NVIDIA：从$40,000 创业到$4 万亿市值的 AI 帝国
1993 年，年仅 30 岁的黄仁勋在硅谷创立 NVIDIA。从$40,000 起步到 2025 年全球首个市值突破$4 万亿和$5 万亿的公司，NVIDIA 的崛起之路是技术押注、标准定义与生态建设的完美组合。
GPU 产品线概览
NVIDIA 的 GPU 产品分为三大系列：
- GeForce RTX – 消费级显卡（RTX 30/40/50 系列），面向游戏和创作者
- 数据中心组（Data Center） – A100/H100/B200/Rubin，用于 AI 训练、HPC 和云计算
- 工作站组（Pro/Quadro） – RTX PRO 6000/Ampere Ada 系列，面向专业工作流
消费级 GeForce RTX GPU 完整型号清单
按架构演进路线整理的消费者显卡产品线
Pascal (GF10x/11x, TSMC 28nm, 2015-2016)
- GTX 970/980 – High-end gaming (Turing-era predecessors)
- GTX 1050 – Entry/mid-range gaming (2GB/3GB GDDR5/GDDR5x PCIe)
- GTX 1050 Ti – Enhanced entry/mid-range (4GB GDDR5/GDDR5x PCIe)
- GTX 1060 – Mid/high-end gaming (3/5/6GB GDDR5 PCIe)
- GTX 1070 – High-end gaming (8GB GDDR5 PCIe)
- GTX 1070 Ti – Flagship Pascal (8GB GDDR5x PCIe)
- GTX 1080 – Ultra-high-end gaming (8GB GDDR5X PCIe)
- GTX 1080 Ti – Pascal flagship (11GB GDDR5X PCIe)
- Titan X (Pascal) – Professional (12GB GDDR5X SXM/PCIe)
- Titan XP – Enhanced Pascal professional (12GB GDDR5X SXM/PCIe)
Volta (GV104, TSMC 12nm, 2017-2018)
- Titan V – Volta professional (32GB HBM2 SXM)
Turing (TU10x/11x, TSMC 12nm, 2018-2019)
- RTX 2060 – Mid/high-end gaming (6GB GDDR6 PCIe)
- RTX 2060 Super – Enhanced mid-range (8GB GDDR6 PCIe)
- RTX 2070 – High-end gaming (8GB GDDR6 PCIe)
- RTX 2070 Super – Flagship Turing high-end (8GB GDDR6 PCIe)
- RTX 2080 – Ultra-high-end gaming (8GB GDDR6 PCIe)
- RTX 2080 Super – Enhanced ultra-high-end (8GB GDDR6 PCIe)
- RTX 2080 Ti – Turing flagship (11GB GDDR6X PCIe)
- Titan RTX – Professional (24GB GDDR6 SXM/PCIe)
Ampere (GA10x/106/107, TSMC 8nm, 2020-2021)
- GTX 1650 – Entry-level gaming (4GB GDDR5 PCIe)
- GTX 1660 – Mid-range gaming (6GB GDDR6 PCIe)
- GTX 1660 Super/Ti – Enhanced mid-range (6/8GB GDDR6 PCIe)
- RTX 3050 – Entry/mid-range gaming (4/6/8GB GDDR6 PCIe)
- RTX 3060 – Mid/high-end gaming (12GB GDDR6 PCIe)
- RTX 3060 Ti – Enhanced mid-range (8GB GDDR6 PCIe)
- RTX 3070 – High-end gaming (8GB GDDR6X PCIe)
- RTX 3070 Ti – Flagship high-end (8GB GDDR6X PCIe)
- RTX 3080 – Ultra-high-end gaming (10GB GDDR6X PCIe)
- RTX 3080 Ti – Enhanced ultra-high-end (12GB GDDR6X PCIe)
- RTX 3090 – Professional/flagship (24GB GDDR6X PCIe)
- Titan RTX Ada – Professional Ampere (48GB GDDR6X SXM)
Ada Lovelace (AD10x, TSMC 5nm, 2022-2024)
- RTX 4060 – Entry-level gaming (8GB GDDR6 PCIe)
- RTX 4060 Ti – Enhanced entry/mid-range (8/16GB GDDR6 PCIe)
- RTX 4070 – Mid/high-end gaming (12GB GDDR6X PCIe)
- RTX 4070 Ti – Enhanced mid-range (12GB GDDR6X PCIe)
- RTX 4070 Ti Super – Flagship mid-high-end (16GB GDDR6X PCIe)
- RTX 4080 – Ultra-high-end gaming (16GB GDDR6X PCIe)
- RTX 4090 – Ada flagship consumer (24GB GDDR6X PCIe)
- Titan RTX Ada – Professional Ada (48GB GDDR6X SXM/PCIe)
Hopper (GH10x, TSMC 4NVM, 2023-至今)
- GeForce RTX 5070/5070 Ti/5080/5090 – Hopper consumer lineup (预计 2024 发布，具体型号待确认)
- 注：Hopper 架构主要面向数据中心 (H100/H200)，消费者产品尚未正式发布
数据中心 GPU 完整型号清单
按架构演进路线整理的数据中心 GPU 产品线
Tesla/Fermi (G9x/XX, TSMC 65/90nm, 2007-2013)
- C870 – 2007 | First CUDA compute card
- S1070 – Server appliance
- C2050/C2070 – Fermi workstation
- M2050/M2070 – Fermi server module
Kepler (GM10x/KS, TSMC 28nm, 2012-2016)
- K20 – Scientific HPC
- K20X – Titan supercomputer GPU
- K40 – 12GB GDDR5 | HPC workhorse
- K80 – 2×12GB GDDR5 | Dual-GPU; dominant early cloud GPU
Maxwell (GM20x/21x, TSMC 28nm, 2014-2016)
- M4 – 4GB GDDR5 | Low-power inference
- M10 – 8GB GDDR5 | VDI/virtual desktop
- M40 – 24GB GDDR5 | Professional workstation
- M60 – 8/12GB GDDR5 ECC | VDI/vGPU (Citrix, VMware)
Pascal (GP10x/PF, TSMC 16nm, 2016-2018)
- P4 – 8GB GDDR5 PCIe | 75W | Low-power inference
- P6 – 16GB GDDR5 MXM | 90W | Server blade
- P40 – 24GB GDDR5 PCIe | 250W | Training/inference
- P100 SXM2 – 2016 | HBM2 | NVLink 3rd gen | First HBM datacenter GPU
- P100 PCIe – 250W
Volta (GV100, TSMC 12nm, 2017-2020)
- V100 SXM2 16GB – 2017 | HBM2 | NVLink 2.0 | DGX-1 GPU
- V100 PCIe 16GB – 250W
- V100 SXM2 32GB – 2018 | HBM2 | DGX-2 only
- V100 PCIe 32GB – 250W
- V100S PCIe 32GB – 2019 | Faster memory variant
Ampere (GA100/A10, TSMC 7nm, 2020)
- A10 – 24GB GDDR6 PCIe | 150W | Inference/graphics | Single-slot
- A16 – 4×16GB GDDR6 PCIe | 250W | VDI/vGPU | Quad-GPU single card
- A30 – 24GB HBM2e PCIe | 165W | Mid-tier HPC/AI
- A40 – 48GB GDDR6 PCIe | 300W | Large-VRAM inference/rendering
- A100 40GB SXM4 – 2020 | HBM2e | 400W | Training flagship
- A100 80GB SXM4 – 2020 | HBM2e | 400W | Extended-memory training
- A100 40GB PCIe – 300W
- A100 80GB PCIe – 2021 | 300W
- A800 80GB SXM – 2022 | China-only export-compliant
- A800 40GB PCIe – 2022 | China-only export-compliant
Turing (TU102/104, TSMC 12nm, 2018-2022)
- T4 – 16GB GDDR6 PCIe | 70W | Cloud inference mainstay (Data Center)
Hopper (GH100, TSMC 4NVM, 2022-至今)
- H100 SXM5 80GB – 2022 | HBM3 | NVLink 4.0 (900 GB/s) | Flagship training
- H100 PCIe 80GB – 2022 | HBM2e | PCIe | Lower bandwidth
- H100 NVL 94GB – 2023 | HBM3 | PCIe | NVLink pair config
- GH200 Grace Hopper – 2023 | 96GB HBM3 + 480GB LPDDR5 | Grace CPU+H100 superchip | 900W
- H800 SXM 80GB – 2023 | China-only export-compliant
- H800 PCIe 80GB – 2023 | China-only export-compliant
- H20 – 2024 | 96GB HBM3 | China-only export-compliant; reduced compute
- H200 SXM 141GB – 2024 | Same GH100 die as H100; ~2× memory
- H200 NVL 141GB – 2024 | NVLink pair; 282 GB combined
Ada Lovelace (AD10x, TSMC 5nm, 2022-2024)
- L4 – Datacenter inference | China-only variants: L2/L40S reduced interconnect
- L40 / L40S – Professional workstation/inference
- RTX PRO 6000 Ada / RTX 6000 Ada – Workstation/professional
Blackwell (GB20x, TSMC 4NVM, 2024-2025)
- B100/B200 SXM – Flagship datacenter training
- GB200/GB300 NVL – Grace Blackwell superchip | GB200 NVL72 (72 GPUs in NVLink domain)
- B300/GB300 – 2025 | Blackwell Ultra; extended to 288GB HBM for heavy inference/training
- RTX PRO 6000 Blackwell – Workstation/professional
GPU 架构演进路线
- Pascal (2015-2017): GTX 9/10, P100, Titan Xp
- Volta (2017): V100, TITAN V – Opened the deep learning era
- Turing (2018-2019): RTX 20, GTX 16, T4 – Introduced Tensor Cores
- Ampere (2020-2021): A100, RTX 30 – Third-gen Tensor Cores, TF32, MIG
- Ada Lovelace (2022-2024): RTX 40, L40S, L4 – Fourth-gen Tensor Cores, FP8
- Hopper (2022-至今): H100, GH200, H200 – Dominant training/HPC architecture
- Blackwell (2024-2025): B200, GB200, RTX 50 – Fifth-gen Tensor Cores, FP4/FP8
- Rubin (2026-): Vera Rubin platform, H2 2026 volume production
市场地位
- 92%离桌面 GPU 市场
- 80%+AI 训练/部署 GPU 市场
- 75%+TOP500 超级计算机芯片供应
- 2023 年第七个市值破$1 万亿美国公司
- 2025 年全球首个市值突破$4 万亿和$5 万亿的公司
