Data Center GPU
The data center GPU lineup that accelerates AI and HPC workloads.
The core computing engines that accelerate generative AI, large language models, graphics, and simulation. Offered by workload — from the latest Blackwell generation (B200 · B300) to proven Hopper (H200 · H100), universal acceleration (L40S), and professional-grade (RTX PRO 6000).
What Are Data Center GPUs
Data center GPUs are accelerators that speed up AI matrix operations with Tensor Cores. The Blackwell generation maximizes training and inference for trillion-parameter models with FP4/FP6 low-precision compute, the 2nd-generation Transformer Engine, and 5th-generation NVLink, while the Hopper generation (H100 · H200) remains widely deployed thanks to its proven reliability and high-capacity HBM.
GPU Lineup (Specs by Line)
B200
Blackwell flagship · 192GB HBM3e · 8TB/s · up to ~20 PFLOPS FP4 — roughly 4× the AI performance of H100
B300 (Blackwell Ultra)
~288GB HBM3e — expanded memory and inference performance over B200, optimized for large-scale inference and agentic AI
GB200 Grace Blackwell Superchip
A superchip combining a Grace CPU with 2× B200 (384GB HBM3e) over NVLink-C2C — the basic unit of the GB200 NVL72 rack
H200
Hopper · 141GB HBM3e · 4.8TB/s — the H100 successor optimized for LLM inference with high-capacity memory
H100
Hopper Tensor Core · 80GB HBM3 · 3.35TB/s · FP8 — the proven standard for large-scale training and inference
L40S
Ada Lovelace · 48GB GDDR6 — a universal GPU spanning AI inference, graphics, and virtualization
RTX PRO 6000 Blackwell
Blackwell · 96GB GDDR7 — the top-tier professional workstation and server GPU
Next-Generation Roadmap — Rubin
Rubin GPU (2026)
Next generation — HBM4 · NVLink 6, paired with the Vera CPU to form the Rubin NVL144 rack
Rubin Ultra (2027) · Feynman (2028)
The roadmap continues with Rubin Ultra NVL576 (2027) followed by the Feynman generation (2028) (per announcements; detailed specs to follow)
Key Spec Comparison
| Model | Architecture | Memory | Bandwidth | AI Performance | Use |
|---|---|---|---|---|---|
| B200 | Blackwell | 192GB HBM3e | 8TB/s | FP4 ~20 PFLOPS * | Large-scale training & inference |
| B300 | Blackwell Ultra | ~288GB HBM3e | ~8TB/s | FP4 uplift * | Large-scale inference & agentic AI |
| H200 | Hopper | 141GB HBM3e | 4.8TB/s | FP8 (H100 class) | High-memory inference |
| H100 | Hopper | 80GB HBM3 | 3.35TB/s | FP8 ~4 PFLOPS (sparse) | Proven training & inference |
| L40S | Ada Lovelace | 48GB GDDR6 | 864GB/s | FP8 ~1.5 PFLOPS (sparse) | Universal · graphics · inference |
| RTX PRO 6000 | Blackwell | 96GB GDDR7 | — | Professional-grade | Workstations & servers |
* Performance figures are NVIDIA-published Tensor Core values and vary with precision (FP4/FP8) and sparsity conditions. Values marked * are approximate or roadmap-based.