RTX 6000 Ada Generation
NVIDIAworkstation| Overview | |
|---|---|
| Architecture | Ada Lovelace (2022) |
| Process node | TSMC 4N |
| Transistor count | 76.3B |
| Compute capability | 8.9 |
| Memory | |
| VRAM | 48 GB GDDR6 (ECC) |
| Memory bandwidth | 960 GB/s |
| Compute — vector | |
| FP32 | 91.1 TFLOPS |
| Compute — matrix / tensor | |
| BF16 | 364.25 TFLOPS |
| FP16 (dense / sparse) | 364.25 / 728.5 TFLOPS |
| FP8 (dense / sparse) | 728.5 / 1,457 TFLOPS |
| Cores & clocks | |
| Streaming Multiprocessors | 142 |
| Shader cores | 18,176 |
| Matrix / Tensor cores | 568 |
| Board & system | |
| TDP | 300 W |
| Form factor | PCIe dual-slot |
| Cooling | Active |
| PCIe | PCIe Gen4 |
| MSRP | $6,800 |
Compatible CUDA Toolkit versions
fp32TFLOPS (91.1) confirmed 2026-09-01 directly from NVIDIA's current product page (nvidia.com/en-us/products/workstations/rtx-6000/, "Single-Precision Performance: 91.1 TFLOPS"). fp8TFLOPS/fp16TFLOPS remain derived, not directly published as separate dense figures: the same product page states "Tensor Performance: 1,457 AI TOPS" footnoted as "Theoretical FP8 TOPS using the sparsity feature" — halving gives dense FP8 (728.5, → fp8TFLOPSSparse: 1457), and halving again gives an estimated dense FP16 (364.25, → fp16TFLOPSSparse: 728.5), cross-checked against the L40S entry (same AD102 die, same 362.05/733 dense FP16/FP8 split) as a sanity check. bf16TFLOPS assumed equal to fp16TFLOPS — not separately published, same assumption made on the L40S entry for the same die family. No tf32TFLOPS — unlike the RTX 4090/5090 whitepaper, no source was found confirming TF32 dense equals the FP32 vector rate specifically for this SKU, so it's omitted rather than assumed.
shaderCoreCount (18,176 CUDA cores), matrixCoreCount (568 4th-gen Tensor Cores), transistorCountBillion (76.3), and processNode ("TSMC 4N") come from a secondary source (an NVIDIA-authored datasheet mirrored by a cloud reseller, acecloud.ai) after NVIDIA's own gated datasheet link (resources.nvidia.com) and the mirror's own compute/TFLOPS table proved corrupted/unreliable on text extraction (looked like leftover Hopper template values, not RTX 6000 Ada's actual figures — discarded rather than used). The core counts, transistor count, and process are treated as reliable despite the secondary source because they're internally consistent (18,176 / 128 CUDA-cores-per-SM = 142 SMs exactly, used as computeUnitCount) and match AD102's well-established public transistor budget shared with the RTX 4090 entry. cacheMB and peakClockMHz omitted — not found in any source consulted.
formFactor, coolingType, pcieGen, and tdpWatts (300W) are directly from NVIDIA's product page Specifications panel. msrpUSD ($6,800) remains an approximate street/channel price, not a datasheet-listed figure — RTX 6000 Ada sells through workstation partners rather than direct retail.