TL;DR
- Workstation card built on AD102 with 48 GB GDDR6 ECC and active cooling — sister product to L40 / L40S.
- Pitched at designers, animators, scientific computing and on-desk AI research.
- 300 W triple-slot PCIe card with four DisplayPort 1.4a outputs; supports FP8 and AV1 encode.
- Frequently used as a desk-side development platform for code that later runs on L40S or H100 in the data centre.
Overview#
The RTX 6000 Ada Generation is the workstation incarnation of AD102 — the same silicon as L40 / L40S but in a triple-slot, actively cooled card with display outputs. NVIDIA positions it for workstations: CAD, 3D rendering, scientific visualisation, and increasingly local AI development.
For AI teams it is the canonical 'powerful desk-side card'. 48 GB of ECC memory comfortably hosts 13B-class models in FP16 or 70B-class models in FP4 quantised form, making it a strong fit for inference development and lightweight fine-tuning before moving to data centre hardware.
Specifications#
| Metric | RTX 6000 Ada |
|---|---|
| Architecture | Ada Lovelace (AD102) |
| Process | TSMC 4N |
| FP32 | 91.1 TFLOPS |
| TF32 (Tensor, sparse) | 364 TFLOPS |
| BF16 / FP16 (Tensor, sparse) | 728 TFLOPS |
| FP8 (Tensor, sparse) | 1,457 TFLOPS |
| RT Cores | Third-generation |
| Memory | 48 GB GDDR6 ECC |
| Memory bandwidth | 960 GB/s |
| TDP | 300 W |
| Form factor | PCIe Gen4 x16, triple-slot, active |
| Display outputs | 4× DisplayPort 1.4a |
| NVLink | Not supported |
When to Pick RTX 6000 Ada#
- Desk-side AI development — fine-tuning 7B models, prototyping inference, profiling kernels.
- Professional content creation (animation, VFX, CAD) where Ada RT performance matters.
- Single-user scientific computing with FP64-light workloads.
- Edge AI appliances where a workstation chassis is acceptable.
- Pick L40S for data-centre deployment of the same silicon.
- Pick a GeForce RTX 5090-class card if you don't need ECC, display drivers or warranty.
Pitfalls#
- Three-slot form factor blocks adjacent PCIe slots in many workstations.
- 300 W active card needs reliable case airflow; thermal throttling can hide as performance variance.
- No NVLink — multi-card configurations rely on PCIe.
- GDDR6 bandwidth limits KV-cache-heavy LLM inference relative to HBM cards.
Software Notes#
Full Studio Driver and NVIDIA Enterprise Driver support, certified for major DCC and CAD applications. AI inference paths (vLLM, TensorRT-LLM, Ollama, Llama.cpp with CUDA) all run unchanged; FP8 paths require recent (2024+) inference server releases.
References
- NVIDIA RTX 6000 Ada Datasheet · NVIDIA