---
title: "F2AI Focus | 36 Must-Know Takeaways from NVIDIA GTC 2024"
author: "Unique Research"
sourcePublication: "Unique Research (FFCAP English library)"
originalPublishedAt: "2024-03-22T16:30:00+08:00"
canonical: "https://ffcap.cn/en/research/20240322a06qs300"
source: "https://ffcap.cn/en/research/20240322a06qs300"
language: "en"
cnSourceUrl: "https://view.inews.qq.com/a/20240322A06QS300"
---

# F2AI Focus | 36 Must-Know Takeaways from NVIDIA GTC 2024

_Unique Research · F2AI Focus · March 22, 2024_

NVIDIA’s GTC 2024 — the “Oscars of AI” — just wrapped. Jensen Huang headlined the ~2-hour opening keynote (Tue ~04:00–06:00 Beijing) and an AI-transformation panel (Thu ~02:00–02:50). Unique Research’s 36 must-know points:

## A. Frontier AI tech & products

1.  **Blackwell AI superchip** — Hopper successor; ~2× Hopper; platform scales to **576** GPUs in less space. Industry endorsements in the launch release: Sundar Pichai, Demis Hassabis, Mark Zuckerberg, Satya Nadella, Sam Altman, Elon Musk. Huang: “the entire industry is preparing for Blackwell” — most successful launch in company history.
2.  **DGX SuperPOD** on Grace Blackwell — “AI learning factories”; **11.5 exaFLOPS** AI at FP4 + **240 TB** fast memory for trillion-param train/infer.
3.  **NIM** (NVIDIA Inference Microservice) — cloud-downloadable microservices to assemble AI software with little/no hand coding.
4.  **Project GR00T** — humanoid/robot foundation stack: learn & imitate human motion/speech; AV, factory workers, talking food kiosks, etc.
5.  **Next-gen DC digital blueprint** — GB200 NVL72 liquid-cooled racks: **18** Grace CPUs + **36** Blackwell GPUs per rack, 4th-gen NVLink switches; sim + digital twin for cooling/design validation before build.
6.  **Omniverse Cloud as APIs** — five new APIs to embed Omniverse into digital-twin / robot & AV validation workflows; warehouse demo with multi-cam perception and autonomous forklifts.

## B. Industry applications

7.  **Telecom — 6G Research Cloud** — AI + Omniverse; Sionna neural radio, Aerial CUDA RAN, Aerial Omniverse Digital Twin for 6G.
8.  **Semiconductors — cuLitho** with TSMC & Synopsys — computational lithography workloads **40–60×** faster.
9.  **Earth climate twin cloud** — interactive hi-res climate/weather; The Weather Company + Taiwan met agency among first Earth-2 Cloud API users; ~**2 km** scale breakthrough cited.
10.  **Healthcare** — imaging, genomics instruments, surgical robotics partners; new biology software + **20+** microservices (imaging, NLP/speech, digital biology gen/predict/sim).

## C. Super-partner moves

11.  **SAP × NVIDIA** — custom LLMs via AI Foundry + NIM on SAP cloud data; Business AI integrations targeted by end-2024. Huang: SAP sits on an enterprise data goldmine for custom genAI agents.
12.  **Oracle × NVIDIA — sovereign AI** — national infra/data/talent/networks to produce AI; Oracle distributed cloud + NVIDIA accel compute/genAI software for “AI factories.”
13.  **Omniverse → Apple Vision Pro** — stream interactive industrial digital twins to the headset; adopters include Ansys, Cadence, Dassault 3DEXCITE, Hexagon, Microsoft, Rockwell, Siemens, Trimble.
14.  **BYD × DRIVE Thor** — next-gen EV fleet on Thor with Blackwell for Transformer/genAI ADAS + IVI workloads.

## D. Soundbites — Jensen Huang (selected)

15.  Mature enterprise platforms sit on data goldmines → genAI copilots.
16.  Next AI wave: learning in the **physical world**.
17.  Need larger multimodal models — text, images, graphs, charts, and massive video (like learning from TV).
18.  Data centers become **AI factories** — mission: revenue + intelligence.
19.  Future work: assemble AI models, assign tasks, review plans — not write software from scratch.
20.  Everything that moves can be a robot; NVIDIA already in cars, trucks, delivery bots, robotaxis.
21.  Graphics, physics, AI are NVIDIA’s soul; genAI changes everything.
22.  Accelerated computing at inflection; general-purpose computing lost momentum.
23.  Need a sim engine gym for robots — Omniverse.
24.  Building software: less greenfield Python; more assembling AI teams.

## E. Soundbites — Fei-Fei Li (fireside with Bill Dally)

25.  Blind big-data scale sells chips; she wants structured / 3D-perception-biased modeling + big data.
26.  Nature took ~540M years for perceiving brains vs ~hundreds of thousands for language — perception is deep.
27.  Task-jobs (grab object, fry egg) may be automated; humanity/creativity/empathy/unique contribution won’t be fully replaced — machines amplify.
28.  Human interaction exceeds compute/mechanics — those jobs evolve with machine superpowers.
29.  You cannot train human intelligence or creativity into machines by training alone.
30.  AI is a tool; humans are creators, decision-makers, voters on governance — many civic roles in AI.

## F. Transformer paper authors reunion (7 of 8 with Huang)

Present: Noam Shazeer (Character.ai), Aidan Gomez (Cohere), Ashish Vaswani (Essential AI), Llion Jones (Sakana), Illia Polosukhin (NEAR), Jakob Uszkoreit (Inceptive), Lukasz Kaiser (OpenAI). Niki Parmar (Essential) absent.

31.  Name “Transformer” was somewhat arbitrary — captures converting the whole signal; ML as disruption.
32.  Adaptive compute next — know how much compute a problem deserves (2+2 shouldn’t burn a trillion params).
33.  Smarter world = multipath feedback from the world; multi-task parallel.
34.  Learning always wins — generative AI should help understand the universe.
35.  Teach machines to code so programming is accessible; embrace the generative era.
36.  Reasoning matters; much is still hand-engineered — interactive Q&A toward strong reasoning; unify video/text/3D generation.

_Ref:_ Tencent Tech Chinese transcript of Huang × Transformer authors. Column intro omitted.

_15 CN images; attach on ingest._

— END —

---

Editorial English edition published on UniqueCapital: https://ffcap.cn/en/research/20240322a06qs300
On-site reading page: https://ffcap.cn/en/research/20240322a06qs300