Justin Sullivan / Getty Images News
Monitoring the evolution of the computing industry over the last few years has been a fascinating exercise. After focusing almost exclusively on one type of chip (CPU (Central Processing Unit)) and measuring the enhancements made by improving its internal architecture, it is possible to play on multiple chip types, especially GPU (Graphics Processing Unit). There was a change. High-speed connectivity between components improves performance.
This has never been clearer than Nvidia’s (NASDAQ: NVDA) latest GTC, or GPU technology conference. At the event’s keynote, the company’s CEO, Jensen Huang, announced the latest GPU architecture (named Hopper after computing pioneer Grace Hopper) and various forms of high-speed chip-to-device and device. We have announced many new advances, including inter-connection options. .. Collectively, the company has taken advantage of these key technological advances to offer many other options, from giant Eos supercomputers to the H100 CNX Converged Accelerator, a PCIe card designed for existing servers. I introduced what I had.
Nvidia’s focus is driven by the industry’s constant pursuit of advances in AI and machine learning. In fact, most of the company’s many chip, hardware, and software announcements from the show are with these key trends such as HPC (High Performance Computing) type supercomputing applications, automated driving systems, and embedded robot applications. It is tied.
By the way, Nvidia also emphasized that it’s not just a chip company for the 2022 Spring GTC, it’s offering key software updates for existing tools and platforms, especially Omniverse 3D collaboration and simulation suites. To facilitate further use of the tool, Nvidia has announced Omniverse Cloud. This allows anyone to try Omniverse in just a browser.
For hyperscalers and large enterprises looking to deploy advanced AI applications, the company has several clouds, including Merlin 1.0 for recommended systems and version 2.0 for both Riva speech recognition and text-to-speech. We have also debuted a new or updated version of the native application service. Services and AI enterprises for a variety of data science and analytics applications. New in AI Enterprise 2.0 are virtualization support and the ability to use containers on multiple platforms such as VMware (VMW) and Red Hat. Overall, these products reflect the company’s growth as a software provider. From a tool-focused approach to an approach that provides SaaS-style applications that can be deployed to all major public clouds, Dell Technologies (DELL), Hewlett Packard Enterprise (HPE), and Lenovo (OTCPK: LNVGY). ).
But never forget its roots, Nvidia’s latest GTC has featured a new hopper GPU architecture and a GPU focused on the company’s H100 data center. Boasting a whopping 80 billion transistors, the 4nm process-based Nvidia H100 supports several key architectural advances. First, the company claims that the H100 will offer a six-fold improvement over the previous Ampere architecture to accelerate the performance of new Transformer-based AI models (such as those driving the GPT-3 natural language engine). Includes Transformer engine. It also includes a new instruction set called DPX designed to accelerate dynamic programming. This is a technique used in applications such as genomics and proteomics that previously ran on the CPU or FPGA.
For privacy-conscious applications, the H100 was also the first GPU or accelerator to support sensitive computing (previous implementations only worked on the CPU) and through a virtualized and reliable execution environment. You can encrypt and protect your model and data. This architecture allows for federated learning in sensitive computing modes. This means that multiple companies with private datasets can train the same model by passing the model between essentially different secure environments. In addition, thanks to the second generation implementation of multi-instance GPUs (MIGs), a single physical GPU can be split into seven separate isolated workloads to improve chip efficiency in a shared environment. I can do it.
Finally, Hopper also supports the 4th generation version of NVLink’s NVLink. This is a huge leap in enabling the use of NVLink switches with significantly increased bandwidth by a factor of 9 compared to previous technologies, supporting connections to up to 256 GPUs. The latter provides the ability to maintain high-speed connectivity not only within a single system, but to external systems. This enables a new range of DGX pods and DGX superpods, Nvidia’s own brand of supercomputer hardware, and the aforementioned Eos supercomputers.
When it comes to NVLink and physical connectivity, the company also announced support for a new chip-to-chip technology called Nvidia NVLink-C2C. It is designed for chip-to-chip and die-to-die connections at speeds up to 900 Gbps. Nvidia component. In addition, the company has unveiled its previously proprietary NVLink standard to work with other chip vendors, and in particular announced support for the newly announced UCIe standard (for more information, see Semiconductor. For the future of UCIe “). This gives other companies in the semiconductor industry greater flexibility in how they work with other companies to create dissimilar components as well.
Nvidia has chosen to leverage its NVLink-C2C for the new Grace Superchip. It’s a combination of two Arm-based CPUs previously announced by the company, revealing that the GraceHopper Superchip, previewed last year, uses the same interconnect technology to provide high-speed connectivity. Between that single Grace CPU and Hopper GPU. Both “superchips” are targeted at data center applications, but their architecture and underlying technology are well understood where PCs and other mainstream applications are expected to go. NVLink-C2C standards that support industry connectivity standards such as Arm’s AMBACHI protocol and CXL can also be used to interconnect DPUs (data processing units) to accelerate critical data transfers within and between systems. I can do it.
In addition to all these data center-focused announcements, Nvidia also announced updates to the Drive Orin platform for assisted and autonomous driving, and the Jetson and Isaac Orin platforms for robotics, as well as more realistic customers. did.
Anyway, it was an impressive launch of many technologies, chips, systems, and platforms. What is clear is that the future of demanding AI applications, along with other challenging computing challenges, will require multiple different elements that work together to complete a particular task. As a result, increasing the variety of chip types and the mechanisms that allow them to communicate with each other is just as important as progress within individual categories. More simply, we are heading into a clearly connected multi-chip world.
Disclaimer: Some of the author’s clients are vendors in the technology industry.
Disclosure: none.
sauce: author
Editor’s Note: The bullet points in the summary of this article have been selected by the editors of Seeking Alpha.