Summary · The names, and whom they honour
Semiconductors · Terminology · ExplainerNVIDIA's own naming statements, April 2021 to March 2026
Decoding the names · Hopper to Feynman
The Longer the Name, the Bigger the Thing Being Sold
H100 is one GPU, GB200 is a CPU joined to two GPUs, GB200 NVL72 is a whole rack, and the Vera Rubin POD is forty racks. Each tier of an NVIDIA name marks the product boundary moving outwards, and NVIDIA says as much: Blackwell NVL72, in its own words, made the rack “the primary unit of integration”. Inside those names sit scientists — Grace Hopper, David Blackwell, Vera Rubin, Rosalind Franklin, each named by NVIDIA itself. The convention is real and it has three documented breaks: the number after NVL has counted two different things, the newest CPU and GPU no longer share one person, and one chip carries a company's name.
- 4scientists NVIDIA itself names as the source of a product name: Hopper, Blackwell, Rubin, Franklin
- 144 → 72NVIDIA's number for the same Rubin rack, in March 2025 and then from January 2026
- 6 → 7chips in the Vera Rubin platform, January to March 2026 — the seventh is the Groq 3 LPU
- 1chip in that platform named for a company rather than a scientist or an NVIDIA product line
What each name stands for, and who says so Sorted by standing. NVIDIA's own word first; reports after it.
| Standing | Name | What it honours | Who says so |
|---|---|---|---|
| Confirmed by NVIDIA | Grace | Grace Hopper, “the U.S. computer-programming pioneer” — NVIDIA's first data-centre CPU | NVIDIA, April 2021 |
| Confirmed by NVIDIA | Hopper | Grace Hopper again, “a pioneering U.S. computer scientist” — the GPU architecture of H100 | NVIDIA, March 2022 |
| Confirmed by NVIDIA | Blackwell | David Harold Blackwell, mathematician of game theory and statistics, “the first Black scholar inducted into the National Academy of Sciences” | NVIDIA, March 2024 |
| Confirmed by NVIDIA | Vera Rubin | Vera Florence Cooper Rubin, astronomer — one name split across the Vera CPU and the Rubin GPU | NVIDIA, March 2025 and January 2026 |
| Confirmed by NVIDIA | Rosa | Rosalind Franklin, “whose X-ray crystallography revealed the structure of DNA” — the CPU planned for Feynman | NVIDIA's account of its March 2026 keynote |
| Confirmed by NVIDIA | Groq 3 LPU | No scientist: Groq, the company whose inference technology NVIDIA licensed | NVIDIA, March 2026; Groq, December 2025 |
| Reported, not in NVIDIA's words | Feynman | Richard Feynman, physicist — the GPU architecture planned for 2028 | Tom's Hardware, The Next Platform, Business Next |
| Reported, contradicted by NVIDIA | Rosa | Rosalyn Sussman Yalow, medical physicist and 1977 Nobel laureate | The Next Platform |
Timeline · 2021 to 2028
Seven years of names Plans are labelled as plans. One change cannot be dated at all.
- Apr 2021NVIDIA announces Grace, its first data-centre CPU, “named for Grace Hopper”.
- Mar 2022The Hopper architecture, also named for Grace Hopper, and its first GPU, H100: 80 billion transistors on TSMC 4N. A Hopper GPU joined to a Grace CPU is “the Grace Hopper Superchip” — one person, two chips.
- Aug 2023Groq, then independent, announces that Samsung will make its next-generation LPU on the SF4X (4nm) process.
- Mar 2024Blackwell, named for David Harold Blackwell. Its GPU is two dies acting as one, 208 billion transistors on TSMC 4NP. GB200 is two B200 GPUs and one Grace CPU; GB200 NVL72 is 36 of those in one liquid-cooled rack — 72 GPUs, 36 CPUs.
- Mar 2025At GTC, “paying tribute to astronomer Vera Rubin”, NVIDIA calls the next rack Vera Rubin NVL 144. Tom's Hardware reports Huang saying “Blackwell was named wrong”: each Blackwell GPU is two dies, so the rack holds 144. Rubin Ultra NVL576 and a Feynman generation join the roadmap.
- Dec 2025Groq announces a non-exclusive licence of its inference technology to NVIDIA. Its founder, its president and other staff join NVIDIA; Groq stays independent. CNBC reports a price of about $20 billion from one of Groq's investors; neither company states one.
- Jan 2026At CES, the Rubin platform, “named for Vera Florence Cooper Rubin”, launches with six chips. The rack is now Vera Rubin NVL72 — 72 GPUs, 36 CPUs — and the eight-GPU server board is HGX Rubin NVL8.
- Mar 2026At GTC, seven chips: the seventh is “the newly integrated NVIDIA Groq 3 LPU”, in five rack-scale systems. Feynman's CPU is Rosa, “named for Rosalind Franklin”. Vera Rubin Ultra NVL576 is now eight racks of 72 GPUs, and a Kyber rack holds 144.
- 2027 (planned)Rubin Ultra, with the LP35 LPU.
- 2028 (planned)Feynman GPU, Rosa CPU and the LP40 LPU.
- UndatedWhen NVIDIA went back from counting dies to counting packages. It falls between March 2025 and January 2026, and NVIDIA's own pages do not say when, or why.
The argument · The name follows the unit of sale
Four tiers, four kinds of object Names at different tiers describe different things and cannot be ranked against each other.
| Tier | Examples | What NVIDIA says it is |
|---|---|---|
| GPU | H100, B200 | One GPU. The leading letter is the architecture — H for Hopper, B for Blackwell. A Blackwell GPU is already two dies joined “into a single, unified GPU”. |
| Superchip | Grace Hopper, GB200, Vera Rubin | A CPU bonded to GPUs. GB200 is one Grace CPU and two B200 GPUs; the Vera Rubin superchip is one Vera CPU and two Rubin GPUs. |
| Rack | GB200 NVL72, Vera Rubin NVL72 | NVL for NVLink, then a GPU count: 36 superchips, 72 GPUs and 36 CPUs in one “NVLink domain that acts as a single, massive GPU”. |
| POD | Vera Rubin POD | Forty racks: 1,152 Rubin GPUs, “nearly 20,000 NVIDIA dies”, working as one system. |
NVIDIA states the shift itself. Its technical blog of January 2026 says Blackwell NVL72 “was the first rack-scale architecture”, “elevating the rack to the primary unit of integration”, and its March release describes AI infrastructure moving “from discrete chips and standalone servers to fully integrated rack-scale systems, POD-scale deployments, AI factories”. The names follow. Business Next's explainer makes the reader's point plainly: names at different tiers cannot be compared directly, and the digits mainly separate products within one generation rather than rank their performance. What the names do not hold still is the number after NVL. In March 2025 NVIDIA called the next rack Vera Rubin NVL 144, counting dies — 72 GPUs of two dies each; from January 2026 the same rack is Vera Rubin NVL72, counting packages. The name NVL576 has meant, in turn, 576 dies in one rack, as Tom's Hardware reported the 2025 slides, and 576 GPUs across eight racks, as NVIDIA's own blog defines it in 2026. And in six NVIDIA pages announcing Rubin, no single Rubin GPU part number appears at all — where the Hopper and Blackwell releases named H100 and B200. What is named is the superchip, the board, the rack and the POD. That the name moves outward with the product is this page's reading of NVIDIA's documents; each fact under it is NVIDIA's.
- 144 → 72the same Rubin rack, counted in dies and then in packages
- 576one name, two definitions: dies in one rack, then GPUs across eight
- 0Rubin GPU part numbers in six NVIDIA pages announcing Rubin
Surname, forename — and one person, then two
Business Next sets out the convention: GPU architectures mostly take a scientist's surname, CPUs mostly a forename. Twice, the two halves were one person. Grace and Hopper are Grace Hopper; Vera and Rubin are Vera Rubin, which is why that platform carries the full name. Reporting Feynman in March 2025, Tom's Hardware expected the pattern to hold — “Presumably that means we'll get Richard CPUs with Feynman GPUs”. The CPU announced a year later is Rosa, named for Rosalind Franklin. Feynman's GPU and its CPU honour two different people, the first time the pairing splits.
The one name that is a company's
Of the Vera Rubin platform's seven chips, five carry an NVIDIA product-line name — NVLink, ConnectX, BlueField, Spectrum — or a scientist's: Vera, Rubin. The seventh, added between January and March 2026, is the NVIDIA Groq 3 LPU. Groq is the company that licensed its inference technology to NVIDIA in December 2025. This is an observation about names: the one chip named for a company is the one whose technology NVIDIA licensed. Neither company has said who designed it.
Rosa: NVIDIA's answer, and another
NVIDIA's own account of its March 2026 keynote says Rosa is “named for Rosalind Franklin, Huang said”, and Business Next says the same. The Next Platform says Rosalyn Sussman Yalow, the medical physicist, and cites no NVIDIA document; Tom's Hardware's transcription of the keynote renders Huang's words as “Ros, short for Rosalyn”. The company's written statement about its own product is Franklin.
What others add · The people behind the names
Six scientists Four named by NVIDIA; Feynman by the trade press; Yalow by one publication that NVIDIA's account contradicts.
Grace · Hopper
Grace Brewster Murray Hopper
- PhD in mathematics from Yale, 1934
- Her team at Remington Rand built A-0, the first compiler, in 1952
- Retired from the US Navy as a rear admiral at 79
Blackwell
David Harold Blackwell
- Statistician and mathematician: probability, game theory, information theory
- The Rao–Blackwell theorem, and an independent invention of dynamic programming
- First Black member of the National Academy of Sciences; first tenured Black professor at Berkeley
Vera · Rubin
Vera Florence Cooper Rubin
- Carnegie astronomer; with Kent Ford, measured how fast stars orbit in Andromeda and dozens of other galaxies
- Found each spiral galaxy sits in a halo of unseen mass — dark matter
- National Medal of Science
Feynman
Richard P. Feynman
- Nobel Prize in Physics 1965, for work in quantum electrodynamics
- Named as the namesake by Tom's Hardware, The Next Platform and Business Next; not stated in NVIDIA's own pages read here
Rosa
Rosalind Franklin
- X-ray crystallographer at King's College London
- “Photograph 51”, May 1952, showed DNA's helical structure
- Later founded structural virology work at Birkbeck
Rosa, as The Next Platform has it
Rosalyn Sussman Yalow
- Nuclear physicist; developed radioimmunoassay with Solomon Berson
- Nobel Prize in Physiology or Medicine 1977
- NVIDIA's own account names Franklin, not Yalow
Four generations, as NVIDIA documents them A process node appears only where NVIDIA itself states one.
| Architecture | Year | CPU | Process | Transistors per GPU |
|---|---|---|---|---|
| Hopper | 2022 | Grace | TSMC 4N | 80 billion (H100) |
| Blackwell | 2024 | Grace | TSMC 4NP | 208 billion, two dies |
| Rubin | 2026 | Vera | not stated by NVIDIA | 336 billion, two compute dies |
| Feynman | 2028 (planned) | Rosa | not stated by NVIDIA | not announced |
The seventh chip A licence, a reported price, and a reported foundry.
NVIDIA's January 2026 launch and its technical blog of the same day both say six chips. Its March release says seven, adding “the newly integrated NVIDIA Groq 3 LPU”. Behind the addition is Groq's own announcement of 24 December 2025: “a non-exclusive licensing agreement with Nvidia for Groq's inference technology”, with its founder Jonathan Ross, its president Sunny Madra and other staff joining NVIDIA, and Groq continuing “as an independent company”. In an email to staff quoted by CNBC, Huang wrote: “we are not acquiring Groq as a company”. The widely quoted price, about $20 billion in cash, comes from Alex Davis, whose firm led Groq's previous funding round, as CNBC reported; neither company has stated one. Jon Peddie Research calls the deal an asset acquisition and Tom's Hardware a licensing and talent deal. In NVIDIA's rack the chip is the LP30, each with 500 MB of SRAM; 256 of them give the LPX rack its 128GB. Tom's Hardware reports that Samsung makes it on a 4nm process, which fits Groq's 2023 announcement that Samsung would build its next LPU on SF4X; NVIDIA's own pages do not name the foundry. LP35 is planned for Rubin Ultra and LP40 for Feynman — names that follow Groq's numbering, not a scientist's.
- licencethe form Groq itself gives the deal; Groq stays independent
- $20bnreported by CNBC from one investor; neither company confirms a price
- 256 × 500 MBLP30 chips in one LPX rack — 128GB of on-chip SRAM, NVIDIA's figures
Conclusion · Reading a name you have not seen
How to read the next one Four questions, in order.
- What tier is it? A letter and digits is a GPU; a CPU name joined to a GPU name is a superchip; NVL and a number is a rack; POD is many racks. Never compare across tiers.
- What does the NVL number count? Since January 2026, GPU packages. From March 2025 until the change, NVIDIA's own name for a Rubin rack counted dies.
- Whose name is it? Ask NVIDIA's own announcement first. It has named Grace Hopper, David Blackwell, Vera Rubin and Rosalind Franklin; for Feynman, the attribution is the press's.
- Is it a scientist at all? A name that follows no scientist and no NVIDIA product line — Groq — marks technology that came from outside.
In summary
NVIDIA's names lengthen as the thing it sells grows, from a GPU to a rack to a POD, and NVIDIA's own documents say the rack became the unit. The scientists inside the names are real and, for four of them, NVIDIA's own choice in its own words. Hold three things loosely: the number after NVL has not always counted the same thing, the one-person pairing of CPU and GPU ended with Rosa and Feynman, and the deal behind the Groq 3 LPU is a licence whose price neither party has stated.
What could not be verified
Unconfirmed report — Jon Peddie Research: that Huang confirmed Feynman will be made on TSMC's A16 (1.6nm) process. Business Next also gives 1.6nm, marked as reported. No NVIDIA page read states a node for Feynman.
Unconfirmed report — Business Next: that Rubin is made on a TSMC 3nm process. No NVIDIA page read states Rubin's node.
Rumour — that the Groq 3 LPU is a chip NVIDIA did not design. No source read says who designed it.
Take-home and sources
- NVIDIA Newsroom — Grace CPU (12 Apr 2021), Hopper (22 Mar 2022), Blackwell (18 Mar 2024), Rubin at CES (5 Jan 2026), Vera Rubin at GTC (16 Mar 2026)
- NVIDIA GB200 NVL72 product page
- NVIDIA Blog, GTC 2025 and GTC 2026 keynote live updates
- NVIDIA Technical Blog — Aubrey, “Inside the NVIDIA Vera Rubin Platform” (5 Jan 2026), Bhargava, “NVIDIA Vera Rubin POD” (16 Mar 2026), Aubrey, “Inside NVIDIA Groq 3 LPX” (16 Mar 2026)
- Groq Newsroom (24 Dec 2025)
- Groq via PR Newswire (15 Aug 2023)
- Faber, CNBC (24 Dec 2025)
- Walton (18 Mar 2025), Shilov (17 Mar 2026) and James (19 Mar 2026), Tom's Hardware
- Prickett Morgan, The Next Platform (19 Mar 2026)
- Peddie, Jon Peddie Research (17 Mar 2026)
- 王聖華, Business Next (11 Aug 2026)
- Yale News (10 Feb 2017)
- Sanders, Berkeley News (15 Jul 2010)
- Carnegie Science (26 Dec 2016)
- NobelPrize.org, Feynman and Yalow
- King's College London, Rosalind Franklin.