NVIDIA RTX 4070 Ti SUPER

NVIDIA RTX 4070 Ti SUPER: NVIDIA's January 2024 Super refresh gave the RTX 4070 Ti a real memory upgrade, not just a clock bump: this card moves off the RTX 4070 Ti's AD104 die onto the larger AD103-275 die (45.9 billion transistors, up from AD104's 35.8 billion) that also powers the RTX 4080, gaining 4GB of VRAM and a wider 256-bit memory bus in the process (NVIDIA's own GeForce RTX 40 SUPER announcement; Wikipedia's GeForce RTX 40 series spec table cross-checks every figure here). Launched January 24, 2024 at $799, it landed $200 under the RTX 4080's $999 launch price for the exact same 16GB capacity, trading only 45 GB/s of bandwidth (672 vs 717 GB/s) and 1,280 fewer CUDA cores (8,448 vs 9,728) to get there. GamersNexus's launch review measured a real but modest 10-11% gaming uplift over the plain RTX 4070 Ti at 4K, a gap that narrows further at lower resolutions.

This site's calculator puts Llama 3.1 8B at 45.6 tok/s at Q8_0 (10.73 GB, fits with room to spare) and GPT-OSS 20B, a model OpenAI explicitly sized for a 16GB card, at 58.1 tok/s at Q4_K_M (14.55 GB). Both figures land within 10% of the RTX 4080's numbers for the same models (48.7 and 62 tok/s), tracking the two cards' 717-vs-672 GB/s bandwidth gap almost exactly, since decode is bandwidth-bound. The jump over this card's own predecessor is much larger: the plain RTX 4070 Ti needs CPU offload for GPT-OSS 20B (25.23 GB spills past its 12GB ceiling, 3.6 tok/s) and for Qwen 3.6 27B (32.75 GB, 1.3 tok/s), while this card runs both natively, GPT-OSS 20B at 58.1 tok/s and Qwen 3.6 27B at 32.3 tok/s (Q3_K_M, 15.15 GB). Measured against the Blackwell RTX 5070 Ti at the identical Q3_K_M quant, Qwen 3.6 27B decodes at 43.1 tok/s there versus 32.3 tok/s here, a 33.4% gap that matches the two cards' 896-vs-672 GB/s bandwidth difference (33.3%) almost exactly, since both fit the same models at the same quant and the only real difference is bandwidth.

Full CUDA support on Ada Lovelace, a mature target for llama.cpp, Ollama, vLLM, and TensorRT-LLM alike, unlike 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 format is Blackwell-only in this site's calculator, so Q2_K remains the smallest quant on this card's standard ladder, unlike the RTX 5070 Ti at the same 16GB. On value, this is the only card in the RTX 4070 family whose bandwidth actually improved over the base RTX 4070 and RTX 4070 Ti; the RTX 4070 SUPER released the same week kept the identical 504 GB/s memory subsystem as both, gaining only compute. For anyone comparing this card against its direct Blackwell successor, the RTX 5070 Ti, the honest baseline is this card's 672 GB/s, not the original RTX 4070 Ti's 504 GB/s: measured from here, the 5070 Ti's bandwidth gain is a smaller 33.3%, not the 77.8% a comparison against the original RTX 4070 Ti would suggest.