all in the process of expanding production".

The news is coming from Taiwanese media outlet Ctee, reporting that the Chiayi Science Park was finalized this year and is expected to build two advanced packaging plants first. The first phase of Jiake will see the breaking ground ceremony in the coming weeks, and into operation in the second half of the year.

The second phase of Jiake is expected to begin construction in Q2 2024, with operation in Q1 2027 which is still a few years away, which Ctee reports will continue to expand the market share of AI and HPC.

AMD and NVIDIA are both pumping out new generations of AI hardware, with the two GPU giants eating all of TSMC's advanced packaging capacity for not just 2024, but also 2025.

In a new report from UDN, the site reports that TSMC will contract advanced packaging production capacity of its in-house CoWoS and SoIC for this and next year, as AI GPU hardware orders aren't slowing down. TSMC expects revenue contributed by server AI processors to more than double this year, and once NVIDIA's new Blackwell B200 AI GPUs are being pumped out later this year, 2025 is going to be a bananas year for TSMC.

In response to the insatiable AI demand, TSMC is now actively expanding its advanced packaging production capacity, with industry estimates that TSMC's CoWoS monthly production capacity will be between 45,000 and 50,000 by the end of 2024. This is a gigantic 3x increase from the 15,000 in 2023, and it's expected to hit 50,000 in 2025.