Tesla first revealed plans for its Dojo D1 training chip back in 2021, with hopes of it powering self-driving technology in the near future. The automative division has relied mostly on NVIDIA over the ensuing years, but is seemingly keen to move onto proprietary solutions. Media reports from two years ago suggest that 5760 NVIDIA A100 GPUs were in play to develop Tesla's advanced driver-assistance system (Autopilot ADAS). Tom's Hardware believed that a $300 Million AI supercomputer cluster—comprised of roughly 10,000 NVIDIA H100 GPUs—was powered on last month. Recent reports emerging from Taiwan suggest that Tesla is doubling Dojo D1 supercomputer chip orders with TSMC.

An Economic Daily report posits that 10,000 Dojo D1 are in a production queue for the next year, with insiders believing that Tesla is quietly expressing confidence in its custom application-specific integrated circuit (ASIC). An upcoming order count could increase for the next batch (in 2025). The article hints that TSMC's "HPC-related order momentum has increased thanks to Tesla." Both organizations have not publicly commented on these developments, but insider sources have disclosed some technical details—most notably that the finalized Dojo design: "mainly uses TSMC's 7 nm family process and combines it with InFO-level system-on-wafer (SoW) advanced packaging."

Google made a small but important change to their privacy policy over the weekend that effectively lays claim to anything you post publicly online for use to train their AI models. The original wording of the section of their privacy policy claimed that public data would be used for business purposes, research, and for improving Google Translate services. Now however the section has been updated to read the following:

Google uses information to improve our services and to develop new products, features and technologies that benefit our users and the public. For example, we use publicly available information to help train Google's AI models and build products and features like Google Translate, Bard, and Cloud AI capabilities.

Further down in the policy text Google has another section which exemplifies the areas of "publicly available" information they seek to scrape,

For example, we may collect information that's publicly available online or from other public sources to help train Google's AI models and build products and features like Google Translate, Bard, and Cloud AI capabilities. Or, if your business's information appears on a website, we may index and display it on Google services.

Thermaltake, the leading PC DIY premium brand for Case, Power, Cooling, Gaming peripherals, and enthusiast Memory solutions, is excited to hold the Thermaltake Indoor Cycling Esports Tournament at COMPUTEX 2023. In recent years, cycling esports has been gaining traction in the esports industry, and the CYCLEDESK 100 Smart Gaming Desk is a multi-functional desk designed for gaming, working, and professional cycling esports. At COMPUTEX 2023 you can not only watch the Thermaltake Indoor Cycling Esports Tournament, but also experience the CYCLEDESK 100 and other cycling esports gaming gear firsthand.

The CYCLEDESK 100 Smart Gaming Desk is an essential part of any cycling esports competition. Standing out from most desks on the market, the CYCLEDESK 100's small footprint and mobility makes it perfect for a variety of usage scenarios; from cycling esports, to gaming stations, and workplace scenarios. The CYCLEDESK 100's multi-functional design provides ample space for a PC monitor or laptop setup, allowing users to comfortably watch videos while cycling at the same time. Empowered by the TT Smart Control Unit, the CYCELDESK 100 comes with a smart height-adjustment function that can be controlled through your mobile devices via WiFi or PC. It also features an interface which displays the current desk height, and offers easy height-adjustments through a push of a button located at the front panel. The CYCLEDESK 100 takes smart design to the next level, and is perfect for cycling esports competitions and regular workout sessions.

NVIDIA and VMware are helping professionals elevate extended reality (XR) by streaming from the cloud with Workspace ONE XR Hub, which includes an integration of NVIDIA CloudXR. Now available, Workspace ONE XR Hub enhances the user experience for VR headsets through advanced authentication and customization options. Combined with the cutting-edge streaming capabilities of CloudXR, Workspace ONE XR Hub enables more professionals, design teams and developers to quickly and securely access complex immersive environments using all-in-one headsets.

Many organizations across industries use XR technologies to boost productivity and enhance creativity. However, teams often come across challenges when it comes to integrating XR into their workflows. That's why NVIDIA and VMware are working together to make it easier for enterprises to adopt augmented reality (AR) and virtual reality (VR). Whether it's running immersive training sessions or conducting virtual design reviews, professionals can achieve the highest fidelity experience with greater mobility through Workspace ONE XR Hub and CloudXR.

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.

Google has announced that the company is officially merging its subsidiaries focused on artificial intelligence to form a single group. More specifically, Google Brain and DeepMind companies are now joining forces to become a single unit called Google DeepMind. As Google CEO Sundar Pichai notes: "This group, called Google DeepMind, will bring together two leading research groups in the AI field: the Brain team from Google Research, and DeepMind. Their collective accomplishments in AI over the last decade span AlphaGo, Transformers, word2vec, WaveNet, AlphaFold, sequence to sequence models, distillation, deep reinforcement learning, and distributed systems and software frameworks like TensorFlow and JAX for expressing, training and deploying large scale ML models."

As a CEO of this group, Demis Hassabis, a previous CEO of DeepMind, will work together with Jeff Dean, now promoted to Google's Chief Scientist, where he will report to the Sundar. In the spirit of a new role, Jeff Dean will work as a Chief Scientist at Google Research and Google DeepMind, where he will set the goal for AI research at both units. This corporate restructuring will help the two previously separate teams work together on a single plan and help advance AI capabilities faster. We are eager to see the upcoming developments these teams accomplish.

IBM today unveiled new single frame and rack mount configurations of IBM z16 and IBM LinuxONE 4, expanding their capabilities to a broader range of data center environments. Based on IBM's Telum processor, the new options are designed with sustainability in mind for highly efficient data centers, helping clients adapt to a digitized economy and ongoing global uncertainty.

Introduced in April 2022, the IBM z16 multi frame has helped transform industries with real-time AI inferencing at scale and quantum-safe cryptography. IBM LinuxONE Emperor 4, launched in September 2022, features capabilities that can reduce both energy consumption and data center floor space while delivering the scale, performance and security that clients need. The new single frame and rack mount configurations expand client infrastructure choices and help bring these benefits to data center environments where space, sustainability and standardization are paramount.

We've created GPT-4, the latest milestone in OpenAI's effort in scaling up deep learning. GPT-4 is a large multimodal model (accepting image and text inputs, emitting text outputs) that, while less capable than humans in many real-world scenarios, exhibits human-level performance on various professional and academic benchmarks. For example, it passes a simulated bar exam with a score around the top 10% of test takers; in contrast, GPT-3.5's score was around the bottom 10%. We've spent 6 months iteratively aligning GPT-4 using lessons from our adversarial testing program as well as ChatGPT, resulting in our best-ever results (though far from perfect) on factuality, steerability, and refusing to go outside of guardrails.

Over the past two years, we rebuilt our entire deep learning stack and, together with Azure, co-designed a supercomputer from the ground up for our workload. A year ago, we trained GPT-3.5 as a first "test run" of the system. We found and fixed some bugs and improved our theoretical foundations. As a result, our GPT-4 training run was (for us at least!) unprecedentedly stable, becoming our first large model whose training performance we were able to accurately predict ahead of time. As we continue to focus on reliable scaling, we aim to hone our methodology to help us predict and prepare for future capabilities increasingly far in advance—something we view as critical for safety.