GTX 1080 Ti
NVIDIAconsumer| Overview | |
|---|---|
| Architecture | Pascal (2017) |
| Launch date | Mar 10, 2017 (announced Feb 28 at GDC) |
| Transistor count | 12B |
| Compute capability | 6.1 |
| Memory | |
| VRAM | 11 GB GDDR5X |
| Memory bandwidth | 484 GB/s |
| Compute — vector | |
| FP32 | 11.3 TFLOPS |
| Cores & clocks | |
| Streaming Multiprocessors | 28 |
| Shader cores | 3,584 |
| Peak clock | 1,582 MHz |
| Board & system | |
| TDP | 250 W |
| PCIe | PCIe Gen3 x16 |
| MSRP | $699 |
Compatible CUDA Toolkit versions
Added 2026-09 — the user asked for older Titan/GTX cards specifically; GTX 1080 Ti was one of the most widely adopted early deep-learning GPUs thanks to strong FP32 throughput per dollar, predating this collection's previous oldest consumer-tier Pascal-class part. Core facts (3,584 CUDA cores, 11GB GDDR5X at 11Gbps, 12 billion transistors, $699 launch price, March 10 2017 availability) are from NVIDIA's own official press release (nvidianews.nvidia.com, "NVIDIA Introduces the Beastly GeForce GTX 1080 Ti", Feb 28 2017), read via pdftotext.
memoryBandwidthGBs (484), peakClockMHz (1582 boost), tdpWatts (250), and the 352-bit memory bus width (not itself stored as a field) are not in that press release's own text but are the same figures NVIDIA distributed on its press spec sheet at the same launch event, reported consistently by outlets present at the briefing (e.g. GamersNexus's "Official nVidia GTX 1080 Ti Specs" writeup) and corroborated unanimously across independent GPU databases — treated as reliable despite not being quoted verbatim in the press-release PDF itself. fp32TFLOPS (11.3) is derived from those same clock/core figures using this site's standard vector-throughput formula (cores × boost clock × 2), not independently published as a headline TFLOPS number by NVIDIA for this SKU. No Tensor/ matrix-throughput fields — Pascal predates Tensor Cores (introduced with Volta the following year), and GP102 has no dedicated FP16 Matrix path (fp16TFLOPSVector also omitted — not stated at 2x rate for this gaming-tier Pascal die, unlike Tesla P100's GP100).
2026-09-15 — added computeUnitCount (28 SMs). NVIDIA publishes the 3,584 CUDA-core total (already in shaderCoreCount) but not the SM count; 28 follows from Pascal's fixed 128 FP32 cores per SM on the consumer GP10x dies (3,584 / 128 = 28) — GP102 with 28 of its 30 SMs enabled. matrixCoreCount stays correctly empty: no Tensor Cores on Pascal.