GeForce, RTX Pro, Tesla & A/H/L/B — Decoding NVIDIA's Product Lineup

Topics We'll Cover

An end-to-end map of NVIDIA's product lineup — the names, the silicon shared between them, the licence boundaries, and how to pick the right SKU for any role.

• The Three Pillars Today
• The Naming System Decoded
• GeForce — Gaming First, AI by Accident
• RTX Pro — The Workstation Line
• Datacenter — A/H/L/B Series
• The China-Export Variants
• Jetson & Tegra — Edge AI Family
• The EULA — What You Can and Can't Do
• The "Same Die, Different Card" Phenomenon
• Driver Branches & Compute Capability
• Interactive: Card-to-Use-Case Mapper

The Three Pillars Today

NVIDIA today sells GPU silicon under three commercial pillars. The same physical die can appear in all three with different firmware, driver, BIOS, board layout, and licence — the differences are real and they matter.

What separates them physically

The Naming System Decoded

Datacenter SKUs follow a strict letter-and-number scheme. Once you know the rules, the part number tells you the architecture, the tier, and whether it's a China-export variant.

The numeric tier code

Reading a model number

GeForce — Gaming First, AI by Accident

GeForce has been NVIDIA's consumer line since 1999. It's marketed at gamers, sold through Best Buy and Amazon, and runs the GeForce driver branch. The fact that the same silicon is wildly useful for LLM inference is a side effect — one that NVIDIA accepts but does not encourage.

Lineage of the flagship

The tier grid — every Ada/Blackwell SKU

RTX Pro — The Workstation Line

The professional line has had three names in five years. Same line; new branding each generation.

Flagship by generation

What makes a card "Pro"

• ECC GDDR. Single-bit errors detected and corrected; double-bit errors flagged. Costs ~6% of nominal capacity.
• Larger VRAM. Same die as the consumer flagship but with twice or three times the memory: 24 GB on a 4090 vs 48 GB on RTX 6000 Ada (both AD102), 32 GB on 5090 vs 96 GB on RTX Pro 6000 Blackwell (both GB202).
• Certified ISV drivers. Validated against Autodesk Maya/3ds Max, Dassault SolidWorks/CATIA, PTC Creo, Adobe Premiere/After Effects, ANSYS, DaVinci Resolve. Bug fixes go through QA against those apps.
• NVIDIA Studio + RTX Enterprise driver branches. Long-life-support, slower cadence than the gaming GeForce driver.
• Datacenter EULA permitted. No "no datacenter" clause — this is a workstation/datacenter card.
• Sometimes NVLink bridge. A6000 Ampere has it; RTX 6000 Ada and RTX Pro 6000 Blackwell do not.
• Blower-style 2-slot cooling. Fits in racked workstations and 4U servers where 4-slot axial GeForce cards don't.

Datacenter — A/H/L/B Series

The datacenter line is split into three rough lanes: top compute (HBM, training-class), inference value (GDDR), and edge/specialist. All datacenter cards share passive cooling, no display outputs, full ECC, and a datacenter-permitting EULA.

Form factors

The China-Export Variants

US Bureau of Industry and Security (BIS) export controls have driven a parallel SKU family aimed exclusively at the Chinese market. The rules don't ban GPU sales — they cap the capability of any single package.

Timeline of the rules

• October 2022. First BIS rule. Bandwidth threshold: chips with bidirectional interconnect > 600 GB/s and a “total processing performance” (TPP = 2 × MAC rate × bit length) of 4800 or more need a licence to ship to China. A100 and H100 caught.
• NVIDIA response (Nov 2022). A800 — an A100 with NVLink slowed from 600 GB/s to 400 GB/s. Same compute. Compliant. H800 — H100 with NVLink slowed from 900 GB/s to 400 GB/s. Same FP8 throughput.
• October 2023. Tighter rule. Now constrains "performance density" (FLOPS / mm²) and total compute. A800/H800 now also banned. New rule explicitly anti-circumvention.
• NVIDIA response (Dec 2023). H20 — an H100 die with most SMs disabled, FP8 throughput cut to ~0.3 PFLOPS (296 TFLOPS dense, vs ~1.98 PFLOPS on H100 SXM), but full 96 GB HBM3 retained. Counter-intuitively a great pure-inference card because LLM decode is bandwidth-bound, not compute-bound.
• 2025+. Blackwell-era B20/B30/B300A variants pending US rule clarifications; specifications change as the export licence regime evolves.

Jetson & Tegra — Edge AI Family

NVIDIA's mobile/embedded line has two names. Tegra is the SoC silicon family (originally for phones and tablets, today for automotive and embedded); Jetson is the developer kit / module family that ships those SoCs as edge AI hardware. Same chip, different packaging and distribution channel.

The Orin family today

The software stack

• JetPack. NVIDIA's edge SDK bundle: CUDA, cuDNN, TensorRT, VPI (vision), DeepStream (video analytics) — all built for ARM64.
• Linux for Tegra (L4T). Ubuntu-based BSP customised for Tegra hardware (display, camera, ISP, codec).
• Holoscan. Real-time streaming framework for sensor pipelines — targeted at medical edge, surgical robotics, smart instruments.
• Isaac. Robotics platform — Isaac SIM, Isaac ROS GEMs, Isaac Manipulator. Runs on Jetson AGX as the on-board compute.
• NVIDIA Drive. Automotive variant with Drive OS (functional-safety Linux) and Drive AGX hardware (multi-Orin or Thor).

The EULA — What You Can and Can't Do

NVIDIA enforces product segmentation through software licences, not silicon. Two distinct EULAs govern almost every GPU sold today.

Enforcement reality

• By audit, not by driver. The driver does not refuse to load on consumer hardware in a colo — it has no way of knowing. NVIDIA enforces by reading customers' marketing material, monitoring procurement (via OEMs and distributors), and occasional cease-and-desist when caught.
• Big-cloud datacenters use only datacenter SKUs as a rule. AWS, GCP, Azure, OCI exclusively rack T4, A10, A100, H100, H200, L4, L40S, B200. Not because it's literally illegal otherwise but because NVIDIA's volume contracts are explicit about it.
• Small-scale commercial use is largely tolerated. An indie SaaS running a handful of 4090s in a closet typically doesn't draw attention. NVIDIA has rarely (publicly) enforced against small commercial users.
• Cloud GPU rental on consumer cards has been enforced. The 2017 wave of crypto-mining-on-Tesla and consumer-cards-as-cloud-rental triggered the first explicit "no datacenter" clause.
• Educational and government carve-outs. Some EULA versions exempt blockchain mining and educational use explicitly — check your specific driver version's EULA.

The "Same Die, Different Card" Phenomenon

The single most useful insight in this entire deck: marketing tier ≠ silicon tier. NVIDIA sells the same physical GPU die in three pillars at vastly different prices, with the differences driven by VRAM count, clocks, BIOS strap, and licence — not silicon.

GA102 (Ampere) — one die, six cards

AD102 (Ada Lovelace) — one die, five cards

GB100 / GB202 (Blackwell)

Driver Branches & Compute Capability

NVIDIA's driver isn't a single artefact — it's at least two parallel streams, and matching the driver to the silicon and the workload is essential operational hygiene.

Driver version ↔ CUDA toolkit support

Compute Capability and kernel binaries

Every CUDA kernel is compiled for a specific Compute Capability (CC). A wheel built only for sm_80 won't use Hopper's FP8 tensor cores even on an H100; you'd need sm_90. The same wheel running on Blackwell needs sm_100 (B100/B200) or sm_120 (consumer Blackwell).

# nvcc explicit target list:
nvcc -gencode arch=compute_80,code=sm_80 \
     -gencode arch=compute_86,code=sm_86 \
     -gencode arch=compute_89,code=sm_89 \
     -gencode arch=compute_90,code=sm_90 \
     -gencode arch=compute_100,code=sm_100 \
     -gencode arch=compute_120,code=sm_120 \
     -O3 kernel.cu -o kernel

# PyTorch / torch.compile environment hint:
export TORCH_CUDA_ARCH_LIST="8.0;8.6;8.9;9.0;10.0;12.0"

# Verify the GPU's CC at runtime:
nvidia-smi --query-gpu=compute_cap --format=csv

Interactive: Card-to-Use-Case Mapper

Pick a SKU. Get the family classification, the firmware features, the licence boundary, and a recommended role.