RTX 3090
NVIDIAconsumer| Overview | |
|---|---|
| Architecture | Ampere (2020) |
| Launch date | Sep 24, 2020 |
| Compute capability | 8.6 |
| Memory | |
| VRAM | 24 GB GDDR6X |
| Memory bandwidth | 936 GB/s |
| Cache | 6 MB |
| Compute — vector | |
| FP32 | 35.6 TFLOPS |
| Compute — matrix / tensor | |
| TF32 | 35.6 TFLOPS |
| BF16 | 71.2 TFLOPS |
| FP16 (dense / sparse) | 71.2 / 142.4 TFLOPS |
| INT8 (dense / sparse) | 284.7 / 569.4 TOPS |
| Cores & clocks | |
| Streaming Multiprocessors | 82 |
| Shader cores | 10,496 |
| Matrix / Tensor cores | 328 |
| Board & system | |
| TDP | 350 W |
| PCIe | PCIe Gen4 |
| Interconnect | NVLink — 112.5 GB/s |
| MSRP | $1,499 |
Compatible CUDA Toolkit versions
Added 2026-09 alongside V100/P100/T4 — RTX 3090 was this site's own README-flagged example of a missing older-but-still-common AI/ML GPU (it remains popular for local LLM inference/fine-tuning due to its 24GB VRAM at consumer pricing). Core compute figures, CUDA/Tensor/SM counts, memory config, L2 cache, and TGP are all read directly from NVIDIA's own "NVIDIA RTX Blackwell GPU Architecture" whitepaper (images.nvidia.com/ aem-dam/Solutions/geforce/blackwell/nvidia-rtx-blackwell-gpu-architecture.pdf, Appendix A, Table 3: "GeForce RTX 5090 vs GeForce RTX 4090 vs GeForce RTX 3090 Specs" — the same table this site's RTX 4090 entry already cites for its own figures, extended here to its RTX 3090 column), read via pdftotext, 2026-09. Board-level fields (boost clock, memory config, PCIe gen, TGP) cross-checked against nvidia.com's own current RTX 3090 product spec page and matched exactly.
fp32TFLOPS/tf32TFLOPS (35.6) match because Ampere's non-Tensor FP32 and dense TF32-Tensor rate are numerically identical on this die, same pattern already documented on this site's RTX 4090 entry; fp16TFLOPSVector is omitted for the same reason (whitepaper's "Peak FP16 TFLOPS (non-Tensor)" is also 35.6 — redundant with fp32TFLOPS). bf16TFLOPS/ fp16TFLOPS (71.2 dense, 142.4 sparse) use the whitepaper's "...with FP32 Accumulate" row, matching this site's established convention (see the RTX 4090 entry's notes) rather than the higher "...with FP16 Accumulate" row (142.3/284.6) also present in the same table. fp8TFLOPS/fp4TFLOPS are correctly omitted, not zero — the whitepaper lists both as "N/A" for RTX 3090 (FP8 Tensor arrived with Ada, FP4 with Blackwell). int4TOPS is likewise not published for this generation.
interconnectBandwidthGBs (112.5, NVLink) is a well-corroborated secondary-source figure (multiple independent reviewer/community sources), not read off an NVIDIA table directly — RTX 3090 is one of the few GeForce cards with an NVLink bridge connector (dropped starting with 3090 Ti), and NVIDIA's own marketing material confirms NVLink support without stating the exact GB/s figure. msrpUSD (1499) is the original Sept 2020 Founders Edition launch price, widely corroborated across contemporaneous outlets rather than quoted verbatim from NVIDIA's own launch page (which confirms the launch but not the price in the specific text pulled). No transistor count or process node — not in the sourced whitepaper table (publicly reported elsewhere as Samsung 8N, but not confirmed here directly, so omitted rather than assumed).