반도체·하드웨어 · 리포트

NVIDIA GPU 로드맵 2026-2030: Rubin, Rubin Ultra, Feynman

업계 전문지Slicast · 2026년 10월 26일 21:09 UTC · 글로벌 · 출처: VRLA Tech

중요도 78

NVIDIA confirmed at GTC 2025, and reaffirmed at CES 2026, GTC 2026, and Computex 2026, that it has moved to an annual architecture cadence: Blackwell now, Rubin in H2 2026, Rubin Ultra in H2 2027, Feynman in 2028, and Rosa Feynman in 2029–2030. This roadmap covers the full timeline by tier (datacenter, workstation pro, consumer, and PC platform) and separates confirmed dates from NVIDIA's own announcements from rumored timelines reported by industry leakers.

## Reading This Roadmap

Every entry carries one of three tags:

- **Confirmed:** Officially announced by NVIDIA with a public timeline.

- **Rumored:** Reported by industry leakers (kopite7kimi and others) and covered by Tom's Hardware, VideoCardz, KitGuru, or similar outlets, but not confirmed by NVIDIA.

- **Projected:** Based on historical NVIDIA product cycles. Not announced or rumored, but a defensible estimate.

## Current State: Mid-2026

As of June 2026, the NVIDIA lineup shipping in volume is as follows:

| Tier | Architecture | Top Product |

|---|---|---|

| Datacenter (rack-scale) | Blackwell Ultra | GB300 NVL72: 1.1 EFLOPS FP4 per rack, 130 TB/s NVLink, 132–140 kW per rack |

| Datacenter (SXM) | Blackwell Ultra, Blackwell, Hopper | B300 (288GB HBM3e), B200 (192GB HBM3e), H200 (141GB HBM3e), H100 SXM5 (80GB HBM3) |

| Workstation Pro | Blackwell | RTX PRO 6000 Blackwell: 96GB GDDR7 ECC, 1.79 TB/s, 24,064 CUDA cores, 752 Tensor cores, 600W |

| Consumer Flagship | Blackwell | RTX 5090: 32GB GDDR7, 1.79 TB/s, 575W |

| RTX Spark (mini-PC/laptop) | Grace Blackwell | Launching Fall 2026: Grace Blackwell Spark with up to 128GB LPDDR5X unified memory |

Blackwell Ultra B300 is the current production datacenter GPU, with 288GB HBM3e (50% more than the B200), 8 TB/s memory bandwidth, 15 PFLOPS dense FP4 compute, and a 1,400W TDP. DGX B300 systems are shipping with 8–12 week lead times. The RTX PRO 6000 Blackwell is the current top professional workstation GPU. Hopper (H100 and H200) remains in production for customers with existing Hopper infrastructure and software stacks.

## H2 2026: Vera Rubin Launches (Confirmed)

### Vera Rubin Platform (Datacenter)

NVIDIA confirmed that Vera Rubin entered full production at GTC Taipei on June 1, 2026. Partner availability begins in H2 2026.

- **Rubin GPU:** 336B transistors, TSMC N3, 288GB HBM4, 22 TB/s, 50 PFLOPS NVFP4 inference, 35 PFLOPS NVFP4 training, NVLink 6 at 3.6 TB/s

- **Vera CPU:** 88 Olympus ARM cores, 176 threads, 227B transistors, NVLink-C2C 1.8 TB/s

- **Vera Rubin NVL72 rack:** 72 Rubin GPUs, 36 Vera CPUs, 260 TB/s aggregate NVLink, 3.6 EFLOPS NVFP4 inference, 2.5 EFLOPS training, 100% liquid cooled

**First cloud deployments:** AWS, Google Cloud, Microsoft Azure, Oracle Cloud, CoreWeave, Lambda, Nebius, and Nscale.

**HBM4 suppliers (NVIDIA-certified June 2026):** SK Hynix, Samsung, and Micron.

For deep technical specifications, see *NVIDIA Vera Rubin Architecture Explained*.

### Vera Rubin NVL144 CPX (Specialized Inference) (Confirmed)

NVIDIA Rubin CPX is a specialized GPU class with 128GB of GDDR7 memory, optimized for the compute-bound prefill phase of million-token context inference. It is paired with standard Rubin GPUs in the NVL144 CPX rack.

- **Per-rack compute:** 8 exaflops NVFP4 (7.5x the GB300 NVL72)

- **Per-rack memory:** 100TB

- **Per-rack memory bandwidth:** 1.7 PB/s

- **Target workloads:** Million-token coding, generative video, and agentic AI

- **Availability:** End of 2026, per the NVIDIA Newsroom

- **Early AI partners:** Cursor, Runway, and Magic

## 2026–2027 Window: RTX Spark Generation Transition (Confirmed on Roadmap)

### RTX Spark Platform

NVIDIA confirmed at Computex 2026 that its RTX Spark platform (Windows on ARM mini-PCs and laptops) will follow a multi-generation roadmap:

- **Grace Blackwell Spark (Fall 2026):** 20 ARM CPU cores, RTX 5070-class graphics, up to 128GB LPDDR5X unified memory

- **Vera Rubin Spark (2027–2028):** Vera CPU and Rubin GPU, LPDDR6 memory

- **Rosa Feynman Spark (2029–2030):** Rosa CPU and Feynman GPU

**DGX Station for Windows,** based on the GB300 Grace Blackwell Ultra Desktop Superchip with 748GB of coherent memory and up to 20 PFLOPS FP4 performance, is expected in Q4 2026 from ASUS and other OEMs.

## H2 2027: Rubin Ultra (Kyber Rack) (Confirmed)

### Rubin Ultra Platform

NVIDIA showcased Rubin Ultra hardware at GTC 2026, where CEO Jensen Huang demonstrated a system featuring 1TB of HBM4e memory on the new Kyber rack (per coverage from Tom's Hardware, DCD, and VideoCardz).

- **Per-GPU package:** Four reticle-sized compute dies (versus two on standard Rubin), ~100 PFLOPS NVFP4, 1TB HBM4e, ~32 TB/s memory bandwidth, 3.6 kW per package

- **Vera CPU:** Continues, with 88 Olympus ARM cores

- **NVLink generation:** NVLink 7 (~10.8 TB/s per GPU, a 6x improvement over NVLink 6)

- **Kyber NVL576 rack:** 144 quad-die Rubin Ultra packages, equal to 576 GPU compute dies; ~600 kW per rack; 800V DC distribution; vertical-tray modular design

- **Per-rack performance:** 15 exaflops FP4 inference, 5 exaflops FP8 training, 365TB total memory, 4.6 PB/s HBM4e bandwidth

- **Networking:** ConnectX-9, with 115.2 TB/s between racks

- **Cooling:** Direct liquid cooling required at 600 kW per rack

Rubin Ultra is designed for hyperscaler AI factory deployment. The Kyber rack architecture is incompatible with traditional Blackwell NVL72 infrastructure and requires new facility power and cooling design.

## 2027 (Rumored): Consumer RTX 60-Series

### RTX 60-Series (GR20x Family)

This is a rumor and has not been confirmed by NVIDIA. Leaker kopite7kimi reported in January 2026 that NVIDIA will use the Rubin architecture for consumer RTX 60-series cards under the GR20x die family naming convention, targeting an H2 2027 launch.

*Source: kopite7kimi on X (January 2026), as reported by Tom's Hardware, KitGuru, VideoCardz, and TechPowerUp.*

- **Rumored architecture:** Rubin (TSMC N3)

- **Rumored top SKU:** RTX 6090, with ~40% rendering uplift over the RTX 5090

- **Rumored timeline:** H2 2027, possibly slipping to late 2027 or early 2028

- **Cited drivers of delay:** AI datacenter demand for TSMC N3 wafers, HBM supply constraints, and a shelved RTX 50 SUPER refresh

NVIDIA has not officially announced any consumer GPU launch beyond the RTX 50-series. The CES 2026 and GTC 2026 keynotes did not include consumer GPU announcements. Until NVIDIA confirms a product and timeline, the RTX 60-series remains a rumor.

## 2027–2028 (Speculation): RTX PRO Rubin Workstation Card (Projected)

### RTX PRO Workstation Successor

NVIDIA has not announced a Rubin-based RTX PRO workstation card. The following is a projection based on historical NVIDIA product cycle patterns, not a rumor or a confirmation.

- **Historical pattern:** NVIDIA's RTX PRO workstation cards follow the consumer flagship by 3 to 9 months (for example, the RTX 5090 launched in January 2025, and the RTX PRO 6000 Blackwell launched in early 2025).

- **Projected timing:** If the consumer Rubin (RTX 60-series) ships in H2 2027 as rumored, a corresponding RTX PRO Rubin workstation card would realistically land between late 2027 and mid-2028.

- **Projected positioning:** Likely 96GB+ of HBM4 or GDDR7 ECC memory, datacenter-rated for sustained 24/7 operation, with full RTX PRO driver validation.

- **Likely channels:** Workstation OEMs and custom system integrators, including VRLA Tech.

For current workstation procurement, the RTX PRO 6000 Blackwell remains the top option. VRLA Tech configures Threadripper PRO 9000WX workstations with RTX PRO 6000 Blackwell, shipping today.

## 2028: Feynman (Confirmed)

### Feynman Architecture

NVIDIA confirmed Feynman as the architecture following Rubin Ultra. It is named after physicist Richard Feynman and was announced at GTC 2025, with additional details shared at GTC 2026.

- **Successor to:** Rubin Ultra

- **Paired CPU:** Rosa (next-generation custom NVIDIA ARM CPU)

- **Confirmed technology directions:** Advanced 3D stacking for higher die density, LP40 memory, BlueField-5 DPU, NVLink-8 interconnect, and "custom HBM" (likely a successor to HBM4e)

- **Release window:** 2028, per the NVIDIA roadmaps shown at GTC 2025 and GTC 2026

Detailed specifications have not yet been publicly disclosed. Feynman represents NVIDIA's commitment to an annual architectural cadence through 2028.

## 2029–2030: Rosa Feynman (Confirmed on Roadmap)

### Rosa Feynman Generation

According to VideoCardz and OC3D coverage of the NVIDIA Computex 2026 roadmap slide, the architecture following Feynman is Rosa Feynman, scheduled for 2029–2030.

- **Successor to:** Feynman

- **Paired CPU:** Rosa

- **Release window:** 2029–2030

- **Public detail:** Limited as of mid-2026. The architecture appears on roadmap slides without architectural specifics.

## The Full Timeline at a Glance

|---|---|---|---|

| End 2026 | Vera Rubin NVL144 CPX | — | — |

| 2029–2030 | Rosa Feynman | — | — |

## Capital Planning by Buyer Type

### Enterprise On-Prem AI Infrastructure

For deployments through Q2 2027: Blackwell B300 or B200 SXM in EPYC GPU servers. VRLA Tech configures EPYC

원문 보기