Have you ever dreamed of having a real conversation with an NPC in a video game? Not just one gated within a dialogue tree of pre-determined answers, but an actual conversation, conducted through spontaneous action and reaction? Lately, a small R&D team at Ubisoft's Paris studio, in collaboration with Nvidia's Audio2Face application and Inworld's Large Language Model (LLM), have been experimenting with generative AI in an attempt to turn this dream into a reality. Their project, NEO NPC, uses GenAI to prod at the limits of how a player can interact with an NPC without breaking the authenticity of the situation they are in, or the character of the NPC itself.

Considering that word—authenticity—the project has had to be a hugely collaborative effort across artistic and scientific disciplines. Generative AI is a hot topic of conversation in the videogame industry, and Senior Vice President of Production Technology Guillemette Picard is keen to stress that the goal behind all genAI projects at Ubisoft is to bring value to the player; and that means continuing to focus on human creativity behind the scenes. "The way we worked on this project, is always with our players and our developers in mind," says Picard. "With the player in mind, we know that developers and their creativity must still drive our projects. Generative AI is only of value if it has value for them."

MAINGEAR, a leading provider of high-performance custom PC systems, and Phison, a global leader in NAND controllers and storage solutions, today unveiled groundbreaking MAINGEAR PRO AI workstations with Phison's aiDAPTIV+ technology. Specifically engineered to democratize Large Language Model (LLM) development and training for small and medium-sized businesses (SMBs), these ultra-powerful workstations incorporate aiDAPTIV+ technology to deliver supercomputer LLM training capabilities at a fraction of the cost of traditional AI training servers.

As the demand for large-scale generative AI models continues to surge and their complexity increases, the potential for LLMs also expands. However, this rapid advancement in LLM AI technology has led to a notable boost in hardware requirements, making model training cost-prohibitive and inaccessible for many small to medium businesses.

AMD has caught up with NVIDIA and Intel in the race to get a locally run AI chatbot up and running on its respective hardware. Team Red's community hub welcomed a new blog entry on Wednesday—AI staffers published a handy "How to run a Large Language Model (LLM) on your AMD Ryzen AI PC or Radeon Graphics Card" step-by-step guide. They recommend that interested parties are best served by downloading the correct version of LM Studio. Their CPU-bound Windows variant—designed for higher-end Phoenix and Hawk Point chips—compatible Ryzen AI PCs can deploy instances of a GPT based LLM-powered AI chatbot. The LM Studio ROCm technical preview functions similarly, but is reliant on Radeon RX 7000 graphics card ownership. Supported GPU targets include: gfx1100, gfx1101 and gfx1102.

AMD believes that: "AI assistants are quickly becoming essential resources to help increase productivity, efficiency or even brainstorm for ideas." Their blog also puts a spotlight on LM Studio's offline functionality: "Not only does the local AI chatbot on your machine not require an internet connection—but your conversations stay on your local machine." The six-step guide invites curious members to experiment with a handful of large language models—most notably Mistral 7b and LLAMA v2 7b. They thoroughly recommend that you select options with "Q4 K M" (AKA 4-bit quantization). You can learn about spooling up "your very own AI chatbot" here.

Apple has been experimenting with Large Language Models (LLMs) that power most of today's AI applications. The company wants these LLMs to serve the users best and deliver them efficiently, which is a difficult task as they require a lot of resources, including compute and memory. Traditionally, LLMs have required AI accelerators in combination with large DRAM capacity to store model weights. However, Apple has published a paper that aims to bring LLMs to devices with limited memory capacity. By storing LLMs on NAND flash memory (regular storage), the method involves constructing an inference cost model that harmonizes with the flash memory behavior, guiding optimization in two critical areas: reducing the volume of data transferred from flash and reading data in larger, more contiguous chunks. Instead of storing the model weights on DRAM, Apple wants to utilize flash memory to store weights and only pull them on-demand to DRAM once it is needed.

Two principal techniques are introduced within this flash memory-informed framework: "windowing" and "row-column bundling." These methods collectively enable running models up to twice the size of the available DRAM, with a 4-5x and 20-25x increase in inference speed compared to native loading approaches on CPU and GPU, respectively. Integrating sparsity awareness, context-adaptive loading, and a hardware-oriented design pave the way for practical inference of LLMs on devices with limited memory, such as SoCs with 8/16/32 GB of available DRAM. Especially with DRAM prices outweighing NAND Flash, setups such as smartphone configurations could easily store and inference LLMs with multi-billion parameters, even if the DRAM available isn't sufficient. For a more technical deep dive, read the paper on arXiv here.

AMD (NASDAQ:AMD) today announced revenue for the third quarter of 2023 of $5.8 billion, gross margin of 47%, operating income of $224 million, net income of $299 million and diluted earnings per share of $0.18. On a non-GAAP basis, gross margin was 51%, operating income was $1.3 billion, net income was $1.1 billion and diluted earnings per share was $0.70.

"We delivered strong revenue and earnings growth driven by demand for our Ryzen 7000 series PC processors and record server processor sales," said AMD Chair and CEO Dr. Lisa Su. "Our data center business is on a significant growth trajectory based on the strength of our EPYC CPU portfolio and the ramp of Instinct MI300 accelerator shipments to support multiple deployments with hyperscale, enterprise and AI customers."

Lenovo Group today announced first quarter results, reporting Group revenue of US$12.9 billion and net income of US$191 million on a non-Hong Kong Financial Reporting Standards (HKFRS) basis. Revenue from the non-PC businesses accounted for 41% of Group revenue, with the service-led business achieving strong growth and sustained profitability - further demonstrating the effectiveness of Lenovo's intelligent transformation strategy.

The Group continues to take proactive actions to keep its Expenses-to-Revenue (E/R) ratio resilient and drive sustainable profitability, whilst also investing for growth and transformation. It remains committed to doubling investment in innovation in the mid-term, including an additional US$1 billion investment over three years to accelerate artificial intelligence (AI) deployment for businesses around the world - specifically AI devices, AI infrastructure, and AI solutions.

Cerebras Systems, the pioneer in accelerating generative AI, and G42, the UAE-based technology holding group, today announced Condor Galaxy, a network of nine interconnected supercomputers, offering a new approach to AI compute that promises to significantly reduce AI model training time. The first AI supercomputer on this network, Condor Galaxy 1 (CG-1), has 4 exaFLOPs and 54 million cores. Cerebras and G42 are planning to deploy two more such supercomputers, CG-2 and CG-3, in the U.S. in early 2024. With a planned capacity of 36 exaFLOPs in total, this unprecedented supercomputing network will revolutionize the advancement of AI globally.

"Collaborating with Cerebras to rapidly deliver the world's fastest AI training supercomputer and laying the foundation for interconnecting a constellation of these supercomputers across the world has been enormously exciting. This partnership brings together Cerebras' extraordinary compute capabilities, together with G42's multi-industry AI expertise. G42 and Cerebras' shared vision is that Condor Galaxy will be used to address society's most pressing challenges across healthcare, energy, climate action and more," said Talal Alkaissi, CEO of G42 Cloud, a subsidiary of G42.

When OpenAI announced its GPT-4 model, it first became a part of ChatGPT, behind the paywall for premium users. The GPT-4 is the latest installment in the Generative Pretrained Transformer (GPT) Large Language Models (LLMs). The GPT-4 aims to be a more capable version than the GPT-3.5 that powered ChatGPT at first, which was capable once it launched. However, it seems like the performance of GPT-4 has been steadily dropping since its introduction. Many users noted the regression, and today we have researchers from Stanford University and UC Berkeley, who benchmarked the GPT-4 performance in March 2023, and the model's performance in June 2023 in tasks like solving math problems, visual reasoning, code generation, and answering sensitive questions.

The results? The paper shows that GPT-4 performance has been significantly degraded in all the tasks. This could be attributed to improving stability, lowering the massive compute demand, and much more. What is unexpected, GPT-3.5 experienced a significant uplift in the same period. Below, you can see the examples that were benchmarked by the researchers, which also compare GTP-4 and GPT-3.5 performance in all cases.

Oracle today announced the addition of generative AI-powered capabilities within Oracle Fusion Cloud Human Capital Management (HCM). Supported by the Oracle Cloud Infrastructure (OCI) generative AI service, the new capabilities are embedded in existing HR processes to drive faster business value, improve productivity, enhance the candidate and employee experience, and streamline HR processes.

"Generative AI is boosting productivity and unlocking a new world of skills, ideas, and creativity that can have an immediate impact in the workplace," said Chris Leone, executive vice president, applications development, Oracle Cloud HCM. "With the ability to summarize, author, and recommend content, generative AI helps to reduce friction as employees complete important HR functions. For example, with the new embedded generative AI capabilities in Oracle Cloud HCM, our customers will be able to take advantage of large language models to drastically reduce the time required to complete tasks, improve the employee experience, enhance the accuracy of workforce insights, and ultimately increase business value."

ASUS today announced a showcase of the latest HPC solutions to empower innovation and push the boundaries of supercomputing, at ISC High Performance 2023 in Hamburg, Germany on May 21-25, 2023. The ASUS exhibition, at booth H813, will reveal the latest supercomputing advances, including liquid-cooling and AI solutions, as well as outlining a slew of sustainability breakthroughs - plus a whole lot more besides.

Comprehensive Liquid-Cooling Solutions
ASUS is working with Submer, the industry-leading liquid-cooling provider to demonstrate immersion-cooling solutions at ISC High Performance 2023, focused on ASUS RS720-E11-IM - the Intel -based 2U4N server that leverages our trusted legacy server architecture and popular features to create a compact new design. This fresh outlook improves the accessibility on I/O ports, storage and cable routing, and strengthens the structure to allow the server to be placed vertically in the tank, with durability assured.

The recent growth in demand for training Large Language Models (LLMs) like Generative Pre-trained Transformer (GPT) has sparked the interest of many companies to invest in GPU solutions that are used to train these models. However, countries like China have struggled with US sanctions, and NVIDIA has to create custom models that meet US export regulations. Carrying two GPUs, H800 and A800, they represent cut-down versions of the original H100 and A100, respectively. We reported about H800; however, it remained as mysterious as A800 that we are talking about today. Thanks to MyDrivers, we have information that the A800 GPU performance is within 70% of the regular A100.

The regular A100 GPU manages 9.7 TeraFLOPs of FP64, 19.5 TeraFLOPS of FP64 Tensor, and up to 624 BF16/FP16 TeraFLOPS with sparsity. A rough napkin math would suggest that 70% performance of the original (a 30% cut) would equal 6.8 TeraFLOPs of FP64 precision, 13.7 TeraFLOPs of FP64 Tensor, and 437 BF16/FP16 TeraFLOPs with sparsity. MyDrivers notes that A800 can be had for 100,000 Yuan, translating to about 14,462 USD at the time of writing. This is not the most capable GPU that Chinese companies can acquire, as H800 exists. However, we don't have any information about its performance for now.

NVIDIA's H100 has recently become available to use via Cloud Service Providers (CSPs), and it was only a matter of time before someone decided to benchmark its performance and compare it to the previous generation's A100 GPU. Today, thanks to the benchmarks of MosaicML, a startup company led by the ex-CEO of Nervana and GM of Artificial Intelligence (AI) at Intel, Naveen Rao, we have some comparison between these two GPUs with a fascinating insight about the cost factor. Firstly, MosaicML has taken Generative Pre-trained Transformer (GPT) models of various sizes and trained them using bfloat16 and FP8 Floating Point precision formats. All training occurred on CoreWeave cloud GPU instances.

Regarding performance, the NVIDIA H100 GPU achieved anywhere from 2.2x to 3.3x speedup. However, an interesting finding emerges when comparing the cost of running these GPUs in the cloud. CoreWeave prices the H100 SXM GPUs at $4.76/hr/GPU, while the A100 80 GB SXM gets $2.21/hr/GPU pricing. While the H100 is 2.2x more expensive, the performance makes it up, resulting in less time to train a model and a lower price for the training process. This inherently makes H100 more attractive for researchers and companies wanting to train Large Language Models (LLMs) and makes choosing the newer GPU more viable, despite the increased cost. Below, you can see tables of comparison between two GPUs in training time, speedup, and cost of training.

ChatGPT has surged in popularity over a few months, and usage of this software has been regarded as one of the fastest-growing apps ever. Based on a Large Language Model (LLM) called GPT-3.5/4, ChatGPT uses user input to form answers based on its extensive database used in the training process. Having billions of parameters, the GPT models used for GPT can give precise answers; however, sometimes, these models hallucinate. Given a question about a non-existing topic/subject, ChatGPT can induce hallucination and make up the information. To prevent these hallucinations, NVIDIA, the maker of GPUs used for training and inferencing LLMs, has released a software library to put AI in place, called NeMo Guardrails.

As the NVIDIA repository states: "NeMo Guardrails is an open-source toolkit for easily adding programmable guardrails to LLM-based conversational systems. Guardrails (or "rails" for short) are specific ways of controlling the output of a large language model, such as not talking about politics, responding in a particular way to specific user requests, following a predefined dialog path, using a particular language style, extracting structured data, and more." These guardrails are easily programmable and can stop LLMs from outputting unwanted content. For a company that invests heavily in the hardware and software landscape, this launch is a logical decision to keep the lead in setting the infrastructure for future LLM-based applications.