The hardware giant Anvidia Held its first artificial intelligence conference for 2022, GTC, this week, and as part of it unveiled a host of new developments and technologies: a supercomputer for AI calculations, new chips, new architecture for graphics processors and more. Some of the announcements published by the company’s CEO and founder, Jensen Huang, were developed at Anvidia’s development center in Israel, which has about 3,000 employees. Globes is making an order.
Anvidia has announced a supercomputer. What it means?
One of the company’s significant announcements this week is NVIDIA Eos – a supercomputer for artificial intelligence performance. Anvidia boasts that it is the most powerful supercomputer in the world for AI calculations, and is expected to work this year. Today, the most powerful computer is Hitachi’s Fugaku Japanese supercomputer, and Anvidia states that their supercomputer will provide performance of up to 18.4 Exaflops (ten to 18) for artificial intelligence calculations, which is four times that of the Japanese supercomputer. For standard scientific computing performance, the supercomputer will provide up to 275 Petaflops (Ten to the power of 15).
Explain – FLOPS is an acronym for Floating Point Operations Per Second, meaning floating point operations per second, the unit of measurement for computers that deal with heavy processing. Floating point is a method of representing real numbers that on the one hand are intended for accurate registration but on the other hand are numbers that are too large and need to be considered in the limited space for them.
According to the company, its new supercomputer will be based on Anvidia’s new Hopper architecture, and will contain 576 new DGX H100 systems – using the new H100 graphics processor (below). Each DGX H100 system has eight H100 processors, making up a total of 4,608 graphics processors in the overall system.
Photo: Courtesy of Anvidia
The supercomputer will also integrate the high-speed communication technology developed in the Israeli center. The system incorporates 1,152 NVIDIA BlueField-3 DPU chips – information processing units designed to handle huge data – and their job is to reduce workloads from the central processor and thus speed up processing and store the data securely and quickly. The supercomputer will incorporate 500 Quantum-2 NDR switches that will allow data transfer rates of up to 400GB per second. This computer will drive the work of Anvidia researchers working in the fields of climate research, digital biology and the future of artificial intelligence.
What is a new architecture?
Another announcement presented this week is a new architecture called NVIDIA Hopper. Anvidia says that in order to successfully power its new supercomputer, as well as all artificial intelligence data processing centers, it needs improved graphics processor architecture. Let us be clear – this is actually a design of the executive structure of the processor. The new architecture is named after Grace Hopper, a pioneer of computer scientists in the United States.
The new Hopper architecture will serve as the basis for NVIDIA’s new NVIDIA H100 graphics center (GPU) processor. According to the company’s announcement, the new processor includes 80 billion transistors and is unnecessary in the 4nm procedure of the TSMC company. Anvidia boasts of the power of the new processor, saying that 20 H100 processors can provide processing power equivalent to the entire global Internet traffic. In other words – this is the most powerful video card in the world today.
Anvidia is already revealing the partnership with its H100-based servers – companies like Cisco, Dell, Gigabyte, Lenovo and more, alongside support from various cloud providers like Microsoft, AWS, Oracle and more. The new processor will be available starting in the third quarter of this year.
Alongside this, the company also introduced the fourth generation of DGX systems, the NVIDIA DGX H100 supercomputer systems, based on the graphics processor in question. Each of these systems features 8 H100 processors connected to each other via an NVIDIA NVLink connection with 32 pethalops AI performance. According to the company, this is six times more than the previous generation. The system can transfer 400GB per second for communication with the computer and storage units, double the speed of the previous generation, and 900GB per second for communication between all the graphics processors in the system. The combination of components yields 70 terabytes per second of bandwidth – 11 times the previous generation.
How does Anvidia’s new super processor look like Apple’s new M1 Ultra?
The announcement that undoubtedly comes in the background of Apple’s latest announcement, is a new processor (CPU) for data centers and accelerated computing called the NVIDIA Grace CPU Superchip. The new processor, the first based on Arm architecture, contains 144 Arm cores and 1 terabyte per second of memory – which provides better performance, meaning higher bandwidth and low latency. The new processor consists of two CPU chips connected together using NVLinke-C2C technology. This idea was introduced earlier this month by Apple, when it connected two M1 Max processors to one large Ultra processor.
Photo: Courtesy of Anvidia
The chip’s energy consumption will be only 500 watts (including memory), and will be based on the Arm v9 data center architecture. According to the company, the new processor, which will be available at the beginning of next year, will run the entire software package of Anvidia, including its Omniverse (Metavors, virtual and augmented reality, of the company Anvidia).
What has Anvidia announced to the worlds of the Meta-Verse?
One of the terms that has come up more and more in recent months, mainly thanks to the ambitious announcement of a meta-company (formerly Facebook), is Meta-Verse. These are the virtual worlds – augmented reality, virtual reality and augmented reality. Anvidia has the Omni-Verse – which is its infrastructure for the virtual world, in other words: simulations in virtual reality for different uses. In fact according to the use required by the company or organization, Anvidia provides the infrastructure for this realization.
This week Anvidia announced the NVIDIA Omniverse Cloud service – cloud services that will allow creators and artists to collaborate on its Omni-Versus platforms. For example, among the various tools is the Nucleus Cloud – a tool that allows creators to access and edit large-scale 3D scenes from anywhere, without having to move large sets of data.
Photo: Courtesy of Anvidia
Alongside this, Anvidia has announced NVIDIA OVX – a large-scale digital twin computing system. The intention is to create a digital parallel model of an object, product, location or anything else in the virtual world. According to the company, this will make it possible to build models of entire buildings, factories, cities and even digital coordination of the earth to examine different verticals and properties before establishing them in the physical world.
Anvidia has actually announced the NVIDIA Scientific Modulus for Digital Twins, a platform for creating physically accurate digital twins for scientific research. This way you can create real-time AI-based simulations that accurately reflect what is happening in reality. Thus, for example, scientists will be applied to examine extreme weather events.
Bottom line, what’s the good news here?
At the conceptual level, Anvidia’s announcements this week amount to two interesting stories: The first – a supercomputer, which according to the company is the most powerful in the world for using AI calculations, and is flooded with Israeli technology – hundreds of chips developed here in Israel, DPU coming from here and a variety of other components. The second – aligning a particular line to develop a “super-chip”, which is basically a combination of chips using connection technology, similar to what Apple introduced earlier this month. Anvidia presents here improved capabilities that can affect the entire industry, precisely because of its market power and the widespread use of its processors.