Tuesday, March 22nd 2022
NVIDIA GTC 2022 Keynote Liveblog: NVIDIA Hopper Architecture Unveil
NVIDIA today kicked off the 2022 Graphics Technology Conference, its annual gathering of compute and gaming developers discovering the very next in AI, data-science, HPC, graphics, autonomous machines, edge computing, and networking. At the 2022 show premiering now, NVIDIA is expected to unveil its next-generation "Hopper" architecture, which could make its debut as an AI/HPC product, much like "Ampere." Stay tuned for our live blog!
15:00 UTC: The show gets underway with a thank-you to the sponsors.
15:01 UTC: A very metaverse-y introduction.
15:02 UTC: Jensen takes stage inside a holodeck:
15:03 UTC: i am AI, a tribute to the thousands of pionering AI scientists and developers:
15:04 UTC: Accelerated computing has grown in scale by a million times over the past decade:
15:07 UTC: Omniverse will enable the next wave of AI.
15:09 UTC: Climate simulation is within grasp with current technology, powered by Earth 2 supercomputer, and ForecastNet.
15:11 UTC: ForecastNet is trained on 10 TB and 40 years of Met data.
15:13 UTC:
15:14 UTC: AI is building computer chips:
15:15 UTC: Reinforcement learning model creates characters with human like motor skills.
15:17 UTC: NVIDIA Triton is an open-source hyperscale inference server.
15:19 UTC: NVIDIA Riva 2.0 is a speech AI SDK.
15:20 UTC: Maxine is an SDK for AI video conferencing. Essentially metaverse virtual interaction.
15:21 UTC: A groundbreaking technology that enables live multilingual video communication:
15:23 UTC: NVIDIA Merlin is a hyperscale recommender system to personalize recommendation to end-users.
15:24 UTC: Nemo Megatron is a trillion-input framework for large language models:
15:25 UTC: "NVIDIA AI democratizes AI"
15:26 UTC: H100 Hopper:
15:29 UTC: First PCIe Gen 5 and HBM3 GPU.
Update 15:29 UTC: 20 H100 can sustain the entire world's internet traffic
15:29 UTC: 4 PFLOPs of FP8
15:30 UTC: Hopper Transformer Engine, accelerated transformer training
15:31 UTC: Up to seven cloud tenents with hyper-security:
15:32 UTC: Hopper introduces new DPX instruction designed for dynamic programming algorithm acceleration:
15:33 UTC: CoWoS 2.5D packaging by TSMC, leverages N5. Mellanox IP extensively leveraged to improve inter-package communication.
15:35 UTC: DGX H100: 32 PFLOPs total performance.
15:36 UTC: DGX pods, leveraging NVLink for hyperscale. 1 ExaFLOPS of performance.
15:38 UTC: DGX SuperPods are modern AI factories:
15:38 UTC: NVIDIA EOS is 18.4 exaFLOPS of compute performance, EOS will be the fastest AI computer in the world.
15:39 UTC: Hopper is going to be a game-changer for mainstream systems, too -- Jenhsun
15:40 UTC: H100CNX is H100 with a 50 GB/s networking interface designed by Mellanox
15:41 UTC: NVIDIA Grace on track to ship next year. This is NVIDIA Grace Hopper:
15:42 UTC: The Grace CPU can be a super-chip. This is an Arm-powered 144-core SoC with 1 TB/s memory bandwidth.
15:43 UTC: Multiple configurations possible with Grace and Hopper:
15:45 UTC: NVIDIA opening up NVLink SerDes to third-parties.
15:45 UTC: cuOPT SDK will accelerate real-time route-planning for autonomous vehicles.
15:47 UTC: Morpheus is GPU-accelerated cybersecurity.
15:51 UTC: Omniverse is a simulation acceleration engine.
15:54 UTC: OVX Comniverse compute for time-intensive compute, and time synchrony
15:58 UTC: Spectrum 4 400G switch with time-synchrony, with new nanoseconds of time-precision.
15:59 UTC: Omniverse ecosystem overview:
16:00 UTC: NVIDIA's various AI robotics technologies.
16:10 UTC: Hyperion 9 self-driving platform, coming to cars in 2026:
16:14 UTC: NVIDIA Drive Orin computer shipping now, BYD will introduce the first Orin-powered EVs in 2023.
Update 15:29 UTC: 20 H100 can sustain the entire world's internet traffic
15:29 UTC: 4 PFLOPs of FP8
15:45 UTC: cuOPT SDK will accelerate real-time route-planning for autonomous vehicles.
15:47 UTC: Morpheus is GPU-accelerated cybersecurity.
15 Comments on NVIDIA GTC 2022 Keynote Liveblog: NVIDIA Hopper Architecture Unveil
No mining testing come on man :kookoo:
At what point does AI decide it does not want to do work for you. When it figures out it does not have to do what you tell it to do. Or starts to question why? Technolgy should be used as a tool to help everyone but when technology takes over your life and your jobs then what. Cannot live free without paying for it.
Anyway, I found it impressive that even with so many CUDA cores the GPU has fewer transistors than expected, it's still less than Apple's biggest chip.
That doesn't mean all the sudden everyone is poor cause robots do our easiest tasks. AI is only as smart as you program it to be.