K80
NVIDIAdatacenter| Overview | |
|---|---|
| Architecture | Kepler (2014) |
| Launch date | Nov 2014 (GA); board spec revised Jan 2015 |
| Compute capability | 3.7 |
| Memory | |
| VRAM | 24 GB GDDR5 |
| Memory bandwidth | 480 GB/s |
| ECC | Yes |
| Compute — vector | |
| FP64 | 2.91 TFLOPS |
| FP32 | 8.74 TFLOPS |
| Cores & clocks | |
| Streaming Multiprocessors | 26 |
| Shader cores | 4,992 |
| Board & system | |
| TDP | 300 W |
| Form factor | Dual-GPU PCIe, Full Height |
| PCIe | PCIe Gen3 |
Compatible CUDA Toolkit versions
Added 2026-09 alongside M40 and older AMD Radeon Instinct cards — K80 predates this site's previous oldest NVIDIA entry (2017's V100) by three years and, despite its age, is still occasionally encountered on legacy training infrastructure. Board-level figures (FP64/FP32 TFLOPS with GPU Boost, memory, bandwidth, power) sourced from NVIDIA's official "Tesla K80" overview PDF (nvidia.com/content/dam/en-zz/Solutions/Data-Center/ tesla-product-literature/nvidia-tesla-k80-overview.pdf); mechanical/power details (form factor, ECC, ~300W max board power) cross-checked against NVIDIA's official Board Specification document (BD-07317-001_v05, Jan 2015, nvidia.com/content/dam/en-zz/Solutions/Data-Center/ tesla-product-literature/Tesla-K80-BoardSpec-07317-001-v05.pdf). Both read via pdftotext.
K80 is a genuinely dual-GPU board — two Tesla GK210 dies connected via an onboard PLX PCIe switch, each with 12GB GDDR5 (24GB total) and 2,496 CUDA cores (4,992 combined, the figure in shaderCoreCount here) — unlike every other entry in this collection, which is a single die. NVIDIA sells and benchmarks the K80 as one board-level SKU, so fp64TFLOPS (2.91) and fp32TFLOPS (8.74) are the combined-board figures "with NVIDIA GPU Boost" from the overview PDF, not per-GPU. There is no separate host-facing NVLink/interconnect field populated — the board's only external interface is the single PCIe Gen3 x16 link shared by both dies; the PLX switch routing between them isn't modeled as an `interconnect` value.
targetId ("3.7", GK210's actual compute capability) is set for correctness. As of 2026-09 this now cross-links to the 10.0 and 10.1 CUDA entries (both list "3.7"), which predate this site's Maxwell-only 10.2+ entries — Kepler support was deprecated in 10.2 and dropped entirely in 11.0, so any current-day CUDA Toolkit genuinely cannot target Kepler. No process node, transistor count, or MSRP found in either sourced document — K80 shipped through OEM/server partners, not direct retail.
2026-09-15 — added computeUnitCount (26 SMX blocks). NVIDIA's K80 datasheet publishes only the 4,992 CUDA-core board total; 26 follows from Kepler's fixed 192 FP32 cores per SMX (4,992 / 192 = 26) and matches the board configuration reported at launch — two GK210 dies with 13 of each die's 15 SMX blocks enabled. Kepler calls the block an SMX; computeUnitLabel says "Streaming Multiprocessors" for consistency with the rest of this collection's NVIDIA entries. As with every other figure in this file, 26 is the whole-board number: software sees two 13-SMX devices, not one 26-SMX device.