NVIDIA RTX 4080

NVIDIA RTX 4080: Reviewers were broadly critical of this card's $1,199 launch price, a real jump from the $699 the RTX 3080 and RTX 2080 each launched at in the two generations before it (widely covered in launch-week press). The more durable story became clear fourteen months later: the RTX 4070 Ti SUPER launched January 24, 2024 at $799, on the identical AD103 die with this card's same 16GB capacity and same 256-bit bus, giving up only 1,280 CUDA cores (8,448 vs 9,728) and 45 GB/s of bandwidth (672 vs 717 GB/s, a 6.7% gap) to get there for $400 less. NVIDIA replaced this card outright the same month with the RTX 4080 SUPER, a near-full AD103 configuration (10,240 CUDA cores) at $999, $200 under this card's original price.

This site's calculator puts Qwen3 8B at its recommended Q5_K_M quant at 76.3 tok/s (7.73 GB, fits with room for context) and Qwen3 14B at its recommended Q5_K_M at 47.1 tok/s (13.31 GB, fits but with only about 1.9 GB of headroom left in the roughly 15.2 GB this card actually has usable). GPT-OSS 20B, a model OpenAI explicitly sized to fit a 16GB card, fits at its recommended Q4_K_M (14.55 GB, 173.4 tok/s), with barely 0.6 GB of that usable budget left over. Qwen 3.6 27B draws the real ceiling: its recommended Q4_K_M build (19.02 GB) overshoots this card's usable VRAM and drops to a CPU-offloaded 10.1 tok/s, but the next quant down, Q3_K_M (15.15 GB), still fits natively at 39.6 tok/s, far below GPT-OSS 20B's 173.4 on the same card, which activates only 3.6B parameters per token. Qwen3 32B doesn't fit natively at any quantization on this site's standard ladder: even Q3_K_M needs 19.17 GB against roughly 15.2 GB of usable VRAM, landing at 9.6 tok/s offloaded rather than a native fit, so 27B-class is closer to this card's real dense-model ceiling than the 32B some marketing copy implies. GPT-OSS 120B doesn't fit at all, even with full system-RAM offload: its 62.6 GB MXFP4 weight floor alone is well past this card's roughly 41 GB usable VRAM+RAM budget. Measured against the RTX 4090 at the identical Q5_K_M quant, Qwen3 8B decodes at 100.6 tok/s there versus 76.3 tok/s here, a 40.6% gap that tracks the two cards' bandwidth difference (1,008 vs 717 GB/s, 40.6%) almost exactly, since decode is bandwidth-bound. Against the RTX 4070 Ti SUPER at the same Q5_K_M quant, the gap is much smaller, 67.5 vs 72.3 tok/s, a 6.6% edge closely matching the two cards' bandwidth difference (6.7%): the whole practical difference between paying $1,199 for this card and $799 for that one.

Full CUDA support on Ada Lovelace, the same mature llama.cpp, Ollama, vLLM, and TensorRT-LLM coverage every Ada card in this site's lineup gets, well ahead of Blackwell's rockier early sm_120 rollout. This card's 4th-gen Tensor Cores add hardware FP8 support, accelerated natively by TensorRT-LLM and vLLM on Ada and Hopper, but not NVFP4: that 4-bit format is Blackwell-only in this site's calculator, so Q2_K remains this card's smallest quant on the standard ladder. On value, the RTX 4070 Ti SUPER is the sharper comparison than the raw specs suggest: identical 16GB capacity on the identical AD103 die, giving up just 6.7% of this card's bandwidth (672 vs 717 GB/s) for $400 less at launch, and since local decode speed is bandwidth-bound, that 6.7% gap (67.5 vs 72.3 tok/s on Qwen3 8B at Q5_K_M above) is close to the entire practical difference between the two cards for LLM inference specifically. The RTX 4060 Ti 16GB undercuts both further at the same 16GB ceiling, but on a much narrower 288 GB/s bus: this site's calculator puts the same Qwen3 8B build at 34 tok/s there, about 40% of this card's speed, tracking the 288-vs-717 GB/s ratio (40.2%) closely. The same models fit either card; they just decode far more slowly on the cheaper one.