Note / GPU
The NVIDIA GPU Product Lineage
A map from Fermi to Rubin that separates consumer, professional, and data-center GPUs from NVIDIA's system-level products.
NVIDIA names are easy to confuse because GPU architectures, board products, market brands, modules, superchips, and rack-scale systems often appear in the same conversation. This guide separates discrete GPUs, embedded SoCs and modules, CPU-GPU superchips, and server or rack systems.
The main chronology follows one strict rule: products share a row only when they use the same GPU architecture.
1. Start with the naming layers
| Layer | Typical names | Intended use |
|---|---|---|
| Consumer GPU brand | GeForce, GeForce GTX, GeForce RTX | Gaming, personal creation, and personal AI computing |
| Flagship consumer / developer sub-line | TITAN | An ultra-high-end line from 2013-2018; part of its market position later moved to GeForce RTX 90-class products |
| Professional GPU brand | Quadro, Quadro RTX, RTX A, RTX Ada, RTX PRO | CAD, DCC, professional visualization, enterprise AI, and workstations |
| Data-center GPU brand | Tesla, A series, H series, B series, Rubin GPU | AI training, inference, HPC, and data-center acceleration |
RTX is not a single brand at the same level as GeForce, Quadro, or Tesla. GeForce RTX is a consumer line. Quadro RTX, RTX A, RTX Ada, and RTX PRO are professional lines.
2. Discrete GPU evolution since 2010
This table contains only discrete GPU or accelerator-card products. GH200, GB200, GB300, NVL72, DGX, TITAN, and similar system or superchip names are intentionally excluded.
| Years / architecture | Consumer GeForce | Professional Quadro / NVIDIA RTX | Tesla / Data Center GPU |
|---|---|---|---|
| 2010-2011 · Fermi | GeForce GTX 480 GeForce GTX 580 | Quadro 4000 Quadro 5000 Quadro 6000 | Tesla C2050 Tesla C2070 Tesla M2090 |
| 2012-2014 · Kepler | GeForce GTX 680 GeForce GTX 780 Ti | Quadro K5000 Quadro K6000 | Tesla K10 Tesla K20 Tesla K40 |
| 2014-2016 · Maxwell | GeForce GTX 980 GeForce GTX 980 Ti | Quadro M4000 Quadro M5000 Quadro M6000 | Tesla M4 Tesla M40 Tesla M60 |
| 2016-2017 · Pascal | GeForce GTX 1080 GeForce GTX 1080 Ti | Quadro P5000 Quadro P6000 | Tesla P4 Tesla P40 Tesla P100 |
| 2017-2018 · Volta | - | Quadro GV100 | Tesla V100 |
| 2018-2020 · Turing | GeForce RTX 2060 GeForce RTX 2080 Ti | Quadro RTX 4000 Quadro RTX 5000 Quadro RTX 6000 Quadro RTX 8000 | NVIDIA Tesla T4 |
| 2020-2022 · Ampere | GeForce RTX 3060 GeForce RTX 3080 GeForce RTX 3090 GeForce RTX 3090 Ti | NVIDIA RTX A2000 RTX A4000 RTX A4500 RTX A5000 RTX A5500 RTX A6000 | NVIDIA A2 A10 A16 A30 A40 A100 |
| 2022-2024 · Hopper | - | - | NVIDIA H100 NVIDIA H200 |
| 2022-2024 · Ada Lovelace | GeForce RTX 4060 GeForce RTX 4070 GeForce RTX 4080 GeForce RTX 4090 | NVIDIA RTX 2000 Ada RTX 4000 Ada RTX 4500 Ada RTX 5000 Ada RTX 6000 Ada | NVIDIA L4 L40 L40S |
| 2024-2025 · Blackwell | - | - | NVIDIA B100 NVIDIA B200 |
| 2025-2026 · Blackwell | GeForce RTX 5050 GeForce RTX 5060 GeForce RTX 5060 Ti GeForce RTX 5070 GeForce RTX 5070 Ti GeForce RTX 5080 GeForce RTX 5090 | NVIDIA RTX PRO 2000 Blackwell RTX PRO 4000 Blackwell RTX PRO 4500 Blackwell RTX PRO 5000 Blackwell RTX PRO 6000 Blackwell Workstation Edition RTX PRO 6000 Blackwell Max-Q Workstation Edition | RTX PRO 6000 Blackwell Server Edition |
| 2025-2026 · Blackwell Ultra | - | - | NVIDIA B300 |
| 2026 · Rubin | - | - | NVIDIA Rubin GPU NVIDIA Rubin CPX |
3. The shortest useful family tree
Early graphics
NV1 -> RIVA
|
|- GeForce -> GTX -> GeForce RTX -> RTX 90 class
| \- TITAN --------------------^
|
|- Quadro -> Quadro RTX -> RTX A -> RTX Ada -> RTX PRO
|
\- CUDA -> Tesla -> A100 -> H100/H200 -> B100/B200
-> B300 -> Rubin GPU
Mobile and embedded
Tegra -> Jetson / DRIVE / SHIELD
Cloud and virtualization
GRID -> NVIDIA vGPU -> RTX Virtual Workstation
AI systems (a parallel system layer, not a GPU generation)
Data Center GPU --combined into--> DGX / HGX / MGX / NVL
Data-center networking
Mellanox -> ConnectX / BlueField / Spectrum-X / Quantum
4. Ampere-and-newer PCIe GPUs for AI training
This table includes major NVIDIA products that use a standard PCIe x16 add-in-card form factor, have strictly more than 16 GB of memory per GPU, support CUDA and Tensor Cores, and can be used for AI training or fine-tuning.
The compute column uses theoretical FP32 peak throughput as a consistent cross-line reference. Real training performance also depends on BF16, FP16, TF32, and FP8 Tensor Core throughput, memory bandwidth, software support, and workload shape.
| Product | Memory | FP32 peak | PCIe |
|---|---|---|---|
| GeForce RTX 3090 | 24GB GDDR6X | ≈35.6 TFLOPS | PCIe 4.0 x16 |
| GeForce RTX 3090 Ti | 24GB GDDR6X | ≈40.0 TFLOPS | PCIe 4.0 x16 |
| NVIDIA RTX A4500 | 20GB GDDR6 ECC | 23.7 TFLOPS | PCIe 4.0 x16 |
| NVIDIA RTX A5000 | 24GB GDDR6 ECC | 27.8 TFLOPS | PCIe 4.0 x16 |
| NVIDIA RTX A5500 | 24GB GDDR6 ECC | 34.1 TFLOPS | PCIe 4.0 x16 |
| NVIDIA RTX A6000 | 48GB GDDR6 ECC | 38.7 TFLOPS | PCIe 4.0 x16 |
| NVIDIA A10 | 24GB GDDR6 | 31.2 TFLOPS | PCIe 4.0 x16 |
| NVIDIA A30 | 24GB HBM2 | 10.3 TFLOPS | PCIe 4.0 x16 |
| NVIDIA A40 | 48GB GDDR6 ECC | 37.4 TFLOPS | PCIe 4.0 x16 |
| NVIDIA A100 PCIe | 40GB HBM2 | 19.5 TFLOPS | PCIe 4.0 x16 |
| NVIDIA A100 80GB PCIe | 80GB HBM2e | 19.5 TFLOPS | PCIe 4.0 x16 |
| GeForce RTX 4090 | 24GB GDDR6X | 83.0 TFLOPS | PCIe 4.0 x16 |
| NVIDIA RTX 4000 SFF Ada Generation | 20GB GDDR6 ECC | 19.2 TFLOPS | PCIe 4.0 x16 |
| NVIDIA RTX 4000 Ada Generation | 20GB GDDR6 ECC | 26.7 TFLOPS | PCIe 4.0 x16 |
| NVIDIA RTX 4500 Ada Generation | 24GB GDDR6 ECC | 39.6 TFLOPS | PCIe 4.0 x16 |
| NVIDIA RTX 5000 Ada Generation | 32GB GDDR6 ECC | 65.3 TFLOPS | PCIe 4.0 x16 |
| NVIDIA RTX 6000 Ada Generation | 48GB GDDR6 ECC | 91.1 TFLOPS | PCIe 4.0 x16 |
| NVIDIA L4 | 24GB GDDR6 | 30.3 TFLOPS | PCIe 4.0 x16 |
| NVIDIA L40 | 48GB GDDR6 ECC | ≈90.5 TFLOPS | PCIe 4.0 x16 |
| NVIDIA L40S | 48GB GDDR6 ECC | 91.6 TFLOPS | PCIe 4.0 x16 |
| NVIDIA H100 PCIe | 80GB HBM2e | 51.0 TFLOPS | PCIe 5.0 x16 |
| NVIDIA H100 NVL | 94GB HBM3 | 60.0 TFLOPS | PCIe 5.0 x16 |
| NVIDIA H200 NVL | 141GB HBM3e | 60.0 TFLOPS | PCIe 5.0 x16 |
| GeForce RTX 5090 | 32GB GDDR7 | ≈104.9 TFLOPS | PCIe 5.0 x16 |
| NVIDIA RTX PRO 4000 Blackwell SFF Edition | 24GB GDDR7 ECC | 24.0 TFLOPS | PCIe 5.0 x16 |
| NVIDIA RTX PRO 4000 Blackwell | 24GB GDDR7 ECC | 40.0 TFLOPS | PCIe 5.0 x16 |
| NVIDIA RTX PRO 4500 Blackwell Server Edition | 32GB GDDR7 ECC | 51.0 TFLOPS | PCIe 5.0 x16 |
| NVIDIA RTX PRO 5000 Blackwell | 48GB GDDR7 ECC | 70.0 TFLOPS | PCIe 5.0 x16 |
| NVIDIA RTX PRO 5000 72GB Blackwell | 72GB GDDR7 ECC | 70.0 TFLOPS | PCIe 5.0 x16 |
| NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition | 96GB GDDR7 ECC | 110 TFLOPS | PCIe 5.0 x16 |
| NVIDIA RTX PRO 6000 Blackwell Workstation Edition | 96GB GDDR7 ECC | 125 TFLOPS | PCIe 5.0 x16 |
| NVIDIA RTX PRO 6000 Blackwell Server Edition | 96GB GDDR7 ECC | 120 TFLOPS | PCIe 5.0 x16 |
How to read the list
Architecture, brand, board form factor, and system product answer different questions:
- the architecture identifies a generation of GPU design;
- the product line defines a market, driver, certification, and support boundary;
- the board form factor determines host compatibility, power, cooling, and interconnect constraints;
- a system product combines GPUs with CPUs, interconnects, and networking into a node or rack.
FP32 peak throughput is only a common comparison point. For training and inference, memory capacity, memory bandwidth, low-precision Tensor Core throughput, interconnect, power, and software support often determine practical performance more directly.