Memory Bandwidth per Dollar
Formula: memory bandwidth (GB/s) ÷ launch MSRP ($) × 1000.
Price class: launch MSRP (USD, each row links its source). Launch MSRP is historical — it is not a current street price, and we publish no used-market prices because none are verified. See methodology.
Price class: launch MSRP (USD, each row links its source). Launch MSRP is historical — it is not a current street price, and we publish no used-market prices because none are verified. See methodology.
GPUs ranked by memory bandwidth per dollar of launch MSRP — a spec-level value proxy for bandwidth-bound LLM token generation, not a measured performance ranking. Computed live from our database — every row's launch price links to its source.
| # | GPU | GB/s per $1000 | Memory GB | Bandwidth | TDP | Launch MSRP | Year |
|---|---|---|---|---|---|---|---|
| 1 | Arc A580 | 2860.34 | 8 GB | 512 GB/s | 185 W | $179 src | 2023 |
| 2 | Arc B580 | 1831.33 | 12 GB | 456 GB/s | 190 W | $249 src | 2024 |
| 3 | Arc A750 | 1771.63 | 8 GB | 512 GB/s | 225 W | $289 src | 2022 |
| 4 | Arc B570 | 1735.16 | 10 GB | 380 GB/s | 150 W | $219 src | 2025 |
| 5 | Arc A770 16GB | 1604.58 | 16 GB | 560 GB/s | 225 W | $349 src | 2022 |
| 6 | GeForce RTX 5060 | 1498.33 | 8 GB | 448 GB/s | 145 W | $299 src | 2025 |
| 7 | GeForce RTX 5050 | 1285.14 | 8 GB | 320 GB/s | 130 W | $249 src | 2025 |
| 8 | Radeon RX 7800 XT | 1250.5 | 16 GB | 624 GB/s | 263 W | $499 src | 2023 |
| 9 | Radeon RX 7900 XT | 1232.67 | 20 GB | 800 GB/s | 315 W | $649 src | 2022 |
| 10 | GeForce RTX 5070 | 1224.04 | 12 GB | 672 GB/s | 250 W | $549 src | 2025 |
| 11 | GeForce RTX 5070 Ti | 1196.26 | 16 GB | 896 GB/s | 300 W | $749 src | 2025 |
| 12 | Radeon RX 9070 | 1165.76 | 16 GB | 640 GB/s | 220 W | $549 src | 2025 |
| 13 | GeForce RTX 3060 12GB | 1094.22 | 12 GB | 360 GB/s | 170 W | $329 src | 2021 |
| 14 | GeForce RTX 3080 10GB | 1087.27 | 10 GB | 760 GB/s | 320 W | $699 src | 2020 |
| 15 | Radeon RX 7600 | 1070.63 | 8 GB | 288 GB/s | 165 W | $269 src | 2023 |
| 16 | Radeon RX 9070 XT | 1068.45 | 16 GB | 640 GB/s | 304 W | $599 src | 2025 |
| 17 | GeForce RTX 5060 Ti 16GB | 1044.29 | 16 GB | 448 GB/s | 180 W | $429 src | 2025 |
| 18 | Radeon RX 7700 XT | 1031.03 | 12 GB | 432 GB/s | 245 W | $419 src | 2023 |
| 19 | GeForce RTX 5080 | 960.96 | 16 GB | 960 GB/s | 360 W | $999 src | 2025 |
| 20 | Radeon RX 7900 XTX | 960.96 | 24 GB | 960 GB/s | 355 W | $999 src | 2022 |
| 21 | GeForce RTX 4070 | 918.03 | 12 GB | 504 GB/s | 200 W | $549 src | 2023 |
| 22 | GeForce RTX 4060 | 909.7 | 8 GB | 272 GB/s | 115 W | $299 src | 2023 |
| 23 | GeForce RTX 3070 | 897.8 | 8 GB | 448 GB/s | 220 W | $499 src | 2020 |
| 24 | GeForce RTX 5090 | 896.45 | 32 GB | 1792 GB/s | 575 W | $1,999 src | 2025 |
| 25 | Radeon RX 7600 XT | 875.38 | 16 GB | 288 GB/s | 190 W | $329 src | 2024 |
| 26 | GeForce RTX 4070 SUPER | 841.4 | 12 GB | 504 GB/s | 220 W | $599 src | 2024 |
| 27 | GeForce RTX 4070 Ti SUPER | 841.05 | 16 GB | 672 GB/s | 285 W | $799 src | 2024 |
| 28 | Radeon RX 6800 XT | 788.91 | 16 GB | 512 GB/s | 300 W | $649 src | 2020 |
| 29 | GeForce RTX 3080 Ti | 760.63 | 12 GB | 912 GB/s | 350 W | $1,199 src | 2021 |
| 30 | GeForce RTX 4080 SUPER | 736.74 | 16 GB | 736 GB/s | 320 W | $999 src | 2024 |
| 31 | GeForce RTX 4060 Ti 8GB | 721.8 | 8 GB | 288 GB/s | 160 W | $399 src | 2023 |
| 32 | GeForce RTX 4090 | 630.39 | 24 GB | 1008 GB/s | 450 W | $1,599 src | 2022 |
| 33 | GeForce RTX 3090 | 624.42 | 24 GB | 936 GB/s | 350 W | $1,499 src | 2020 |
| 34 | GeForce RTX 2080 Ti | 616.62 | 11 GB | 616 GB/s | 250 W | $999 src | 2018 |
| 35 | GeForce RTX 4080 | 597.83 | 16 GB | 716.8 GB/s | 320 W | $1,199 src | 2022 |
| 36 | GeForce RTX 4060 Ti 16GB | 577.15 | 16 GB | 288 GB/s | 160 W | $499 src | 2023 |
| 37 | GeForce RTX 3090 Ti | 504.25 | 24 GB | 1008 GB/s | 450 W | $1,999 src | 2022 |
| 38 | NVIDIA RTX 5000 Ada | 256 | 32 GB | 576 GB/s | 250 W | $2,250 src | 2023 |
| 39 | AMD Radeon Pro W7800 | 230.49 | 32 GB | 576 GB/s | 260 W | $2,499 src | 2022 |
| 40 | AMD Radeon Pro W7900 | 216.05 | 48 GB | 864 GB/s | 295 W | $3,999 src | 2022 |
| 41 | M3 Ultra (Mac Studio) | 204.8 | 512 GB | 819 GB/s | 270 W | $3,999 src | 2025 |
| 42 | M4 Max (MacBook Pro) | 170.68 | 128 GB | 546 GB/s | — | $3,199 src | 2024 |
| 43 | RTX A6000 | 170.67 | 48 GB | 768 GB/s | 300 W | $4,500 src | 2021 |
| 44 | RTX 6000 Ada | 141.18 | 48 GB | 960 GB/s | 300 W | $6,800 src | 2023 |
Method: this is a spec-level ranking, not a measured performance ranking — scores divide one database spec by launch MSRP (or rated TDP). Only GPUs with a source-linked launch price appear; 20 GPUs are excluded for lacking sourced pricing. These are launch MSRPs, not current street prices; no used-market prices are tracked, so none are shown. Unified-memory systems list GPU-addressable unified memory, not discrete VRAM. See methodology.
Measured performance per dollar — Llama-3.1-8B (Q4_K_M), llama.cpp
The only true performance-per-dollar ranking on this page: real measured tokens/s divided by launch MSRP. It exists only where sourced, non-estimated benchmark rows exist — estimated rows never enter this table.
| # | GPU | Measured tok/s | tok/s per $1000 launch MSRP | Software stack | Source |
|---|---|---|---|---|---|
| 1 | Arc B580 | 41 tok/s | 164.7 | llama.cpp (Vulkan) | src |
| 2 | GeForce RTX 4070 SUPER | 92 tok/s | 153.6 | llama.cpp b3500 (CUDA) | src |
| 3 | GeForce RTX 5080 | 132 tok/s | 132.1 | llama.cpp via Ollama | src |
| 4 | GeForce RTX 3060 12GB | 38 tok/s | 115.5 | llama.cpp b3500 (CUDA) | src |
| 5 | GeForce RTX 5090 | 215 tok/s | 107.6 | llama.cpp b3500 (CUDA) | src |
| 6 | GeForce RTX 4080 SUPER | 102 tok/s | 102.1 | llama.cpp b3500 (CUDA) | src |
| 7 | GeForce RTX 4090 | 125 tok/s | 78.2 | llama.cpp b3500 (CUDA) | src |
| 8 | Radeon RX 7900 XTX | 76 tok/s | 76.1 | llama.cpp b3500 (ROCm) | src |
| 9 | GeForce RTX 3090 | 85 tok/s | 56.7 | llama.cpp b3500 (CUDA) | src |
| 10 | RTX 6000 Ada | 110 tok/s | 16.2 | llama.cpp b3500 (CUDA) | src |
Method: measured tokens/s ÷ launch MSRP × 1000 — price class launch MSRP, not current street price. Single-stream Llama-3.1-8B Q4_K_M token generation from sourced benchmark rows; software stacks differ (CUDA/ROCm/Metal/Vulkan/Ollama), so treat cross-stack gaps as approximate rather than like-for-like. No used-market prices are tracked, so no used-price performance ranking exists.
What to do with these rankings
High VRAM capacity per dollar means more model capacity for the money — start there when deciding which models fit your budget. High memory bandwidth per dollar tracks the spec that most limits token generation speed. High memory bandwidth per TDP watt suits always-on servers and small-form-factor builds.
Then compare finalists head-to-head in the GPU comparison tool, check which models fit each card, or browse all sourced benchmark results.