Edited by 
Short answer
Rubin is the NVIDIA data-center GPU that follows Blackwell. Each Rubin GPU has 288 GB of HBM4 at 19.2 TB/s, 35 PFLOPS of dense NVFP4 for training (50 PFLOPS for inference), 17.5 PFLOPS of FP8 and 3 TB/s of NVLink 6. It ships in Vera Rubin NVL72 racks that pair 72 Rubin GPUs with 36 Vera CPUs. NVIDIA put the platform into full production in January 2026, and production shipments began in the second half of 2026. Against a GB300 GPU, Rubin has about 2.4 times the memory bandwidth and 2.3 times the dense FP4 throughput.
Rubin GPU specifications
All figures are per GPU and dense (without 2:4 sparsity) unless stated. Rubin is a two-die package, like Blackwell, and NVIDIA counts the package as one GPU.
| Specification | Rubin (Vera Rubin NVL72) | GB300 (Blackwell Ultra) | Change |
|---|---|---|---|
| Transistors | 336 billion | 208 billion | 1.6× |
| Memory | 288 GB HBM4 | 288 GB HBM3e | same capacity |
| Memory bandwidth | 19.2 TB/s | 8 TB/s | 2.4× |
| NVFP4 | 35 PFLOPS (training) · 50 PFLOPS (inference) | 15 PFLOPS | 2.3× |
| FP8 / FP6 | 17.5 PFLOPS | 5 PFLOPS | 3.5× |
| FP16 / BF16 | 4 PFLOPS | 2.5 PFLOPS | 1.6× |
| FP64 | 33 TFLOPS | 1.25 TFLOPS | 26× |
| NVLink per GPU | 3 TB/s (NVLink 6) | 1.8 TB/s (NVLink 5) | 1.7× |
| Streaming multiprocessors | 224 | 160 | 1.4× |
| CPU partner | Vera (88 cores) | Grace (72 cores) |
Compare Rubin with 27 other chips, including AMD MI455X, in the AI GPU specs database.
What changed from Blackwell
HBM4 memory
Rubin is NVIDIA's first GPU with HBM4. HBM4 doubles the interface width of each stack to 2,048 bits, which is how Rubin reaches 19.2 TB/s with the same 288 GB as Blackwell Ultra. Bandwidth matters most for generating tokens, because every new token has to read the model's active weights and its KV cache. At the same model and precision, 2.4 times the bandwidth means up to 2.4 times the tokens per second for one user.
Faster low-precision Tensor Cores
Rubin's Tensor Cores push hardest on the formats used for large language models. Dense FP8 rises 3.5 times and NVFP4 about 2.3 times, while FP16 rises only 1.6 times. The 50 PFLOPS inference figure relies on what NVIDIA calls adaptive compression in the Transformer Engine, so the 35 PFLOPS training figure is the fairer comparison with other chips. Read NVFP4 explained for how the 4-bit format keeps accuracy.
FP64 returns
Blackwell Ultra cut FP64 to about 1.25 TFLOPS per GPU to spend the area on AI formats. Rubin brings it back to 33 TFLOPS, close to B200's 37 TFLOPS, which matters for scientific computing customers who buy the same systems.
NVLink 6 and a bigger rack fabric
Each Rubin GPU gets 3 TB/s of NVLink 6, against 1.8 TB/s of NVLink 5 on Blackwell. Across the 72 GPUs of an NVL72 rack that adds up to 216 TB/s of scale-up bandwidth. See NVLink and NVSwitch explained for why this matters for mixture-of-experts models.
The Vera CPU
Vera replaces Grace as the CPU in NVIDIA's superchips. It has 88 custom Arm cores (NVIDIA calls them Olympus) running 176 threads, up to 1.5 TB of LPDDR5X memory at up to 1.2 TB/s, and a 1.8 TB/s NVLink-C2C link to the GPUs. That link is twice the 900 GB/s of Grace, which helps when models spill data such as the KV cache into CPU memory.
Vera Rubin NVL72 rack
| Rack total | Vera Rubin NVL72 |
|---|---|
| GPUs and CPUs | 72 Rubin GPUs, 36 Vera CPUs |
| NVFP4 | 3,600 PFLOPS inference · 2,520 PFLOPS training |
| FP8 | 1,260 PFLOPS |
| FP16 / BF16 | 288 PFLOPS |
| HBM4 | 20.7 TB |
| CPU memory | Up to 54 TB LPDDR5X |
| NVLink bandwidth | 216 TB/s |
| Cooling | Liquid, with 45 °C supply water |
The rack was first announced as "Vera Rubin NVL144", counting GPU dies rather than packages. NVIDIA's 2026 materials count packages instead, so the same 72-package rack is called NVL72, matching GB200 NVL72.
The 2026 bandwidth revision
When NVIDIA detailed Rubin in early 2026 it quoted 22 TB/s of HBM4 bandwidth and 3.6 TB/s of NVLink 6 per GPU. In September 2026 NVIDIA's product pages changed to 19.2 TB/s and 3 TB/s, without a press release. Reports linked the change to HBM4 supply. The compute figures did not change. This article and our specs database use the current figures.
The chips in the Rubin platform
NVIDIA presents Rubin as a set of chips designed together, not just a GPU. At CES 2026 it named six:
- Rubin GPU, the accelerator described above.
- Vera CPU, the host processor.
- NVLink 6 Switch, which connects the 72 GPUs in a rack.
- ConnectX-9 SuperNIC, giving each GPU 1.6 Tb/s of networking.
- BlueField-4 DPU, for storage, security and networking offload.
- Spectrum-6 Ethernet switch, for scale-out networking between racks.
At GTC in March 2026 NVIDIA described the platform as seven chips.
Timeline and availability
- March 2025: Rubin and Rubin Ultra shown on NVIDIA's roadmap at GTC.
- September 2025: Rubin CPX, a GDDR7 chip for long-context prefill, announced.
- January 5, 2026: Rubin launched at CES and declared in full production. NVIDIA claimed up to 10 times lower inference cost per token and 4 times fewer GPUs to train large models than Blackwell. These are NVIDIA's figures for its chosen workloads.
- March 2026: Architecture details at GTC. Micron and Samsung began HBM4 volume production in the first quarter.
- August 2026: NVIDIA reported production shipments starting. Microsoft's Fairwater AI data centers and CoreWeave are among the first named users.
- September 2026: Memory and NVLink figures revised down, as described above.
What comes after Rubin
NVIDIA's roadmap places Rubin Ultra in 2027. It is planned with four GPU dies per package, up to 1 TB of HBM4e and about 100 PFLOPS of FP4 per package, in NVIDIA's denser Kyber rack design. Feynman is the architecture planned for 2028. The status of Rubin CPX is unclear: it was missing from NVIDIA's GTC 2026 presentation, and NVIDIA has not confirmed a ship date.
What Rubin means in practice
- For inference providers: the jump in bandwidth and FP4 throughput lowers cost per token, especially for large mixture-of-experts models that need the whole 72-GPU NVLink domain.
- For training: 3.5 times the FP8 throughput per GPU means fewer GPUs for the same job, if software keeps the Tensor Cores busy.
- For everyone else: Rubin is sold as complete racks through NVIDIA's partners and clouds. Most people will use it by renting capacity, not buying a card.
To see how AMD's rival compares, read AMD MI355X vs NVIDIA B200 or open the specs database and compare Rubin with MI455X.
Frequently asked questions
When is NVIDIA Rubin released?
NVIDIA launched Rubin at CES on January 5, 2026, and said it was in full production. Production shipments to partners began in the second half of 2026, with Vera Rubin NVL72 racks going to cloud providers and AI labs first.
How much memory does a Rubin GPU have?
Each Rubin GPU has 288 GB of HBM4 with 19.2 TB/s of bandwidth. A full Vera Rubin NVL72 rack has 20.7 TB of HBM4.
Is Rubin faster than Blackwell?
Per GPU, Rubin has about 2.4 times the memory bandwidth of B300 or GB300, 3.5 times the dense FP8 throughput and 2.3 times the dense FP4 throughput. Real speedups depend on the model and software. NVIDIA's own claim is up to 10 times lower inference cost per token.
What is the difference between Vera and Rubin?
Rubin is the GPU and Vera is the CPU. NVIDIA names the combined platform and rack after both, as it did with Grace Blackwell.
Why did Vera Rubin NVL144 become NVL72?
Both names describe the same 72-package rack. NVL144 counted the two dies inside each Rubin package as separate GPUs. NVIDIA's 2026 materials count packages, which gives NVL72.
