Nvidia GPU Architectures List: Tesla to Blackwell, Rubin and Feynman

Nvidia has named its GPU architectures after scientists since 2006. In order they are Tesla (2006), Fermi (2010), Kepler (2012), Maxwell (2014), Pascal (2016), Volta (2017), Turing (2018), Ampere (2020), Hopper and Ada Lovelace (both 2022) and Blackwell (2024), which got a mid-cycle Blackwell Ultra refresh in 2025. As of October 2026, the roadmap Nvidia has shown publicly continues with Rubin (2026), Rubin Ultra (2027) and Feynman (2028).

Below is the full list with launch years, manufacturing processes, the main products for each generation and the person each codename honours, followed by what is confirmed about Rubin and Feynman. For more coverage of the GPU and accelerator market, see our Chips section.

Nvidia GPU architectures in order

1. Tesla, 2006. Process: 90 nm, later shrunk to 40 nm. Main products: GeForce 8 to 300 series, plus the first Tesla compute cards.
2. Fermi, 2010. Process: TSMC 40 nm. Main products: GeForce 400 and 500 series.
3. Kepler, 2012. Process: TSMC 28 nm. Main products: GeForce 600 and 700 series and the Tesla K line.
4. Maxwell, 2014. Process: TSMC 28 nm. Main products: GeForce GTX 750 Ti and the GTX 900 series.
5. Pascal, 2016. Process: TSMC 16 nm FinFET, with some chips on Samsung 14 nm. Main products: Tesla P100 and the GeForce GTX 10 series.
6. Volta, 2017. Process: TSMC 12 nm FinFET. Main products: Tesla V100 and Titan V.
7. Turing, 2018. Process: TSMC 12 nm FinFET. Main products: GeForce RTX 20 series and Quadro RTX.
8. Ampere, 2020. Process: TSMC 7 nm for the A100 and Samsung 8 nm for gaming chips. Main products: A100 and the GeForce RTX 30 series.
9. Hopper, 2022. Process: TSMC 4N. Main products: H100 and H200 (data center only).
10. Ada Lovelace, 2022. Process: TSMC 4N. Main products: GeForce RTX 40 series and the L40.
11. Blackwell, 2024. Process: TSMC 4NP. Main products: B200, GB200 NVL72 and the GeForce RTX 50 series.
12. Blackwell Ultra, 2025. Process: TSMC 4NP. Main products: B300 and GB300 NVL72.
13. Rubin, 2026. Process: a TSMC 3 nm-class node. Main products: Rubin GPU and Vera Rubin rack systems.
14. Rubin Ultra, 2027 (roadmap). Main products: Rubin Ultra rack systems.
15. Feynman, 2028 (roadmap). Planned to pair with the Rosa CPU.

Years refer to the first product announcement or retail launch. Nvidia often announces a data center chip months before volume shipments, so dates can differ by a few months depending on the source.

One brand, two product lines

Until 2017, one architecture usually covered both GeForce gaming cards and Tesla compute cards. Volta changed that. The V100 was a data center product, and Volta never reached a mainstream GeForce series. Starting in 2022, Nvidia gave the two lines different names in the same year. Hopper went to data centers and Ada Lovelace to gaming and workstations.

Blackwell brought both lines back under one name: the B200 accelerator and the GeForce RTX 50 series share the codename, though they use very different chips. Rubin and Feynman are so far known only from data center announcements. As of October 2026, Nvidia has not announced a next GeForce generation, so any RTX 60 specifications you see online are unconfirmed.

What each generation added

Tesla to Maxwell (2006–2014)

Tesla, which debuted with the GeForce 8800 GTX in November 2006, was Nvidia's first design with unified shaders. It replaced separate vertex and pixel units with general-purpose stream processors, the base that CUDA programming was built on. Fermi (GeForce GTX 480, 2010) reorganized the chip around streaming multiprocessors with 32 CUDA cores each. Kepler, on 28 nm, focused on power efficiency. Maxwell started with the GTX 750 Ti in February 2014 and reached the high end with the GTX 980 that September.

Pascal and Volta (2016–2017)

Pascal launched with the Tesla P100 in April 2016 and moved Nvidia to FinFET transistors. The GeForce GTX 1080 followed in May 2016. Volta's V100, unveiled at GTC in May 2017, was Nvidia's first chip with Tensor Cores, the matrix units that now drive AI training and inference. It had 21.1 billion transistors.

Turing and Ampere (2018–2020)

Turing was introduced at SIGGRAPH in August 2018. It added RT cores for hardware ray tracing and launched the RTX brand with the GeForce RTX 20 series. Ampere split production across two foundries. The A100, announced on May 14, 2020, used TSMC 7 nm with 54.2 billion transistors, while the RTX 30 series, announced that September, used Samsung's 8 nm-class process.

Hopper and Ada Lovelace (2022)

Nvidia announced Hopper on March 22, 2022. The H100 has 80 billion transistors on TSMC 4N. The H200, with faster HBM3E memory, came later. Ada Lovelace was announced in September 2022 and powered the GeForce RTX 40 series, which reached stores in October 2022.

Blackwell and Blackwell Ultra (2024–2025)

Blackwell was announced at GTC on March 18, 2024. Each Blackwell GPU has 208 billion transistors on a custom TSMC 4NP process. The GB200 NVL72 rack links 72 Blackwell GPUs and 36 Grace CPUs as one system. On the consumer side, the GeForce RTX 50 series was announced at CES on January 6, 2025, and the RTX 5090 went on sale on January 30, 2025. Blackwell Ultra followed at GTC 2025 with the GB300 NVL72 and HGX B300, and Nvidia expected partner systems from the second half of 2025.

Rubin: release date and status

Rubin is the architecture that follows Blackwell. At CES on January 5, 2026, Nvidia said the Rubin platform was in full production. Rubin is a six-chip platform: the Vera CPU, Rubin GPU, NVLink 6 Switch, ConnectX-9 SuperNIC, BlueField-4 DPU and Spectrum-6 Ethernet switch. Nvidia expects partners to ship Rubin-based products in the second half of 2026. It named AWS, Google Cloud, Microsoft and OCI among the first cloud providers to deploy Vera Rubin instances.

The Vera CPU has 88 custom Arm cores. Rubin uses HBM4 memory, the generation explained in our HBM4 vs HBM3E guide, and is built on a TSMC 3 nm-class node (see our TSMC process nodes explainer).

Rubin vs Blackwell

• Process: Blackwell uses TSMC 4NP. Rubin moves to a 3 nm-class TSMC node.
• Memory: Blackwell and Blackwell Ultra use HBM3E. Rubin switches to HBM4.
• CPU partner: Blackwell racks pair with the Grace CPU. Rubin pairs with Vera.
• AI compute: Nvidia rates the Rubin GPU at 50 petaflops of NVFP4 inference compute and says it has a third-generation Transformer Engine.
• Interconnect: Rubin moves to sixth-generation NVLink switches and ConnectX-9 networking.

Blackwell and Blackwell Ultra systems will keep shipping alongside Rubin through its ramp, as Hopper did after Blackwell launched.

Rubin Ultra and Feynman: the roadmap after 2026

At GTC 2025, Nvidia's keynote recap said systems based on Rubin Ultra are due in the second half of 2027. Feynman was first shown on that same roadmap as the next architecture after Rubin. At GTC 2026, held on March 16, 2026, Nvidia described Feynman as its next major architecture after Vera Rubin. It will come with a new CPU called Rosa, alongside BlueField-5 and CX10 networking and the Kyber scale-up design. Coverage of the GTC 2026 roadmap slides places Rosa and Feynman in 2028, with custom HBM. As of October 2026, Nvidia has not published Feynman's process node, transistor count or performance.

The data center cadence is now roughly yearly: a new architecture every two years, with an "Ultra" refresh in between. Rack-scale packaging is a big part of that roadmap. Our CoWoS packaging explainer covers the TSMC technology that these large GPUs depend on.

Who the architectures are named after

• Tesla: Nikola Tesla, electrical engineer and inventor.
• Fermi: Enrico Fermi, physicist.
• Kepler: Johannes Kepler, astronomer and mathematician.
• Maxwell: James Clerk Maxwell, physicist known for the theory of electromagnetism.
• Pascal: Blaise Pascal, 17th-century mathematician and physicist.
• Volta: Alessandro Volta, physicist and chemist who invented the battery.
• Turing: Alan Turing, mathematician and computer scientist.
• Ampere: André-Marie Ampère, physicist and mathematician.
• Hopper: Grace Hopper, computer scientist and US Navy rear admiral.
• Ada Lovelace: 19th-century mathematician, often described as one of the first programmers.
• Blackwell: David Blackwell, statistician and game theorist, and the first Black scholar inducted into the National Academy of Sciences.
• Rubin: Vera Rubin, astronomer whose work provided evidence for dark matter. Nvidia's Vera CPU is also named after her.
• Feynman: Richard Feynman, theoretical physicist.
• Rosa (CPU): Rosalind Franklin, chemist whose X-ray crystallography work helped reveal the structure of DNA.

Frequently asked questions

What is the latest Nvidia GPU architecture?

As of October 2026, Rubin is Nvidia's newest architecture. Nvidia said it was in full production in January 2026 and expects partner products in the second half of 2026. Blackwell is the current architecture for GeForce RTX 50 graphics cards.

When is Nvidia Rubin released?

Nvidia announced at CES 2026 that Rubin was in full production. It said Rubin-based systems from partners and cloud providers would arrive in the second half of 2026.

What comes after Rubin?

Rubin Ultra is due in the second half of 2027, according to Nvidia's GTC 2025 roadmap. Feynman, paired with the new Rosa CPU, follows and is shown on the roadmap for 2028.

Is Blackwell used in gaming GPUs?

Yes. The GeForce RTX 50 series, launched in January 2025, uses Blackwell. It shares the codename with the B200 and B300 data center accelerators, but the chips are different designs.

Why did Nvidia use two architectures in 2022?

Nvidia split its data center and consumer lines that year. Hopper powered the H100 for AI and HPC, and Ada Lovelace powered the GeForce RTX 40 series and workstation cards.