TITAN RTX
NVIDIAconsumer| Overview | |
|---|---|
| Architecture | Turing (2018) |
| Launch date | Dec 2018 |
| Compute capability | 7.5 |
| Memory | |
| VRAM | 24 GB GDDR6 |
| Memory bandwidth | 672 GB/s |
| Compute — vector | |
| FP32 | 16.3 TFLOPS |
| Compute — matrix / tensor | |
| FP16 | 130 TFLOPS |
| Cores & clocks | |
| Streaming Multiprocessors | 72 |
| Shader cores | 4,608 |
| Matrix / Tensor cores | 576 |
| Peak clock | 1,770 MHz |
| Board & system | |
| TDP | 280 W |
| PCIe | PCIe Gen3 x16 |
| Interconnect | NVLink (optional bridge) — 100 GB/s |
| MSRP | $2,499 |
Compatible CUDA Toolkit versions
Added 2026-09 alongside GTX 1080 Ti, TITAN X (Pascal), and TITAN V — the last of NVIDIA's consumer/prosumer TITAN line (no successor was released after Ampere), and notable for pairing Turing's Tensor Cores with 24GB of VRAM, unusually large for a non-datacenter card at the time. CUDA core count (4,608), Tensor core count (576), boost clock (1,770 MHz), memory config (24GB GDDR6), memory bandwidth (672 GB/s), Tensor performance ("130 Tensor TFLOPs"), and the optional NVLink bridge (100 GB/s, doubling effective VRAM to 48GB across two cards) are all quoted directly from NVIDIA's own official product page (nvidia.com/en-us/titan/titan-rtx/), read 2026-09.
fp32TFLOPS (16.3) and tdpWatts (280) aren't on that specific page but are consistently corroborated across independent GPU databases and contemporaneous press coverage of the Dec 2018 launch. int8TOPS is intentionally omitted, not guessed — secondary sources disagree sharply on this figure (some report 130 TOPS, equal to the FP16 Tensor rate; others report roughly double), and no NVIDIA-published number was found to resolve the conflict. msrpUSD (2499) is the original launch price, well-corroborated but not itself quoted on the current product page (it no longer lists a price). RT Core count/TFLOPS are published but out of scope for this AI/ML-focused schema, matching this collection's existing convention (see RTX 4090's notes for the same treatment).
2026-09-15 — added computeUnitCount (72 SMs). NVIDIA's TITAN RTX page gives 4,608 CUDA cores, 576 Tensor Cores and 72 RT Cores; on Turing's fixed per-SM structure all three independently give 72 SMs (4,608 / 64, 576 / 8, and one RT Core per SM). Recorded as a derived value rather than left blank because three published figures agree on it.