Artificial Intelligence News - Page 3

AMD's new Instinct MI300X AI accelerator with 192GB of HBM3E has had a deep dive at Hot Chips 2024 this week, as well as the company teasing its refreshed MI325X with 288GB of HBM3E later this year.

Inside, AMD's new Instinct MI300X AI Accelerator features a total of 153 billion transistors, using a mix of TSMC's new 5nm and 6nm FinFET process nodes. There are 8 chiplets that feature 4 shared engines, and each shared engine contains 10 compute units.

The entire chip packs 32 shader engines, with a total of 40 shader engines inside of a single XCD and 320 in total across the entire package. Each individual XCD has its dedicated L2 cache, and out the outskirts of the package, features the Infinity Fabric Link, 8 HBM3 IO sites, and a single PCIe Gen5 link with 128GB/sec of bandwidth that connects the MI300X to an AMD EPYC CPU.

Intel has announced new details on its new Xeon 6 SoC, Lunar Lake mobile processor, and Gaudi 3 AI accelerator and its OCI chiplet at Hot Chips 2024 this week.

First off, is the new Intel 6 SoC that combines the compute chiplet from Intel Xeon 6 processors with an edge-optimized I/O chiplet built on Intel 4 process technology. This enables the Xeon 6 SoC to deliver performance boosts over previous-generation Xeon CPUs with improved power efficiency and transistor density compared to previous-gen tech.

Intel will have more details about its next-gen AI PC processor, Lunar Lake, with Arik Gihon, the lead client CPU SoC architect, to talk about the new Lunar Lake CPU and how it's designed to "set a new bar for x86 power efficiency while delivering leading core, graphics and client AI performance".

IBM has just unveiled its new Telum II processor and Spyre AI accelerator, which it plans to use inside of its new IBM Z mainframe systems powering AI workloads.

The company provided details of the architecture of its new Telum II processor and Spyre AI accelerator, which are designed for AI workloads on the next-gen IBM Z mainframes. The new mainframes will accelerate traditional AI workloads, as well as LLMs using a brand new ensemble method of AI.

IBM's new Telum II processor features 8 high-performance cores running at 5.5GHz, with 36MB L2 cache per core and a 40% increase in on-chip cache capacity for a total of 360MB. The virtual level-4 cache of 2.88GB per processor drawer provides a 40% increase over the previous generation. The integrated AI accelerator allows for low-latency, high-throughput in-transaction AI inferencing, for example enhancing fraud detection during financial transactions, and provides a fourfold increase in compute capacity per chip over the previous generation.

SK hynix is aiming to have its HBM4 memory tape-out in Q4 2024, ready for NVIDIA's next-gen Rubin R100 AI GPU coming in 2025.

In a new report from ZDnet, we're learning that SK hynix is nearing the final stage of commercializing its next-generation HBM4 memory, with the design drawings to be transferred to the manufacturing process, or "tape out". According to ZDnet's industry sources, SK hynix plans to complete the tape out of its HBM4 for NVIDIA in October, so we're just weeks away.

HBM4 offers a huge 40% increase in bandwidth, and a reduced power consumption of a rather incredible 70% to HBM3E, the fastest memory in the world. HBM4 density will be 1.3x higher, with all of these advancements combined, the leap in performance and efficiency is a key driver in NVIDIA's continued AI GPU dominance.

TSMC is expected to make over NT $1 trillion (around $31 billion USD or so) in revenue from its 3nm and 5nm process nodes, in just a span of 9 months.

DigiTimes reports that TSMC will generate around $31 billion from just two of its high-end semiconductor nodes, thanks to their unstoppable demand -- customers like Apple, AMD, NVIDIA, Intel, Qualcomm, MediaTek -- with TSMC seeing huge revenue increases for Q2 2024 to NT$336.7 billion, or around 40% of their total revenue in Q1 2023.

TSMC estimates it will generate NT $754 billion (around $23 billion USD or so) from its 3nm and 5nm process nodes in Q3 2024, with major customers in Apple and NVIDIA.

In a recently shared data center AI chip roadmap posted on X, we get a good look at what companies have on the market already, and what's in the AI chip pipeline through to 2027. Check it out:

The list includes chip makers NVIDIA, AMD, Intel, Google, Amazon, Microsoft, Meta, ByteDance, and Huawei. You can see the list of NVIDIA AI GPUs includes the Ampere A100 through to the Hopper H100, GH200, H200 AI GPUs, and into the Blackwell B200A, B200 Ultra, GB200 Ultra and GB200A. But after that -- which we all know is coming -- is Rubin and Rubin Ultra, both rocking next-gen HBM4 memory.

We also have AMD's growing line of Instinct MI series AI accelerators, with the MI250X through to the new MI350 and the upcoming MI400 listed in there for 2026 and beyond.

Mistral-NeMo-Minitron 8B is a "miniaturized version" of the new highly accurate Mistral NeMo 12B AI model. It is tailor-made for GPU-accelerated data centers, the cloud, and high-end workstations with NVIDIA RTX hardware. Accuracy is often sacrificed to ensure performance regarding scalable AI models; Mistral AI and NVIDIA's new Mistral-NeMo-Minitron 8B deliver the best of both worlds.

Small enough to run in real-time on a workstation or desktop rig with a high-end GeForce RTX 40 Series graphics card, with NVIDIA, noting that the 8B or 8 billion variant excels when it comes to benchmarks for AI chatbots, virtual assistant, content generation, and educational tools.

Available and packaged as an NVIDIA NIM microservice (downloadable via Hugging Face), Mistral-NeMo-Minitron 8B is currently outperforming Llama 3.1 8B and Gemma 7B in the all-important accuracy category in at least nine popular benchmarks for AI language models.

NVIDIA will be hosting a Hot Chips Talk next week, deep diving into its new Blackwell GPU architecture while reminding the world that its Blackwell GPU has the highest AI compute, memory bandwidth, and interconnect bandwidth ever in a single GPU.

At Hot Chips 2024 next week, NVIDIA will go into more detail about the Blackwell GPU architecture while also reminding us that it features not one, but two reticle-limited GPUs merged into one. One of the limitations of lithographic chipmaking tools is that they've been designed to make ICs (integrated circuits) that are no bigger than around 800 square millimeters, which is referred to as the "reticle limit".

NVIDIA has two reticle-limited AI GPUs together (104 billion transistors per chip, 208 billion transistors in total). NVIDIA will discuss building to the reticle limit, and how it delivers on the highest communication density, lowest latency, and optimal energy efficiency during its Hot Chips Talk.

NVIDIA will discuss using AI to build next-generation chips for AI at the upcoming Hot Chips 2024 event next week.

The company will discuss how NVIDIA designs some of the most complex products on the planet, with its new Blackwell B200 AI GPU featuring 208 billion transistors, made on TSMC's new 4NP process node. At Hot Chips, NVIDIA will discuss using generative AI to generative optimized Verilog code.

What's Verilog code? Verilog is a hardware description language that describes circuits in the form of code. It's used for design and verification of processors, with NVIDIA building an Agentic AI LLM application to accelerate Computer Aided Engineering (CAE) that generates Verilog code, which can:

NVIDIA has powered up its new Blackwell AI GPUs and run them in real-time inside of their data centers, while teasing it will provide more details about its Blackwell GPU architecture at Hot Chips next week.

The company has been embroiled in rumors of its Blackwell AI GPUs having issues big enough to require a redesign, and issues with Blackwell AI servers leaking through their water-cooling setups. NVIDIA has now shown its new Blackwell AI GPUs running in real-time, with Blackwell on-track to ramp into production and ship (in small quantities) to customers in Q4 2024.

NVIDIA also teased new pictures of various trays available in the Blackwell family, with these first images of Blackwell trays being teased showing just how much engineering and design work goes into these things. It's truly incredible, almost like a work of art... and that's on the outside.