{"meta":{"title":"D-Central AI & Local-Inference GPU Database","description":"GPU and AI-accelerator reference for local inference \u2014 VRAM, bandwidth, FP16\/INT8, TDP, and suitability tier. CC BY 4.0.","generated":"2026-07-15T13:44:01+00:00","version":"1.0","date_as_of":"2026-06-15","license":"https:\/\/creativecommons.org\/licenses\/by\/4.0\/","license_name":"CC BY 4.0","source":"https:\/\/d-central.tech\/data\/ai-gpu-database\/","record_count":30,"disclaimer":"Specs as of June 2026. Verify at source before purchasing decisions. FP16 methodology varies by card class \u2014 see plugin docs.","downloads":{"csv":"https:\/\/d-central.tech\/wp-content\/uploads\/data\/dcentral-ai-gpu-database.csv","json":"https:\/\/d-central.tech\/wp-content\/uploads\/data\/dcentral-ai-gpu-database.json"}},"rows":[{"id":"h100-sxm-80gb","name":"NVIDIA H100 SXM 80 GB","manufacturer":"NVIDIA","architecture":"Hopper (GH100)","vram_gb":80,"vram_type":"HBM3","mem_bandwidth_gbs":3350,"fp16_tflops":1979,"int8_tops":3958,"tdp_w":700,"inference_tier":1,"inference_tier_label":"Data Center \u2014 enterprise only","segment":"datacenter","ai_notes":"The frontier inference GPU as of 2026. Not home-deployable. Runs 405B+ models in FP16. Available via cloud; new pricing enterprise-only. Supports NVIDIA Hopper Transformer Engine (FP8). Credit: NVIDIA Corporation.","source_primary":"https:\/\/www.nvidia.com\/en-us\/data-center\/h100\/","last_verified":"2026-06"},{"id":"a100-sxm4-80gb","name":"NVIDIA A100 SXM4 80 GB","manufacturer":"NVIDIA","architecture":"Ampere Data Center (GA100)","vram_gb":80,"vram_type":"HBM2e","mem_bandwidth_gbs":2000,"fp16_tflops":312,"int8_tops":624,"tdp_w":400,"inference_tier":1,"inference_tier_label":"Data Center \u2014 enterprise only","segment":"datacenter","ai_notes":"Widely deployed cloud inference GPU for 70B-class models. Predecessor to H100; available used through secondary market at sharply lower prices than new. Credit: NVIDIA Corporation.","source_primary":"https:\/\/www.nvidia.com\/en-us\/data-center\/a100\/","last_verified":"2026-06"},{"id":"l40s-48gb","name":"NVIDIA L40S 48 GB","manufacturer":"NVIDIA","architecture":"Ada Lovelace (AD102)","vram_gb":48,"vram_type":"GDDR6","mem_bandwidth_gbs":864,"fp16_tflops":183,"int8_tops":366,"tdp_w":350,"inference_tier":1,"inference_tier_label":"Professional Inference \u2014 workstation\/server","segment":"workstation","ai_notes":"48 GB GDDR6 enables 70B-class models without quantisation. Designed for always-on inference servers. PCIe form factor deployable in standard workstation. Ada Lovelace Transformer Engine with FP8. Credit: NVIDIA Corporation.","source_primary":"https:\/\/www.nvidia.com\/en-us\/data-center\/l40s\/","last_verified":"2026-06"},{"id":"l4-24gb","name":"NVIDIA L4 24 GB","manufacturer":"NVIDIA","architecture":"Ada Lovelace (AD104)","vram_gb":24,"vram_type":"GDDR6","mem_bandwidth_gbs":300,"fp16_tflops":121,"int8_tops":242,"tdp_w":72,"inference_tier":2,"inference_tier_label":"Prosumer \u2014 serious home inference server","segment":"workstation","ai_notes":"Exceptional efficiency at 72 W TDP with 24 GB. Fits in a low-profile PCIe slot. Bandwidth-limited for large-batch throughput but outstanding per-watt for single-stream inference of 13B\u201334B quantised models. Credit: NVIDIA Corporation.","source_primary":"https:\/\/www.nvidia.com\/en-us\/data-center\/l4\/","last_verified":"2026-06"},{"id":"rtx-a6000-ampere-48gb","name":"NVIDIA RTX A6000 (Ampere) 48 GB","manufacturer":"NVIDIA","architecture":"Ampere (GA102)","vram_gb":48,"vram_type":"GDDR6","mem_bandwidth_gbs":768,"fp16_tflops":155,"int8_tops":310,"tdp_w":300,"inference_tier":2,"inference_tier_label":"Prosumer \u2014 serious home inference server","segment":"workstation","ai_notes":"48 GB GDDR6 makes this the VRAM champion for home\/small-office workstations. Runs full 65B\u201370B models in INT8. Available used for substantially less than new. NVLink bridge enables 96 GB pooled across two units. Credit: NVIDIA Corporation.","source_primary":"https:\/\/www.nvidia.com\/en-us\/products\/workstations\/rtx-a6000\/","last_verified":"2026-06"},{"id":"rtx-5090-32gb","name":"NVIDIA GeForce RTX 5090","manufacturer":"NVIDIA","architecture":"Blackwell (GB202)","vram_gb":32,"vram_type":"GDDR7","mem_bandwidth_gbs":1792,"fp16_tflops":210,"int8_tops":null,"tdp_w":575,"inference_tier":2,"inference_tier_label":"Prosumer \u2014 serious home inference server","segment":"consumer","ai_notes":"Highest bandwidth consumer GPU ever released (1,792 GB\/s). 32 GB GDDR7 enables unquantised 34B and quantised 70B locally. CUDA ecosystem (llama.cpp, Ollama, vLLM). Launched January 2025. Credit: NVIDIA Corporation.","source_primary":"https:\/\/www.nvidia.com\/en-us\/geforce\/graphics-cards\/50-series\/rtx-5090\/","last_verified":"2026-06"},{"id":"rtx-4090-24gb","name":"NVIDIA GeForce RTX 4090","manufacturer":"NVIDIA","architecture":"Ada Lovelace (AD102)","vram_gb":24,"vram_type":"GDDR6X","mem_bandwidth_gbs":1008,"fp16_tflops":165.19999999999998863131622783839702606201171875,"int8_tops":330.30000000000001136868377216160297393798828125,"tdp_w":450,"inference_tier":2,"inference_tier_label":"Prosumer \u2014 serious home inference server","segment":"consumer","ai_notes":"Gold standard home inference GPU. 24 GB enables 34B at FP16 or 70B at Q4. The highest-bandwidth consumer card before the RTX 5090. CUDA ecosystem (llama.cpp, Ollama, vLLM, LM Studio). Credit: NVIDIA Corporation.","source_primary":"https:\/\/www.nvidia.com\/en-us\/geforce\/graphics-cards\/40-series\/rtx-4090\/","last_verified":"2026-06"},{"id":"apple-m4-max-128gb","name":"Apple M4 Max (40-core GPU, up to 128 GB)","manufacturer":"Apple","architecture":"Apple Silicon (M4, 3 nm)","vram_gb":128,"vram_type":"Unified","mem_bandwidth_gbs":546,"fp16_tflops":18.39999999999999857891452847979962825775146484375,"int8_tops":38,"tdp_w":null,"inference_tier":2,"inference_tier_label":"Prosumer \u2014 serious home inference server","segment":"mobile-soc","ai_notes":"Up to 128 GB unified memory runs 70B models at FP16 without quantisation \u2014 a capability no single consumer discrete GPU matches. Energy-efficient (whole-laptop TGP ~92 W). Metal \/ MLX \/ llama.cpp Metal backend. Neural Engine adds 38 TOPS for INT8\/INT4 tasks. Note: fp16_tflops = GPU shader throughput only; int8_tops = Neural Engine. Credit: Apple Inc.","source_primary":"https:\/\/www.apple.com\/newsroom\/2024\/10\/apple-introduces-m4-pro-and-m4-max\/","last_verified":"2026-06"},{"id":"rtx-3090-24gb","name":"NVIDIA GeForce RTX 3090","manufacturer":"NVIDIA","architecture":"Ampere (GA102)","vram_gb":24,"vram_type":"GDDR6X","mem_bandwidth_gbs":936,"fp16_tflops":71,"int8_tops":142,"tdp_w":350,"inference_tier":3,"inference_tier_label":"Capable \u2014 13B\u201334B models","segment":"consumer","ai_notes":"24 GB VRAM and 936 GB\/s bandwidth make this the best-value used option for large-model inference. Widely available on the second-hand market. Ampere Tensor Cores (3rd gen). CUDA ecosystem. NVLink possible for 48 GB dual-card. Credit: NVIDIA Corporation.","source_primary":"https:\/\/www.nvidia.com\/en-us\/geforce\/graphics-cards\/30-series\/rtx-3090-3090ti\/","last_verified":"2026-06"},{"id":"rtx-5080-16gb","name":"NVIDIA GeForce RTX 5080","manufacturer":"NVIDIA","architecture":"Blackwell (GB203)","vram_gb":16,"vram_type":"GDDR7","mem_bandwidth_gbs":960,"fp16_tflops":112.599999999999994315658113919198513031005859375,"int8_tops":null,"tdp_w":360,"inference_tier":3,"inference_tier_label":"Capable \u2014 13B\u201334B models","segment":"consumer","ai_notes":"High GDDR7 bandwidth narrows the gap vs. 5090 for memory-bound LLM inference. 16 GB limits model size; 34B Q4 fits, 70B does not. Launched January 2025. CUDA ecosystem. Credit: NVIDIA Corporation.","source_primary":"https:\/\/www.nvidia.com\/en-us\/geforce\/graphics-cards\/50-series\/rtx-5080\/","last_verified":"2026-06"},{"id":"rtx-4080-super-16gb","name":"NVIDIA GeForce RTX 4080 Super","manufacturer":"NVIDIA","architecture":"Ada Lovelace (AD103)","vram_gb":16,"vram_type":"GDDR6X","mem_bandwidth_gbs":736,"fp16_tflops":104.400000000000005684341886080801486968994140625,"int8_tops":null,"tdp_w":320,"inference_tier":3,"inference_tier_label":"Capable \u2014 13B\u201334B models","segment":"consumer","ai_notes":"Solid 16 GB Ada card. Better value than RTX 4080 for new buyers. 13B models fit in FP16; 34B requires quantisation. CUDA ecosystem. Credit: NVIDIA Corporation.","source_primary":"https:\/\/www.nvidia.com\/en-us\/geforce\/graphics-cards\/40-series\/rtx-4080-family\/","last_verified":"2026-06"},{"id":"rtx-4080-16gb","name":"NVIDIA GeForce RTX 4080","manufacturer":"NVIDIA","architecture":"Ada Lovelace (AD103)","vram_gb":16,"vram_type":"GDDR6X","mem_bandwidth_gbs":717,"fp16_tflops":97.5,"int8_tops":null,"tdp_w":320,"inference_tier":3,"inference_tier_label":"Capable \u2014 13B\u201334B models","segment":"consumer","ai_notes":"16 GB VRAM, strong bandwidth. Equivalent AI inference to 4080 Super for most workloads. CUDA ecosystem. Credit: NVIDIA Corporation.","source_primary":"https:\/\/www.nvidia.com\/en-us\/geforce\/graphics-cards\/40-series\/rtx-4080\/","last_verified":"2026-06"},{"id":"rx-7900-xtx-24gb","name":"AMD Radeon RX 7900 XTX","manufacturer":"AMD","architecture":"RDNA 3 (Navi 31)","vram_gb":24,"vram_type":"GDDR6","mem_bandwidth_gbs":960,"fp16_tflops":123,"int8_tops":123,"tdp_w":355,"inference_tier":3,"inference_tier_label":"Capable \u2014 13B\u201334B models","segment":"consumer","ai_notes":"24 GB GDDR6 and high bandwidth rival the RTX 4090. AI inference via ROCm (Linux) or Vulkan\/DirectML (Windows). ROCm ecosystem support has improved significantly; not yet as seamless as CUDA. Credit: AMD (Advanced Micro Devices).","source_primary":"https:\/\/www.amd.com\/en\/products\/graphics\/desktops\/radeon\/7000-series\/amd-radeon-rx-7900xtx.html","last_verified":"2026-06"},{"id":"apple-m3-max-128gb","name":"Apple M3 Max (40-core GPU, up to 128 GB)","manufacturer":"Apple","architecture":"Apple Silicon (M3, 3 nm)","vram_gb":128,"vram_type":"Unified","mem_bandwidth_gbs":400,"fp16_tflops":16.39999999999999857891452847979962825775146484375,"int8_tops":18,"tdp_w":null,"inference_tier":3,"inference_tier_label":"Capable \u2014 13B\u201334B models","segment":"mobile-soc","ai_notes":"Same 128 GB unified memory ceiling as M4 Max, at lower bandwidth (400 GB\/s) and slightly lower GPU throughput. Excellent for local 70B at Q4. MLX \/ Metal \/ llama.cpp Metal backend. Note: fp16_tflops = GPU shader; int8_tops = Neural Engine. Credit: Apple Inc.","source_primary":"https:\/\/www.apple.com\/newsroom\/2023\/10\/apple-unveils-m3-m3-pro-and-m3-max-the-most-advanced-chips-for-a-personal-computer\/","last_verified":"2026-06"},{"id":"rtx-4070-ti-super-16gb","name":"NVIDIA GeForce RTX 4070 Ti Super","manufacturer":"NVIDIA","architecture":"Ada Lovelace (AD103)","vram_gb":16,"vram_type":"GDDR6X","mem_bandwidth_gbs":672,"fp16_tflops":88.099999999999994315658113919198513031005859375,"int8_tops":null,"tdp_w":285,"inference_tier":3,"inference_tier_label":"Capable \u2014 13B\u201334B models","segment":"consumer","ai_notes":"16 GB VRAM upgrade over the standard 4070 Ti makes a meaningful difference for AI. Good price-to-bandwidth ratio. 13B models in FP16; 34B in Q4. CUDA ecosystem. Credit: NVIDIA Corporation.","source_primary":"https:\/\/www.nvidia.com\/en-us\/geforce\/graphics-cards\/40-series\/rtx-4070-ti-super\/","last_verified":"2026-06"},{"id":"rx-7900-xt-20gb","name":"AMD Radeon RX 7900 XT","manufacturer":"AMD","architecture":"RDNA 3 (Navi 31)","vram_gb":20,"vram_type":"GDDR6","mem_bandwidth_gbs":800,"fp16_tflops":103,"int8_tops":103,"tdp_w":315,"inference_tier":3,"inference_tier_label":"Capable \u2014 13B\u201334B models","segment":"consumer","ai_notes":"20 GB sweet spot \u2014 runs 20B models in FP16, 34B in Q4; more VRAM headroom than 16 GB cards. ROCm (Linux) or Vulkan\/DirectML. Credit: AMD (Advanced Micro Devices).","source_primary":"https:\/\/www.amd.com\/en\/products\/graphics\/desktops\/radeon\/7000-series\/amd-radeon-rx-7900-xt.html","last_verified":"2026-06"},{"id":"apple-m4-pro-64gb","name":"Apple M4 Pro (20-core GPU, up to 64 GB)","manufacturer":"Apple","architecture":"Apple Silicon (M4, 3 nm)","vram_gb":64,"vram_type":"Unified","mem_bandwidth_gbs":273,"fp16_tflops":9.199999999999999289457264239899814128875732421875,"int8_tops":38,"tdp_w":null,"inference_tier":3,"inference_tier_label":"Capable \u2014 13B\u201334B models","segment":"mobile-soc","ai_notes":"64 GB unified memory at reasonable bandwidth. Runs 34B Q4 locally; 70B Q4 fits but is bandwidth-constrained at 273 GB\/s. Laptop-optimised power envelope. MLX \/ Metal. Note: fp16_tflops = GPU shader estimate (low confidence); int8_tops = Neural Engine official. Credit: Apple Inc.","source_primary":"https:\/\/support.apple.com\/en-us\/121553","last_verified":"2026-06"},{"id":"rtx-4070-super-12gb","name":"NVIDIA GeForce RTX 4070 Super","manufacturer":"NVIDIA","architecture":"Ada Lovelace (AD104)","vram_gb":12,"vram_type":"GDDR6X","mem_bandwidth_gbs":504,"fp16_tflops":71,"int8_tops":null,"tdp_w":220,"inference_tier":4,"inference_tier_label":"Mid-Range \u2014 7B\u201313B models","segment":"consumer","ai_notes":"Best price-to-performance in the Ada mid-tier for AI. 12 GB limits to 7B\u201313B at Q4; some 20B models fit at Q3\/Q4. CUDA ecosystem. Credit: NVIDIA Corporation.","source_primary":"https:\/\/www.nvidia.com\/en-us\/geforce\/graphics-cards\/40-series\/rtx-4070-super\/","last_verified":"2026-06"},{"id":"rx-7800-xt-16gb","name":"AMD Radeon RX 7800 XT","manufacturer":"AMD","architecture":"RDNA 3 (Navi 32)","vram_gb":16,"vram_type":"GDDR6","mem_bandwidth_gbs":576,"fp16_tflops":74.599999999999994315658113919198513031005859375,"int8_tops":74.599999999999994315658113919198513031005859375,"tdp_w":263,"inference_tier":4,"inference_tier_label":"Mid-Range \u2014 7B\u201313B models","segment":"consumer","ai_notes":"16 GB VRAM at mid-range price. Runs 13B in FP16; 34B in heavy quantisation. ROCm (Linux) \/ Vulkan. Good budget option for VRAM-over-compute builds. Credit: AMD (Advanced Micro Devices).","source_primary":"https:\/\/www.amd.com\/en\/products\/graphics\/desktops\/radeon\/7000-series\/amd-radeon-rx-7800-xt.html","last_verified":"2026-06"},{"id":"arc-a770-16gb","name":"Intel Arc A770 16 GB","manufacturer":"Intel","architecture":"Xe-HPG \/ Alchemist (ACM-G10)","vram_gb":16,"vram_type":"GDDR6","mem_bandwidth_gbs":560,"fp16_tflops":39.39999999999999857891452847979962825775146484375,"int8_tops":null,"tdp_w":225,"inference_tier":4,"inference_tier_label":"Mid-Range \u2014 7B\u201313B models","segment":"consumer","ai_notes":"Best-value 16 GB card at launch (2022). Intel XMX matrix engines accelerate INT8\/FP16 inference. OpenVINO + IPEX-LLM ecosystem; llama.cpp SYCL backend. ROCm\/CUDA not available. Ecosystem maturity lower than NVIDIA. Credit: Intel Corporation.","source_primary":"https:\/\/www.intel.com\/content\/www\/us\/en\/products\/sku\/229151\/intel-arc-a770-graphics-16gb\/specifications.html","last_verified":"2026-06"},{"id":"rx-7700-xt-12gb","name":"AMD Radeon RX 7700 XT","manufacturer":"AMD","architecture":"RDNA 3 (Navi 32)","vram_gb":12,"vram_type":"GDDR6","mem_bandwidth_gbs":432,"fp16_tflops":70.2999999999999971578290569595992565155029296875,"int8_tops":70.2999999999999971578290569595992565155029296875,"tdp_w":245,"inference_tier":4,"inference_tier_label":"Mid-Range \u2014 7B\u201313B models","segment":"consumer","ai_notes":"12 GB RDNA 3 at mid-range price. Handles 7B in FP16 and 13B at Q4. ROCm (Linux) \/ Vulkan. Credit: AMD (Advanced Micro Devices).","source_primary":"https:\/\/www.amd.com\/en\/products\/graphics\/desktops\/radeon\/7000-series\/amd-radeon-rx-7700-xt.html","last_verified":"2026-06"},{"id":"rtx-4070-12gb","name":"NVIDIA GeForce RTX 4070","manufacturer":"NVIDIA","architecture":"Ada Lovelace (AD104)","vram_gb":12,"vram_type":"GDDR6X","mem_bandwidth_gbs":504,"fp16_tflops":58.5,"int8_tops":null,"tdp_w":200,"inference_tier":4,"inference_tier_label":"Mid-Range \u2014 7B\u201313B models","segment":"consumer","ai_notes":"Efficient 200 W Ada card. 12 GB handles 7B\u201313B Q4. Prefer the 4070 Super for the same price range if still available. CUDA ecosystem. Credit: NVIDIA Corporation.","source_primary":"https:\/\/www.nvidia.com\/en-us\/geforce\/graphics-cards\/40-series\/rtx-4070\/","last_verified":"2026-06"},{"id":"rtx-4060-ti-16gb","name":"NVIDIA GeForce RTX 4060 Ti 16 GB","manufacturer":"NVIDIA","architecture":"Ada Lovelace (AD106)","vram_gb":16,"vram_type":"GDDR6","mem_bandwidth_gbs":288,"fp16_tflops":44.2000000000000028421709430404007434844970703125,"int8_tops":null,"tdp_w":165,"inference_tier":4,"inference_tier_label":"Mid-Range \u2014 7B\u201313B models","segment":"consumer","ai_notes":"16 GB VRAM at a modest TDP \u2014 the main draw for AI. Memory bandwidth is a bottleneck at 288 GB\/s; token generation will be ~3\u00d7 slower than the RTX 4090 despite the same VRAM. Best for those who need 16 GB on a tight budget. CUDA. Credit: NVIDIA Corporation.","source_primary":"https:\/\/www.nvidia.com\/en-us\/geforce\/graphics-cards\/40-series\/rtx-4060-ti\/","last_verified":"2026-06"},{"id":"rtx-3080-12gb","name":"NVIDIA GeForce RTX 3080 12 GB","manufacturer":"NVIDIA","architecture":"Ampere (GA102)","vram_gb":12,"vram_type":"GDDR6X","mem_bandwidth_gbs":912,"fp16_tflops":61.2999999999999971578290569595992565155029296875,"int8_tops":null,"tdp_w":350,"inference_tier":4,"inference_tier_label":"Mid-Range \u2014 7B\u201313B models","segment":"consumer","ai_notes":"High bandwidth at 912 GB\/s despite 12 GB VRAM cap. Ideal for fast token generation on 7B\u201312B models. Available used at strong value. CUDA ecosystem. 350 W TDP is the main drawback. Credit: NVIDIA Corporation.","source_primary":"https:\/\/www.nvidia.com\/en-us\/geforce\/graphics-cards\/30-series\/rtx-3080-3080ti\/","last_verified":"2026-06"},{"id":"rx-6900-xt-16gb","name":"AMD Radeon RX 6900 XT","manufacturer":"AMD","architecture":"RDNA 2 (Navi 21)","vram_gb":16,"vram_type":"GDDR6","mem_bandwidth_gbs":512,"fp16_tflops":46.10000000000000142108547152020037174224853515625,"int8_tops":null,"tdp_w":300,"inference_tier":4,"inference_tier_label":"Mid-Range \u2014 7B\u201313B models","segment":"consumer","ai_notes":"16 GB GDDR6 at good used prices. RDNA 2 lacks the AI Accelerators of RDNA 3, so FP16 performance is shader-only. ROCm support present but older generation. Vulkan \/ DirectML usable for inference. Credit: AMD (Advanced Micro Devices).","source_primary":"https:\/\/www.amd.com\/en\/products\/graphics\/desktops\/radeon\/6000-series\/amd-radeon-rx-6900-xt.html","last_verified":"2026-06"},{"id":"arc-b580-12gb","name":"Intel Arc B580 12 GB","manufacturer":"Intel","architecture":"Xe2-HPG \/ Battlemage (BMG-G21)","vram_gb":12,"vram_type":"GDDR6","mem_bandwidth_gbs":456,"fp16_tflops":27.300000000000000710542735760100185871124267578125,"int8_tops":null,"tdp_w":190,"inference_tier":4,"inference_tier_label":"Mid-Range \u2014 7B\u201313B models","segment":"consumer","ai_notes":"Best-value 12 GB card at its $249 launch price (December 2024). Xe2 architecture with 160 XMX engines for FP16\/INT8 matrix acceleration. 12 GB VRAM holds 7B in FP16 or 13B in Q4. Intel IPEX-LLM \/ OpenVINO ecosystem; llama.cpp SYCL. Credit: Intel Corporation.","source_primary":"https:\/\/www.intel.com\/content\/www\/us\/en\/products\/sku\/241598\/intel-arc-b580-graphics\/specifications.html","last_verified":"2026-06"},{"id":"rtx-4060-ti-8gb","name":"NVIDIA GeForce RTX 4060 Ti 8 GB","manufacturer":"NVIDIA","architecture":"Ada Lovelace (AD106)","vram_gb":8,"vram_type":"GDDR6","mem_bandwidth_gbs":288,"fp16_tflops":44.2000000000000028421709430404007434844970703125,"int8_tops":null,"tdp_w":160,"inference_tier":5,"inference_tier_label":"Entry \u2014 7B limit, VRAM-constrained","segment":"consumer","ai_notes":"8 GB VRAM limits to 7B Q4 or smaller. Good compute and low TDP but the VRAM wall prevents larger models without CPU offload. Prefer the 16 GB variant or the RTX 4060 Ti 16 GB for AI work. CUDA. Credit: NVIDIA Corporation.","source_primary":"https:\/\/www.nvidia.com\/en-us\/geforce\/graphics-cards\/40-series\/rtx-4060-ti\/","last_verified":"2026-06"},{"id":"rtx-4060-8gb","name":"NVIDIA GeForce RTX 4060","manufacturer":"NVIDIA","architecture":"Ada Lovelace (AD107)","vram_gb":8,"vram_type":"GDDR6","mem_bandwidth_gbs":272,"fp16_tflops":30.10000000000000142108547152020037174224853515625,"int8_tops":null,"tdp_w":115,"inference_tier":5,"inference_tier_label":"Entry \u2014 7B limit, VRAM-constrained","segment":"consumer","ai_notes":"Extremely efficient at 115 W TDP. 8 GB VRAM limits to 7B Q4 or smaller. Good entry point for those starting with local AI on a tight budget. CUDA ecosystem. Credit: NVIDIA Corporation.","source_primary":"https:\/\/www.nvidia.com\/en-us\/geforce\/graphics-cards\/40-series\/rtx-4060\/","last_verified":"2026-06"},{"id":"rx-7600-8gb","name":"AMD Radeon RX 7600","manufacturer":"AMD","architecture":"RDNA 3 (Navi 33)","vram_gb":8,"vram_type":"GDDR6","mem_bandwidth_gbs":288,"fp16_tflops":42.7999999999999971578290569595992565155029296875,"int8_tops":42.7999999999999971578290569595992565155029296875,"tdp_w":165,"inference_tier":5,"inference_tier_label":"Entry \u2014 7B limit, VRAM-constrained","segment":"consumer","ai_notes":"8 GB RDNA 3 budget card. 64 AI Accelerators help FP16\/INT8 throughput. VRAM cap at 8 GB is the binding constraint. ROCm (Linux) \/ Vulkan. Credit: AMD (Advanced Micro Devices).","source_primary":"https:\/\/www.amd.com\/en\/products\/graphics\/desktops\/radeon\/7000-series\/amd-radeon-rx-7600.html","last_verified":"2026-06"},{"id":"rtx-3070-8gb","name":"NVIDIA GeForce RTX 3070","manufacturer":"NVIDIA","architecture":"Ampere (GA104)","vram_gb":8,"vram_type":"GDDR6","mem_bandwidth_gbs":448,"fp16_tflops":40.60000000000000142108547152020037174224853515625,"int8_tops":null,"tdp_w":220,"inference_tier":5,"inference_tier_label":"Entry \u2014 7B limit, VRAM-constrained","segment":"consumer","ai_notes":"Widely available used at low prices. 8 GB VRAM is the main constraint; limits to 7B Q4 or smaller. Better bandwidth than comparable RTX 40-series 8 GB cards. CUDA ecosystem. Credit: NVIDIA Corporation.","source_primary":"https:\/\/www.nvidia.com\/en-us\/geforce\/graphics-cards\/30-series\/rtx-3070-3070ti\/","last_verified":"2026-06"}]}