How to Read NVIDIA GPU Names — Architecture Generations and Product Lines

A100, H100, H200, B200, GB300, L40S, RTX 5090. NVIDIA’s GPU names look like a cipher at first, but two axes decode most of them: the architecture generation (a scientist’s name) and the product line (the intended use). This post lays out that grid in one place, as of mid-2026. GPU internals are covered in Hardware Intermediate #8 and the server-level picture in the previous post.

Axis 1 — architecture generations: scientists’ names #

NVIDIA names each GPU design generation after a scientist, and a product name’s leading letter usually points to that generation.

The published roadmap continues with Rubin Ultra (2027) and Feynman (2028). Note the ordering follows the order in which NVIDIA adopted the names, not the alphabet — “H before B” is correct here.

Axis 2 — product lines: roles within a generation #

Each architecture splits into products by role.

• Datacenter flagships (x100/x200): the A100 → H100 → B200 line, the top training tier. The x200-style parts (H200, B300) are memory-boosted variants of the same generation: H200 is an H100 with larger HBM, and B300 (Blackwell Ultra) extends B200 to 288GB of HBM3e for heavier inference and training. Large models bottleneck on memory capacity and bandwidth before compute, hence these derivatives.
• The GB prefix: GB200 and GB300 pair a Grace CPU (NVIDIA’s arm CPU) with GPUs in one superchip module. Add a suffix like NVL72 and it means 72 GPUs bound into a single NVLink domain — a rack-scale product where the sales unit is the rack, not the server.
• Inference/graphics line (the L series and friends): L4 and L40S are Ada-generation inference and general-acceleration cards. Far cheaper and lower-power than flagships (the L4 draws 72W), they’re everywhere in small-model inference, video processing, and VDI. They lack NVLink, so they’re unsuited to large multi-GPU workloads.
• Consumer GeForce RTX: in RTX 4090 and RTX 5090, the first two digits are the generation (40 = Ada, 50 = Blackwell) and the last two are the tier (90 > 80 > 70…). They share architectures with the datacenter parts, so they serve for development and experiments, but VRAM (32GB on the 5090), the absence of ECC and NVLink, and datacenter-deployment license restrictions separate them from production serving.
• Workstation RTX PRO: the former Quadro line — consumer silicon with enlarged VRAM and ECC for professional workstations.

From name to spec — the three numbers that matter #

Once the name is decoded, comparison narrows to three items on the spec sheet.

1. VRAM capacity — the first gate: does the model fit at all? Within and across generations the spread is wide: H100 80GB / H200 141GB / B200 192GB / B300 288GB / L40S 48GB / L4 24GB.
2. Memory bandwidth — LLM inference token throughput is often bound by memory bandwidth, not compute. HBM generations (HBM3 → HBM3e → HBM4) are effectively bandwidth generations.
3. Compute per precision and supported formats — newer generations add hardware support for lower precisions (FP8, FP4), and marketing numbers usually quote the lowest precision available. When comparing across generations, check that the same precision is being compared — that’s where the exaggeration hides.

When an article claims “N times faster,” the habit worth having is to check the precision conditions and whether a rack-scale (NVL) product is being compared against a single chip.

The confusions worth preempting #

• H200 is Hopper — an older generation than B200. The shared “200” doesn’t mean the same era. The letter is the generation; the number is a derivative within it.
• A10 and A100 are different classes. Both Ampere, but A100 is the flagship and A10 is the inference/graphics line. The digit count signals the class.
• Cloud instance names are yet another layer. AWS’s p families (p4 = A100, p5 = H100) and g families (g5 = A10G, g6 = L4), GCP’s a families — every cloud has its own scheme, so confirm the actual GPU model in the instance spec sheet.
• “Blackwell” alone tells you nothing about class. The B200 (datacenter flagship) and the RTX 5060 (budget consumer) are both Blackwell. Architecture names date the design; product lines set the performance tier.

Summary #

• Read NVIDIA GPU names as a grid: architecture generations (scientists: Volta → Ampere → Hopper → Ada → Blackwell → Rubin) × product lines (datacenter x100/x200, GB superchips, L inference line, RTX consumer).
• x200-style parts (H200, B300) are memory-boosted variants; GB means Grace CPU included; NVL means a rack-scale NVLink domain.
• As of mid-2026, Blackwell (B200/B300) is mainstream, Hopper (H100/H200) remains in production, and Rubin is just beginning to ship.
• After the name, compare on VRAM capacity, memory bandwidth, and precision conditions. Treat marketing multiples with the precision check.
• In the cloud, judge by the actual GPU model in the spec sheet, not the instance name.