Nvidia’s new Spectrum-4 network platform offers 51.2Tbps of switching throughput, four times more than previous generations.
It consists of Spectrum-4 switch family, the ConnectX-7 SmartNIC, the BlueField-3 DPU, and the DOCA Datacenter infrastructure software.
Together, they allow nanosecond (rather than millisecond) timing precision, twice the per-port bandwidth, a reduction in the number of switches required, and 40% less power consumption compared to the previous generation.
|
“A new era of massive cloud technologies, like Omniverse, requires a transformation of the data center architecture,” said Nvidia Vice President of the network Kevin Deierling.
“The extreme performance and robust security of the Spectrum-4 platform will equip data centers to achieve discoveries that push the boundaries of what is possible for the benefit of society.”
The Nvidia Spectrum-4 ASIC and SN5000 switch family uses more than 100 billion transistors to deliver 51.2Tbps aggregated ASIC bandwidth, enabling 128 400GbE ports, with adaptive routing and enhanced congestion control mechanisms to optimize RDMA over converged Ethernet fabrics.
Security features include support for MACsec, VXLANsec, and secure boot as standard over hardware root of trust, but they deliver 12.8Tbps encrypted bandwidth.
BlueField DPUs can run and accelerate Red Hat OpenShift for complex containerized workloads, and are a core feature of VMware Project Monterey (currently in closed beta), which aims to improve the performance, maneuverability, and security of data centers .
The BlueField-3 DPU and Spectrum-4 switching systems will be available later this year. ConnectX-7 is immediately available.
Other GTC announcements include:
DGX H100
The fourth generation Nvidia DGX system is the first based on Nvidia H100 Tensor Core GPUs.
With eight GPUs per system, the DGX H100 offers 32 petaflops of AI performance with new FP8 precision – six times that of its predecessor.
It also includes two BlueField-3 DPUs for downloading and accelerating network, storage and security services, plus eight ConnectX-7 Quantum-2 InfiniBand network adapters that offer 400Gbps throughput.
DGX H100 is used in the next generation DGX Pod and SuperPOD. Thanks to the new NVLink Switch system, the latter can connect 32 nodes for a total of 256 GPUs and 1 Exaflop of FP8 performance.
Owners of existing DGX systems can upgrade to DGX H100 through the new DGX-Ready Lifecycle Management Program.
“AI has fundamentally changed what software can do and how it is produced. Companies that are revolutionizing their industries with AI are realizing the importance of their AI infrastructure,” said Nvidia founder and CEO Jensen Huang.
“Our new DGX H100 systems will enable enterprise AI factories to refine data and our most valuable resource – intelligence.”
Nvidia will build the first H100-based SuperPOD computer – called the Eos – with 4608 GPUs expected to deliver 18.4 exaflops of AI computing power. That’s four times faster than the fastest in the world, the Japanese Fugaku,
Grace CPU
Designed for data center use, Grace is an Arm Neoverse CPU for AI infrastructure and high-performance computing.
It consists of two CPUs connected to the new NVLink-C2C chip-to-chip and chip-to-die high-speed, low-latency interconnect, which offer 144 arm cores in a single socket, with an estimated performance of 1.5 times that of the Dual CPU used in current DGX A100 systems.
Grace will run all Nvidia computing software stacks, including RTX, HPC, AI, and Omniverse.
Grace-based servers will be configurable as CPU-only systems or with one, two, four or eight hopper-based GPUs.
Hopper is designed to accelerate dynamic programming, a technique used in algorithms for genomics, quantum computing, route optimization and other applications. Even SQL queries with multiple join operations can be significantly accelerated by applying dynamic programming.
“A new kind of data center has emerged – AI factories that process and refine data citizens to produce intelligence,” Huang said.
“The Grace CPU Superchip offers the highest performance, memory bandwidth and NVIDIA software platforms in a single chip and will shine as the CPU of the world’s AI infrastructure.”
OVX
OVX is a system specifically designed to operate complex digital twin simulations within Nvidia’s omniverse simulation and 3D design collaboration platform.
An OVX server consists of eight Nvidia A40 GPUs, three Nvidia ConnectX-6 Dx 200Gbps NICs, 1TB system memory and 16TB NVMe storage.
The system scales from a single pod of eight OVX servers to a 32-server OVX SuperPOD connected to Nvidia Spectrum-3 switch, and several OVX SuperPODs can be combined for massive digital twin simulations.
“Physically correct digital twins are the future of how we design and build,” said Nvidia Vice President of Professional Visualization Bob Pette.
“Digital twins will change as every industry and company plans. The OVX portfolio of systems will be able to deliver real-time, real-time, ever-synchronized, industrial-scale digital twins across the industry.”
OVX servers from Inspur, Lenovo and Supermicro will be available later in 2022.
AI Platform
Multiple components of Nvidia’s free AI platform have been updated. Riva Speech AI and Merlin recommendation software are now widely available, and new versions of Triton (Inferencing), NeMo (large language models), Maxine (audio and video enhancement) and Tao Toolkit (Custom tuning of Riva models) have been released.
Nvidia AI Enterprise 2.0 is optimized, certified, and supported across all major data centers and cloud platforms, including bare-metal servers, virtualized infrastructure, and CPU-only systems, and is now supported on Red Hat OpenShift and VMware vSphere with Tanzu.
Cuda-X
Similarly, the Cuda-X libraries, tools, and technologies have been extensively updated to improve performance and extend.
For example, Sionna is a new, GPU-accelerated, open-source library for 6G physical layer research, while the Rapids Accelerator for Apache Spark accelerates Rapids data science processing by a factor of three without code changes.
“Innovation in AI and accelerated computing is leading to major scientific breakthroughs in the creation of new applications and services in virtually every industry,” said Nvidia Vice President of Developer Programs Greg Estes.
“With these updates, NVIDIA makes it easier than ever for researchers and developers to take advantage of the power of CUDA and get the highest performance from our platforms.”
Fuert Orin
Nvidia’s Drive Orin autonomous vehicle computer has gone into production, and will be used by my more than 25 vehicle manufacturers, including new customers BYD and Lucid Motors.
The company has also announced the next generation of Drive Hyperion architecture for vehicles that will begin shipping in 2026. The update increases performance for processing sensor data and extending full-range operating domains. Drive Hyperion 9 will have 14 cameras, nine radars, three lidars and 20 ultrasounds as part of its sensor suite.
Jetson AGX Orin Developer Kit
Designed for use in robotic, autonomous machines, next-generation embedded and edge computing, Jetson AGX Orin has more than eight times the processing power of its former Jetson AGX Xavier, but it has the same form factor, is pin-compatible and cost-effective. the approximately the same ($ 1999).
“As AI transforms manufacturing, healthcare, commerce, transportation, smart cities and other key sectors of the economy, demand for processing continues to rise,” said Nvidia Vice President of Embedded and Edge Computing Deepu Talla.
“One million developers and more than 6,000 companies have already turned to Jetson. The availability of Jetson AGX Orin will surpass the efforts of the entire industry as it builds the next generation of robotics and edge AI products.”
Production modules will be available in 4Q22 starting at US $ 399.
Add Comment