20 GPUs worth comparing.
Gaming and local AI ask different things of a graphics card. This shortlist puts memory, power, and practical tradeoffs side by side—without pretending one GPU is best for everyone.
| GPU / sources | Memory | Board power | Gaming considerations | Price / stock |
|---|---|---|---|---|
GeForce RTX 5090NVIDIA · BlackwellAvailability and US retailer pricing have not been checked. Confirm exact board model, warranty, dimensions, cooling, and connectors before purchase. power_w is NVIDIA reference board power, 575 W (Total Graphics Power / TGP), not typical gaming consumption or a whole-system PSU recommendation; partner boards may differ. Memory interface: 512-bit [3]. Reference GPU specs; board-partner dimensions, cooling, power limits, and connectors can differ. |
32 GBGDDR7 [3] | 575 WReference board [3] | Consider for demanding high-resolution gaming; assess game-specific native-rendering benchmarks rather than treating generated frames as equivalent. Largest VRAM capacity in this shortlist: 32 GB. A capacity-first CUDA candidate, but check Blackwell runtime support and the 575 W reference power requirement. [1] [2] | Not connectedNo price / stock feed |
GeForce RTX 5080NVIDIA · Blackwell |
16 GBGDDR7 [4] | 360 WReference board [4] | Consider for high-resolution gaming; validate ray-tracing demands and settings with game-specific benchmarks. 16 GB CUDA candidate. Choose the exact model, quantization and context first; equal VRAM does not mean equal inference speed. Check Blackwell runtime support. [1] [2] | Not connectedNo price / stock feed |
GeForce RTX 5070 TiNVIDIA · BlackwellAvailability and US retailer pricing have not been checked. Confirm exact board model, warranty, dimensions, cooling, and connectors before purchase. power_w is NVIDIA reference board power, 300 W (Total Graphics Power / TGP), not typical gaming consumption or a whole-system PSU recommendation; partner boards may differ. Memory interface: 256-bit [5]. Reference GPU specs; board-partner dimensions, cooling, power limits, and connectors can differ. |
16 GBGDDR7 [5] | 300 WReference board [5] | Consider for 1440p or higher-resolution gaming after checking the intended games and settings. 16 GB CUDA candidate. Choose the exact model, quantization and context first; equal VRAM does not mean equal inference speed. Check Blackwell runtime support. [1] [2] | Not connectedNo price / stock feed |
GeForce RTX 5070NVIDIA · BlackwellAvailability and US retailer pricing have not been checked. Confirm exact board model, warranty, dimensions, cooling, and connectors before purchase. power_w is NVIDIA reference board power, 250 W (Total Graphics Power / TGP), not typical gaming consumption or a whole-system PSU recommendation; partner boards may differ. Memory interface: 192-bit [5]. Reference GPU specs; board-partner dimensions, cooling, power limits, and connectors can differ. |
12 GBGDDR7 [5] | 250 WReference board [5] | Consider for 1440p gaming; demanding textures and ray tracing warrant game-specific checks. 12 GB CUDA candidate. Choose the exact model, quantization and context first; equal VRAM does not mean equal inference speed. Check Blackwell runtime support. [1] [2] | Not connectedNo price / stock feed |
GeForce RTX 5060 Ti 16GBNVIDIA · BlackwellAvailability and US retailer pricing have not been checked. Confirm exact board model, warranty, dimensions, cooling, and connectors before purchase. power_w is NVIDIA reference board power, 180 W (Total Graphics Power / TGP), not typical gaming consumption or a whole-system PSU recommendation; partner boards may differ. Memory interface: 128-bit [6]. Reference GPU specs; board-partner dimensions, cooling, power limits, and connectors can differ. |
16 GBGDDR7 [6] | 180 WReference board [6] | Only the 16 GB variant is cataloged; additional memory is not a guarantee of higher frame rates. 16 GB CUDA candidate. Choose the exact model, quantization and context first; equal VRAM does not mean equal inference speed. Check Blackwell runtime support. [1] [2] | Not connectedNo price / stock feed |
GeForce RTX 5060NVIDIA · BlackwellAvailability and US retailer pricing have not been checked. Confirm exact board model, warranty, dimensions, cooling, and connectors before purchase. power_w is NVIDIA reference board power, 145 W (Total Graphics Power / TGP), not typical gaming consumption or a whole-system PSU recommendation; partner boards may differ. Memory interface: 128-bit [6]. Reference GPU specs; board-partner dimensions, cooling, power limits, and connectors can differ. |
8 GBGDDR7 [6] | 145 WReference board [6] | Consider for 1080p gaming with settings tailored to the game; check texture-memory demands. 8 GB leaves less memory headroom. Start with smaller quantized workloads; confirm model, context and runtime allocations before purchase. [1] [2] | Not connectedNo price / stock feed |
GeForce RTX 4090NVIDIA · Ada LovelaceOlder-generation option; verify the exact listing, seller, condition, warranty, cooling and connectors. No current price or availability checked. Board-partner specifications may differ. Memory interface: 384-bit [7]. Reference GPU specs; board-partner dimensions, cooling, power limits, and connectors can differ. |
24 GBGDDR6X [7] | 450 WReference board [7] | A previous-generation high-end option to evaluate for demanding gaming; inspect cooling, dimensions, and game-specific benchmarks. 24 GB CUDA candidate. More memory headroom than the 16 GB cards, but a substantial 450 W reference board-power requirement. [1] [2] | Not connectedNo price / stock feed |
GeForce RTX 4080 SUPERNVIDIA · Ada LovelaceOlder-generation option; verify the exact listing, seller, condition, warranty, cooling and connectors. No current price or availability checked. Board-partner specifications may differ. Memory interface: 256-bit [8]. Reference GPU specs; board-partner dimensions, cooling, power limits, and connectors can differ. |
16 GBGDDR6X [8] | 320 WReference board [8] | Evaluate for high-resolution gaming against alternatives using actual game settings and benchmarks, not the SUPER label alone. 16 GB CUDA candidate. Choose the exact model, quantization and context first; equal VRAM does not mean equal inference speed. [1] [2] | Not connectedNo price / stock feed |
GeForce RTX 4070 Ti SUPERNVIDIA · Ada LovelaceOlder-generation option; verify the exact listing, seller, condition, warranty, cooling and connectors. No current price or availability checked. Board-partner specifications may differ. Memory interface: 256-bit [9]. Reference GPU specs; board-partner dimensions, cooling, power limits, and connectors can differ. |
16 GBGDDR6X [9] | 285 WReference board [9] | Consider for 1440p and higher-resolution gaming after checking game-specific performance and settings. 16 GB CUDA candidate. Choose the exact model, quantization and context first; equal VRAM does not mean equal inference speed. [1] [2] | Not connectedNo price / stock feed |
GeForce RTX 4070 SUPERNVIDIA · Ada LovelaceOlder-generation option; verify the exact listing, seller, condition, warranty, cooling and connectors. No current price or availability checked. Board-partner specifications may differ. Memory interface: 192-bit [9]. Reference GPU specs; board-partner dimensions, cooling, power limits, and connectors can differ. |
12 GBGDDR6X [9] | 220 WReference board [9] | Consider for 1440p gaming; validate demanding ray-tracing and texture settings. 12 GB CUDA candidate. Choose the exact model, quantization and context first; equal VRAM does not mean equal inference speed. [1] [2] | Not connectedNo price / stock feed |
GeForce RTX 3090NVIDIA · AmpereLegacy stock/used evaluation; this is not a claim of current availability. Check seller, warranty, condition, cooling, and exact board specifications. power_w is NVIDIA reference board power, 350 W (Founders Edition Graphics Card Power), not typical gaming consumption or a whole-system PSU recommendation; partner boards may differ. Memory interface: 384-bit [10]. Reference GPU specs; board-partner dimensions, cooling, power limits, and connectors can differ. |
24 GBGDDR6X [10] | 350 WReference board [10] | Legacy high-end hardware; evaluate current game-specific results and inspect cooling and condition before considering a used board. 24 GB CUDA option to investigate on the legacy market. Inspect condition, thermals, warranty and seller carefully; no used-price or stock claims here. [1] [2] | Not connectedNo price / stock feed |
GeForce RTX 4060 Ti 16GBNVIDIA · Ada LovelaceOlder-generation option; verify the exact listing, seller, condition, warranty, cooling and connectors. No current price or availability checked. Board-partner specifications may differ. NVIDIA gives 160–165 W for the RTX 4060 Ti family; the exact selected variant is not mapped in the retrieved table. Memory interface: 128-bit [11]. Reference GPU specs; board-partner dimensions, cooling, power limits, and connectors can differ. |
16 GBGDDR6 [11] | Not verifiedReference board [11] | Only the 16 GB variant is cataloged; extra capacity should not be confused with higher gaming compute performance. 16 GB CUDA candidate. Choose the exact model, quantization and context first; equal VRAM does not mean equal inference speed. [1] [2] | Not connectedNo price / stock feed |
Radeon RX 9070 XTAMD · RDNA 4power_w is AMD reference Typical Board Power (Desktop) (304 W), not whole-system draw or PSU capacity. Partner-board power limits, cooling, connectors and dimensions may differ. No live US price, stock or listing condition verified; inspect the actual new/used listing, warranty and seller terms. Memory interface: 256-bit [12] [13]. Reference GPU specs; board-partner dimensions, cooling, power limits, and connectors can differ. |
16 GBGDDR6 [12] [13] | 304 WReference board [12] [13] | A candidate for higher-resolution gaming. Compare native-rendering and ray-tracing results in the games you play; no frame-rate guarantee. 16 GB capacity. Listed in Ollama’s Linux ROCm GPU table; the cited Windows ROCm list does not list this model. Check OS/backend support first. [1] [2] | Not connectedNo price / stock feed |
Radeon RX 9070AMD · RDNA 4power_w is AMD reference Typical Board Power (Desktop) (220 W), not whole-system draw or PSU capacity. Partner-board power limits, cooling, connectors and dimensions may differ. No live US price, stock or listing condition verified; inspect the actual new/used listing, warranty and seller terms. Memory interface: 256-bit [14] [13]. Reference GPU specs; board-partner dimensions, cooling, power limits, and connectors can differ. |
16 GBGDDR6 [14] [13] | 220 WReference board [14] [13] | A candidate for higher-resolution gaming. Compare native-rendering and ray-tracing results in the games you play; no frame-rate guarantee. 16 GB capacity. Listed in Ollama’s Linux ROCm GPU table; the cited Windows ROCm list does not list this model. Check OS/backend support first. [1] [2] | Not connectedNo price / stock feed |
Radeon RX 9060 XT 16GBAMD · RDNA 4power_w is AMD reference Typical Board Power (Desktop) (160 W), not whole-system draw or PSU capacity. Partner-board power limits, cooling, connectors and dimensions may differ. No live US price, stock or listing condition verified; inspect the actual new/used listing, warranty and seller terms. Memory interface: 128-bit [15] [13]. Reference GPU specs; board-partner dimensions, cooling, power limits, and connectors can differ. |
16 GBGDDR6 [15] [13] | 160 WReference board [15] [13] | A candidate for 1440p gaming. Compare native-rendering and ray-tracing results in the games you play; no frame-rate guarantee. 16 GB capacity. Listed in Ollama’s Linux ROCm GPU table; the cited Windows ROCm list does not list this model. Check OS/backend support first. [1] [2] | Not connectedNo price / stock feed |
Radeon RX 7900 XTXAMD · RDNA 3power_w is AMD reference Typical Board Power (Desktop) (355 W), not whole-system draw or PSU capacity. Partner-board power limits, cooling, connectors and dimensions may differ. No live US price, stock or listing condition verified; inspect the actual new/used listing, warranty and seller terms. Memory interface: 384-bit [16] [17]. Reference GPU specs; board-partner dimensions, cooling, power limits, and connectors can differ. |
24 GBGDDR6 [16] [17] | 355 WReference board [16] [17] | A candidate for higher-resolution gaming. Compare native-rendering and ray-tracing results in the games you play; no frame-rate guarantee. 24 GB capacity. Listed in Ollama’s Linux ROCm GPU table; also listed for Windows, with the required driver stack. [1] [2] | Not connectedNo price / stock feed |
Radeon RX 7900 XTAMD · RDNA 3power_w is AMD reference Typical Board Power (Desktop) (315 W), not whole-system draw or PSU capacity. Partner-board power limits, cooling, connectors and dimensions may differ. No live US price, stock or listing condition verified; inspect the actual new/used listing, warranty and seller terms. Memory interface: 320-bit [18] [17]. Reference GPU specs; board-partner dimensions, cooling, power limits, and connectors can differ. |
20 GBGDDR6 [18] [17] | 315 WReference board [18] [17] | A candidate for higher-resolution gaming. Compare native-rendering and ray-tracing results in the games you play; no frame-rate guarantee. 20 GB capacity. Listed in Ollama’s Linux ROCm GPU table; also listed for Windows, with the required driver stack. [1] [2] | Not connectedNo price / stock feed |
Radeon RX 7800 XTAMD · RDNA 3power_w is AMD reference Typical Board Power (Desktop) (263 W), not whole-system draw or PSU capacity. Partner-board power limits, cooling, connectors and dimensions may differ. No live US price, stock or listing condition verified; inspect the actual new/used listing, warranty and seller terms. Memory interface: 256-bit [19] [20]. Reference GPU specs; board-partner dimensions, cooling, power limits, and connectors can differ. |
16 GBGDDR6 [19] [20] | 263 WReference board [19] [20] | A candidate for 1440p gaming. Compare native-rendering and ray-tracing results in the games you play; no frame-rate guarantee. 16 GB capacity. Listed in Ollama’s Linux ROCm GPU table; also listed for Windows, with the required driver stack. [1] [2] | Not connectedNo price / stock feed |
Intel Arc B580Intel · Xe2 (Battlemage)power_w is Intel published reference TBP (Total Board Power) (190 W), not whole-system draw or PSU capacity. Partner-board power limits, cooling, connectors and dimensions may differ. No live US price, stock or listing condition verified; inspect the actual new/used listing, warranty and seller terms. Memory interface: 192-bit [21] [22]. Reference GPU specs; board-partner dimensions, cooling, power limits, and connectors can differ. |
12 GBGDDR6 [21] [22] | 190 WReference board [21] [22] | Investigate for mainstream gaming with current game-specific reviews. Check platform requirements and drivers rather than assuming all titles behave alike. 12 GB capacity. llama.cpp offers an Intel-oriented SYCL backend and Vulkan options. Check this card’s driver/runtime support; not a drop-in CUDA replacement. [1] [2] | Not connectedNo price / stock feed |
Intel Arc B570Intel · Xe2 (Battlemage)power_w is Intel published reference TBP (Total Board Power) (150 W), not whole-system draw or PSU capacity. Partner-board power limits, cooling, connectors and dimensions may differ. No live US price, stock or listing condition verified; inspect the actual new/used listing, warranty and seller terms. Memory interface: 160-bit [23] [22]. Reference GPU specs; board-partner dimensions, cooling, power limits, and connectors can differ. |
10 GBGDDR6 [23] [22] | 150 WReference board [23] [22] | Investigate for mainstream gaming with current game-specific reviews. Check platform requirements and drivers rather than assuming all titles behave alike. 10 GB capacity. llama.cpp offers an Intel-oriented SYCL backend and Vulkan options. Check this card’s driver/runtime support; not a drop-in CUDA replacement. [1] [2] | Not connectedNo price / stock feed |
Buying for gaming?
Start with the games you play, your monitor resolution, and independent benchmarks for those games. VRAM capacity alone does not rank gaming performance.
- Compare native rendering separately from upscaling and frame generation.
- Check your power supply, connectors, and the exact card's physical dimensions.
- A GPU model is not a retailer listing. Partner cards can differ.
Buying for local LLMs?
Decide which model and runtime you want to use before choosing hardware. Model weights, quantization, context length, KV cache, and runtime overhead all affect memory requirements. More VRAM is useful headroom, not a fit guarantee.
Ollama's supported GPU lists differ by OS. llama.cpp offers CUDA, HIP, Vulkan and SYCL backends; a supported backend does not guarantee every card/runtime combination works equally well.[1][2]
CPU/GPU partial offload is possible in llama.cpp, but this is not a claim of full-VRAM fit or a measured speed estimate.[2]
Selection method, limitations & sources
This is an editorial shortlist across current and older desktop GPUs, not the 20 fastest cards or an exhaustive market ranking. We include different VRAM capacities and software ecosystems to help gaming and local-LLM buyers compare tradeoffs. We have not benchmarked these cards. No scores, FPS, tokens/sec, live availability, or price/value rankings are fabricated.
Specs describe the stated reference model and selected VRAM variant. Confirm the exact manufacturer SKU before buying. Older GPUs can involve legacy inventory or used-market risk; inclusion is not a recommendation to buy from an unverified seller. Board power is not total PC power or a PSU recommendation. Missing facts are marked “Not verified.”
Sources were reviewed on 2026-10-05; runtime support and product specifications can change. There are no retailer affiliate links on this catalog yet.
- Ollama hardware support
- llama.cpp supported backends
- GeForce RTX 5090 Graphics Cards | NVIDIA
- NVIDIA GeForce RTX 5080 reference specifications
- GeForce RTX 5070 Family Graphics Cards | NVIDIA
- GeForce RTX 5060 Family Graphics Cards | NVIDIA
- NVIDIA GeForce RTX 4090 reference specifications
- NVIDIA GeForce RTX 4080 family reference specifications
- NVIDIA GeForce RTX 4070 family reference specifications
- 3090 & 3090 Ti Graphics Cards | NVIDIA GeForce
- NVIDIA GeForce RTX 4060 family reference specifications
- Radeon™ RX 9070 XT
- 243019457-B_Radeon 9000 Series QRG_Competitive.indd
- Radeon™ RX 9070
- AMD Radeon™ RX 9060 XT (16GB)
- Radeon™ RX 7900 XTX
- AMD Radeon™ RX 7900 Series Graphics
- Radeon™ RX 7900 XT
- Radeon™ RX 7800 XT
- AMD Radeon RX 7000 Series Graphics Cards
- Intel® Arc™ B580 Graphics
- Intel Arc B-Series Graphics for Desktops
- Intel® Arc™ B570 Graphics