{
 "source": "https://vramcalculator.net",
 "updated": "2026-10-14",
 "license": "Free to use with attribution to VRAMCalculator.net",
 "items": [
  {
   "slug": "rtx-2060-12gb",
   "name": "RTX 2060 12GB",
   "brand": "NVIDIA",
   "type": "consumer",
   "vram": 12,
   "bw": 336,
   "arch": "Turing",
   "released": "2021-12",
   "notes": "A late Turing refresh that doubled the 2060's memory. 12 GB at a low used price makes it a surprising budget option for 12–14B models, though bandwidth and older tensor cores make it slower than an RTX 3060 12GB."
  },
  {
   "slug": "rtx-2080-ti-11gb",
   "name": "RTX 2080 Ti 11GB",
   "brand": "NVIDIA",
   "type": "consumer",
   "vram": 11,
   "bw": 616,
   "arch": "Turing",
   "released": "2018-09",
   "notes": "Still fast for its age thanks to 616 GB/s of bandwidth, but 11 GB is an awkward size: it runs 8B models quickly and 14B models only with short contexts. No FP8 or BF16 support, which matters for some image and training tools."
  },
  {
   "slug": "rtx-3050-8gb",
   "name": "RTX 3050 8GB",
   "brand": "NVIDIA",
   "type": "consumer",
   "vram": 8,
   "bw": 224,
   "arch": "Ampere",
   "released": "2022-01",
   "notes": "The cheapest 8 GB NVIDIA card with CUDA. Its 224 GB/s of bandwidth is the bottleneck: it runs the same models as an RTX 4060 but noticeably slower."
  },
  {
   "slug": "rtx-3060-12gb",
   "name": "RTX 3060 12GB",
   "brand": "NVIDIA",
   "type": "consumer",
   "vram": 12,
   "bw": 360,
   "arch": "Ampere",
   "released": "2021-02",
   "ref": true,
   "notes": "The best-value entry card for local AI. 12 GB fits 14B models at Q4_K_M, which 8 GB cards cannot, and used prices are low. Make sure you get the 12 GB version: the later 8 GB variant has less memory and less bandwidth.",
   "tips": "Setup tips: run 14B models at Q4_K_M with an 8–16K context, or 8–9B models at Q8_0 for best quality. A Q8_0 KV cache roughly doubles the context you can fit. At 170 W it runs quietly in almost any PC, and two of them make a cheap 24 GB layer-split setup for 27–32B models."
  },
  {
   "slug": "rtx-3060-ti-8gb",
   "name": "RTX 3060 Ti 8GB",
   "brand": "NVIDIA",
   "type": "consumer",
   "vram": 8,
   "bw": 448,
   "arch": "Ampere",
   "released": "2020-12",
   "notes": "Faster than the RTX 3060 12GB for models that fit in 8 GB, but it cannot hold 14B models. For local AI, the 12 GB RTX 3060 is usually the better buy."
  },
  {
   "slug": "rtx-3070-8gb",
   "name": "RTX 3070 8GB",
   "brand": "NVIDIA",
   "type": "consumer",
   "vram": 8,
   "bw": 448,
   "arch": "Ampere",
   "released": "2020-10",
   "notes": "A strong gaming card held back by 8 GB for AI. It runs 7B and 8B models quickly but needs CPU offload above that."
  },
  {
   "slug": "rtx-3070-ti-8gb",
   "name": "RTX 3070 Ti 8GB",
   "brand": "NVIDIA",
   "type": "consumer",
   "vram": 8,
   "bw": 608,
   "arch": "Ampere",
   "released": "2021-06",
   "notes": "GDDR6X makes it one of the fastest 8 GB cards for token generation, but memory, not speed, is the limit. It is a good 7–8B model card and needs offloading for anything larger."
  },
  {
   "slug": "rtx-3080-10gb",
   "name": "RTX 3080 10GB",
   "brand": "NVIDIA",
   "type": "consumer",
   "vram": 10,
   "bw": 760,
   "arch": "Ampere",
   "released": "2020-09",
   "notes": "High bandwidth makes it fast, but 10 GB is an awkward size: too small for 14B models with useful context, and only slightly roomier than 8 GB cards."
  },
  {
   "slug": "rtx-3080-12gb",
   "name": "RTX 3080 12GB",
   "brand": "NVIDIA",
   "type": "consumer",
   "vram": 12,
   "bw": 912,
   "arch": "Ampere",
   "released": "2022-01",
   "notes": "The 12 GB refresh of the 3080 with a wider memory bus. It fits the same models as an RTX 3060 12GB but generates tokens about 2.5 times faster."
  },
  {
   "slug": "rtx-3090-24gb",
   "name": "RTX 3090 24GB",
   "brand": "NVIDIA",
   "type": "consumer",
   "vram": 24,
   "bw": 936,
   "arch": "Ampere",
   "released": "2020-09",
   "ref": true,
   "notes": "The most popular used card for local LLMs. 24 GB fits 32B models at Q4_K_M, and two of them (48 GB) run 70B models. It has NVLink, which the 4090 dropped, though most inference runtimes don't need it.",
   "tips": "Setup tips: set a power limit of about 280 W with nvidia-smi; generation speed barely drops and the card runs far cooler. Check memory junction temperatures on used cards, since the rear memory chips run hot; new thermal pads often help. For 70B models, pair two cards in llama.cpp or Ollama, which split layers automatically."
  },
  {
   "slug": "rtx-3090-ti-24gb",
   "name": "RTX 3090 Ti 24GB",
   "brand": "NVIDIA",
   "type": "consumer",
   "vram": 24,
   "bw": 1008,
   "arch": "Ampere",
   "released": "2022-03",
   "notes": "Slightly faster memory than the 3090 with the same 24 GB. For LLMs it performs within about 10% of a 3090 and runs the same models."
  },
  {
   "slug": "rtx-4060-8gb",
   "name": "RTX 4060 8GB",
   "brand": "NVIDIA",
   "type": "consumer",
   "vram": 8,
   "bw": 272,
   "arch": "Ada Lovelace",
   "released": "2023-06",
   "notes": "An efficient 8 GB card. A narrow 128-bit bus limits token speed, and 8 GB limits you to 8B-class models at Q4_K_M with moderate context."
  },
  {
   "slug": "rtx-4060-ti-8gb",
   "name": "RTX 4060 Ti 8GB",
   "brand": "NVIDIA",
   "type": "consumer",
   "vram": 8,
   "bw": 288,
   "arch": "Ada Lovelace",
   "released": "2023-05",
   "notes": "Same memory bus as the 16 GB version but half the memory. For AI, the 16 GB model is worth the extra cost."
  },
  {
   "slug": "rtx-4060-ti-16gb",
   "name": "RTX 4060 Ti 16GB",
   "brand": "NVIDIA",
   "type": "consumer",
   "vram": 16,
   "bw": 288,
   "arch": "Ada Lovelace",
   "released": "2023-07",
   "ref": true,
   "notes": "The cheapest new 16 GB NVIDIA card of its generation. 16 GB fits 14B models at Q6_K or 24B models at Q4_K_M, but the 288 GB/s bus makes it slower than a used RTX 3090 or 3080 12GB.",
   "tips": "Setup tips: the narrow 128-bit bus makes this card slower than its price suggests, so favour mixture-of-experts models such as gpt-oss-20b, which read few parameters per token. It draws only about 165 W, making it a good choice for a quiet always-on home AI server."
  },
  {
   "slug": "rtx-4070-12gb",
   "name": "RTX 4070 12GB",
   "brand": "NVIDIA",
   "type": "consumer",
   "vram": 12,
   "bw": 504,
   "arch": "Ada Lovelace",
   "released": "2023-04",
   "notes": "Same 12 GB as the RTX 3060 with more bandwidth, so it runs the same models about 40% faster.",
   "tips": "Setup tips: excellent for 8–14B models, where its bandwidth gives quick replies. Keep contexts to 8–16K on 14B models, or use a Q8_0 KV cache. For image generation, FLUX in GGUF Q5–Q6 with model swapping fits well."
  },
  {
   "slug": "rtx-4070-super-12gb",
   "name": "RTX 4070 Super 12GB",
   "brand": "NVIDIA",
   "type": "consumer",
   "vram": 12,
   "bw": 504,
   "arch": "Ada Lovelace",
   "released": "2024-01",
   "notes": "More compute than the 4070 but the same memory and bandwidth, so LLM token speed is nearly identical.",
   "tips": "Setup tips: treat it as a fast 12 GB card: 14B models at Q4_K_M, 8–9B models at Q8_0. For more than 16K of context on 14B models, enable flash attention and a Q8_0 KV cache."
  },
  {
   "slug": "rtx-4070-ti-12gb",
   "name": "RTX 4070 Ti 12GB",
   "brand": "NVIDIA",
   "type": "consumer",
   "vram": 12,
   "bw": 504,
   "arch": "Ada Lovelace",
   "released": "2023-01",
   "notes": "A fast 12 GB card. It runs the same models as an RTX 3060 12GB about 40% faster, and its Ada tensor cores support FP8, which helps image generation. The later 4070 Ti Super moved to 16 GB."
  },
  {
   "slug": "rtx-4070-ti-super-16gb",
   "name": "RTX 4070 Ti Super 16GB",
   "brand": "NVIDIA",
   "type": "consumer",
   "vram": 16,
   "bw": 672,
   "arch": "Ada Lovelace",
   "released": "2024-01",
   "notes": "A fast 16 GB card. It runs 14B models at high quality and 24B models at Q4_K_M with a short context.",
   "tips": "Setup tips: a strong 16 GB card for gpt-oss-20b, Gemma 4 26B-A4B and 14B models at high quality. 24B dense models fit at Q4_K_M only with a short context. It handles FLUX.1 in FP8 well thanks to Ada's FP8 support."
  },
  {
   "slug": "rtx-4080-16gb",
   "name": "RTX 4080 16GB",
   "brand": "NVIDIA",
   "type": "consumer",
   "vram": 16,
   "bw": 717,
   "arch": "Ada Lovelace",
   "released": "2022-11",
   "notes": "Nearly identical to the 4080 Super for AI: same 16 GB and slightly less bandwidth. A strong card for 14B models at high quality and 20–24B models at Q4 with a short context, but it cannot hold 32B models with useful context."
  },
  {
   "slug": "rtx-4080-super-16gb",
   "name": "RTX 4080 Super 16GB",
   "brand": "NVIDIA",
   "type": "consumer",
   "vram": 16,
   "bw": 736,
   "arch": "Ada Lovelace",
   "released": "2024-01",
   "notes": "Plenty of compute but only 16 GB. For local AI a used RTX 3090 holds larger models, though the 4080 Super is faster on anything that fits in 16 GB.",
   "tips": "Setup tips: fast for everything that fits in 16 GB. If you mostly run LLMs, compare its price with a used RTX 3090: the 3090 is a little slower but holds 27–35B models, which this card cannot."
  },
  {
   "slug": "rtx-4090-24gb",
   "name": "RTX 4090 24GB",
   "brand": "NVIDIA",
   "type": "consumer",
   "vram": 24,
   "bw": 1008,
   "arch": "Ada Lovelace",
   "released": "2022-10",
   "ref": true,
   "notes": "The fastest 24 GB consumer card of its generation. Token generation is only slightly faster than a 3090 (bandwidth is similar), but prompt processing and image generation are roughly twice as fast.",
   "tips": "Setup tips: the 4090 shines at prompt processing, image generation and fine-tuning. For long coding-agent sessions, pair it with hybrid-attention models such as Qwen 3.6 27B, which fit 64K+ of context. A power limit of 350–400 W costs little LLM speed."
  },
  {
   "slug": "rtx-5060-8gb",
   "name": "RTX 5060 8GB",
   "brand": "NVIDIA",
   "type": "consumer",
   "vram": 8,
   "bw": 448,
   "arch": "Blackwell",
   "released": "2025-05",
   "notes": "The cheapest current Blackwell card. GDDR7 gives it more bandwidth than the RTX 4060, but 8 GB still limits it to 7–9B models and small image models without offloading."
  },
  {
   "slug": "rtx-5060-ti-8gb",
   "name": "RTX 5060 Ti 8GB",
   "brand": "NVIDIA",
   "type": "consumer",
   "vram": 8,
   "bw": 448,
   "arch": "Blackwell",
   "released": "2025-04",
   "notes": "Same chip and bandwidth as the 16 GB 5060 Ti but half the memory. For local AI the 16 GB version is worth the extra money; this one is limited to 8B-class models."
  },
  {
   "slug": "rtx-5060-ti-16gb",
   "name": "RTX 5060 Ti 16GB",
   "brand": "NVIDIA",
   "type": "consumer",
   "vram": 16,
   "bw": 448,
   "arch": "Blackwell",
   "released": "2025-04",
   "notes": "A budget 16 GB Blackwell card with GDDR7. About 55% more bandwidth than the 4060 Ti 16GB at a similar price, plus FP4 support for newer runtimes.",
   "tips": "Setup tips: make sure you buy the 16 GB version. It is the best-value new 16 GB card for local AI: gpt-oss-20b fits fully with a long context, and GDDR7 makes it noticeably faster than the 4060 Ti 16GB. Update to a recent driver and CUDA 12.8 or newer for Blackwell support."
  },
  {
   "slug": "rtx-5070-12gb",
   "name": "RTX 5070 12GB",
   "brand": "NVIDIA",
   "type": "consumer",
   "vram": 12,
   "bw": 672,
   "arch": "Blackwell",
   "released": "2025-03",
   "notes": "Fast GDDR7 memory but only 12 GB. Good for 8B to 14B models; for bigger models a 16 GB card is the better choice."
  },
  {
   "slug": "rtx-5070-ti-16gb",
   "name": "RTX 5070 Ti 16GB",
   "brand": "NVIDIA",
   "type": "consumer",
   "vram": 16,
   "bw": 896,
   "arch": "Blackwell",
   "released": "2025-02",
   "notes": "Nearly RTX 3090-class bandwidth with 16 GB and modern features. It is fast for anything up to about 24B at Q4_K_M.",
   "tips": "Setup tips: fast for 14–20B models and mixture-of-experts models up to about 16 GB. Use a recent llama.cpp or Ollama build compiled for Blackwell. For 27B+ models, a Q3 quant or partial offload works but loses some quality or speed."
  },
  {
   "slug": "rtx-5080-16gb",
   "name": "RTX 5080 16GB",
   "brand": "NVIDIA",
   "type": "consumer",
   "vram": 16,
   "bw": 960,
   "arch": "Blackwell",
   "released": "2025-01",
   "notes": "Very fast, but 16 GB caps model size. For LLMs it runs the same models as a 5070 Ti about 7% faster.",
   "tips": "Setup tips: very fast within 16 GB. For local LLMs, capacity is its limit, so it suits users who mainly run 8–20B models or gpt-oss-20b. FP4 support helps newer image models in FP4 or NVFP4 builds."
  },
  {
   "slug": "rtx-5090-32gb",
   "name": "RTX 5090 32GB",
   "brand": "NVIDIA",
   "type": "consumer",
   "vram": 32,
   "bw": 1792,
   "arch": "Blackwell",
   "released": "2025-01",
   "ref": true,
   "notes": "The first consumer card with 32 GB and nearly double the 4090's bandwidth. It runs 32B models at Q6_K or with long context, and 70B models only at very low quantization.",
   "tips": "Setup tips: plan for a quality 1,000 W+ power supply and good case airflow; a power limit of 450–500 W keeps LLM speed nearly unchanged. 32 GB fits 27–32B models at Q6_K or with 64K+ contexts, and makes FLUX.2 practical with GGUF weights."
  },
  {
   "slug": "rx-6700-xt-12gb",
   "name": "Radeon RX 6700 XT 12GB",
   "brand": "AMD",
   "type": "consumer",
   "vram": 12,
   "bw": 384,
   "arch": "RDNA 2",
   "released": "2021-03",
   "notes": "A cheap used 12 GB card that works well with llama.cpp's Vulkan backend. ROCm does not officially support it, so tools that need ROCm may require workarounds."
  },
  {
   "slug": "rx-6800-xt-16gb",
   "name": "Radeon RX 6800 XT 16GB",
   "brand": "AMD",
   "type": "consumer",
   "vram": 16,
   "bw": 512,
   "arch": "RDNA 2",
   "released": "2020-11",
   "notes": "One of the cheapest ways to get 16 GB on a fast bus. It runs 14B models at high quality and 20–24B models at Q4 with a short context in llama.cpp, Ollama and LM Studio. Image generation and training are slower than on NVIDIA."
  },
  {
   "slug": "rx-7600-8gb",
   "name": "Radeon RX 7600 8GB",
   "brand": "AMD",
   "type": "consumer",
   "vram": 8,
   "bw": 288,
   "arch": "RDNA 3",
   "released": "2023-05",
   "notes": "A budget AMD card. llama.cpp's Vulkan backend runs on it without ROCm setup. Expect 8B-class models at Q4_K_M."
  },
  {
   "slug": "rx-7700-xt-12gb",
   "name": "Radeon RX 7700 XT 12GB",
   "brand": "AMD",
   "type": "consumer",
   "vram": 12,
   "bw": 432,
   "arch": "RDNA 3",
   "released": "2023-09",
   "notes": "A 12 GB RDNA 3 card with decent bandwidth. It handles 12–14B models in llama.cpp, but the 16 GB RX 7800 XT costs little more and fits noticeably bigger models."
  },
  {
   "slug": "rx-7800-xt-16gb",
   "name": "Radeon RX 7800 XT 16GB",
   "brand": "AMD",
   "type": "consumer",
   "vram": 16,
   "bw": 624,
   "arch": "RDNA 3",
   "released": "2023-09",
   "notes": "One of the cheapest ways to get 16 GB with good bandwidth. Works with llama.cpp (ROCm or Vulkan), Ollama and LM Studio."
  },
  {
   "slug": "rx-7900-xt-20gb",
   "name": "Radeon RX 7900 XT 20GB",
   "brand": "AMD",
   "type": "consumer",
   "vram": 20,
   "bw": 800,
   "arch": "RDNA 3",
   "released": "2022-12",
   "notes": "20 GB is an uncommon size that fits 24B models at Q5 and 27B models at Q4 with short contexts. Bandwidth is high, so token generation is fast in llama.cpp with ROCm or Vulkan."
  },
  {
   "slug": "rx-7900-xtx-24gb",
   "name": "Radeon RX 7900 XTX 24GB",
   "brand": "AMD",
   "type": "consumer",
   "vram": 24,
   "bw": 960,
   "arch": "RDNA 3",
   "released": "2022-12",
   "ref": true,
   "notes": "AMD's 24 GB flagship. It holds the same models as an RTX 3090 or 4090, and token generation is competitive, but software support for training and image tools still trails CUDA.",
   "tips": "Setup tips: on Linux, install ROCm and use Ollama or llama.cpp built with HIP; on Windows, LM Studio's ROCm or Vulkan runtime is the easiest route. It runs the same 27–35B models as a 3090 or 4090 at Q4_K_M."
  },
  {
   "slug": "rx-9060-xt-16gb",
   "name": "Radeon RX 9060 XT 16GB",
   "brand": "AMD",
   "type": "consumer",
   "vram": 16,
   "bw": 320,
   "arch": "RDNA 4",
   "released": "2025-06",
   "notes": "The cheapest new 16 GB card. Capacity is the draw: it fits the same models as an RTX 5060 Ti 16GB, but its 128-bit bus makes token generation noticeably slower."
  },
  {
   "slug": "rx-9070-xt-16gb",
   "name": "Radeon RX 9070 XT 16GB",
   "brand": "AMD",
   "type": "consumer",
   "vram": 16,
   "bw": 640,
   "arch": "RDNA 4",
   "released": "2025-03",
   "notes": "AMD's RDNA 4 card with much stronger AI compute than RDNA 3. For LLM inference it sits close to an RTX 5070 Ti on models that fit in 16 GB.",
   "tips": "Setup tips: supported by recent ROCm releases on Linux and by LM Studio's Vulkan and ROCm runtimes on Windows. Its bandwidth makes it quick for 14B models and gpt-oss-20b."
  },
  {
   "slug": "arc-a770-16gb",
   "name": "Intel Arc A770 16GB",
   "brand": "Intel",
   "type": "consumer",
   "vram": 16,
   "bw": 560,
   "arch": "Alchemist",
   "released": "2022-10",
   "notes": "16 GB at a low price with good bandwidth. Support has improved through llama.cpp's SYCL and Vulkan backends and Intel's IPEX-LLM, but expect more setup than on NVIDIA and lower real-world efficiency."
  },
  {
   "slug": "arc-b570-10gb",
   "name": "Intel Arc B570 10GB",
   "brand": "Intel",
   "type": "consumer",
   "vram": 10,
   "bw": 380,
   "arch": "Battlemage",
   "released": "2025-01",
   "notes": "A budget Battlemage card with 10 GB, a little more room than 8 GB cards for 8B models with longer context. The 12 GB Arc B580 is the better local-AI pick if the price gap is small."
  },
  {
   "slug": "arc-b580-12gb",
   "name": "Intel Arc B580 12GB",
   "brand": "Intel",
   "type": "consumer",
   "vram": 12,
   "bw": 456,
   "arch": "Battlemage",
   "released": "2024-12",
   "notes": "The cheapest new 12 GB card. It runs llama.cpp via SYCL or Vulkan and Ollama via IPEX-LLM. Software maturity is behind NVIDIA and AMD.",
   "tips": "Setup tips: use llama.cpp's SYCL or Vulkan backend, LM Studio's Vulkan runtime, or Intel's IPEX-LLM builds of Ollama. Update to the latest Intel driver, since performance has improved with each release."
  },
  {
   "slug": "rtx-6000-ada-48gb",
   "name": "RTX 6000 Ada 48GB",
   "brand": "NVIDIA",
   "type": "pro",
   "vram": 48,
   "bw": 960,
   "arch": "Ada Lovelace",
   "released": "2022-12",
   "ref": true,
   "notes": "The Ada-generation 48 GB workstation card. Holds 70B models at Q4_K_M on a single card with ECC memory and a 300 W blower design."
  },
  {
   "slug": "rtx-a6000-48gb",
   "name": "RTX A6000 48GB",
   "brand": "NVIDIA",
   "type": "pro",
   "vram": 48,
   "bw": 768,
   "arch": "Ampere",
   "released": "2020-10",
   "notes": "A 48 GB workstation card, often found used. One card runs 70B models at Q4_K_M with a modest context. Slower than two RTX 3090s, but one slot, lower power and simpler."
  },
  {
   "slug": "rtx-pro-6000-96gb",
   "name": "RTX PRO 6000 Blackwell 96GB",
   "brand": "NVIDIA",
   "type": "pro",
   "vram": 96,
   "bw": 1792,
   "arch": "Blackwell",
   "released": "2025-04",
   "ref": true,
   "notes": "96 GB on one workstation card with RTX 5090-class bandwidth. It runs 70B models at Q8_0 or gpt-oss-120b natively on one card."
  },
  {
   "slug": "radeon-ai-pro-r9700-32gb",
   "name": "Radeon AI PRO R9700 32GB",
   "brand": "AMD",
   "type": "pro",
   "vram": 32,
   "bw": 640,
   "arch": "RDNA 4",
   "released": "2025-07",
   "notes": "A 32 GB workstation version of the RX 9070 XT chip, priced well below NVIDIA's 32 GB options. It holds 32B models at Q6_K or with long contexts and pairs well for 64 GB multi-GPU setups. ROCm supports it officially."
  },
  {
   "slug": "arc-pro-b60-24gb",
   "name": "Intel Arc Pro B60 24GB",
   "brand": "Intel",
   "type": "pro",
   "vram": 24,
   "bw": 456,
   "arch": "Battlemage",
   "released": "2025-06",
   "notes": "A 24 GB Battlemage workstation card aimed at low-cost inference. It fits 27–32B models at Q4_K_M; bandwidth and software maturity make it slower than a used RTX 3090."
  },
  {
   "slug": "a100-40gb",
   "name": "A100 40GB",
   "brand": "NVIDIA",
   "type": "datacenter",
   "vram": 40,
   "bw": 1555,
   "arch": "Ampere",
   "released": "2020-05",
   "notes": "The original A100. HBM2 bandwidth makes it fast, but 40 GB is too small for 70B models at Q4_K_M."
  },
  {
   "slug": "a100-80gb",
   "name": "A100 80GB",
   "brand": "NVIDIA",
   "type": "datacenter",
   "vram": 80,
   "bw": 2039,
   "arch": "Ampere",
   "released": "2020-11",
   "ref": true,
   "notes": "Still widely rented in the cloud. 80 GB holds 70B models at Q6_K or FP8-sized weights, with room for long contexts."
  },
  {
   "slug": "h100-80gb",
   "name": "H100 80GB SXM",
   "brand": "NVIDIA",
   "type": "datacenter",
   "vram": 80,
   "bw": 3350,
   "arch": "Hopper",
   "released": "2022-09",
   "ref": true,
   "notes": "The standard datacenter GPU for LLM serving, with native FP8. Same memory as the A100 80GB but about 1.6 times the bandwidth. PCIe versions have lower bandwidth (2 TB/s)."
  },
  {
   "slug": "h200-141gb",
   "name": "NVIDIA H200 141GB",
   "brand": "NVIDIA",
   "type": "datacenter",
   "vram": 141,
   "bw": 4800,
   "arch": "Hopper",
   "released": "2024-11",
   "notes": "Hopper with 141 GB of HBM3e. One card holds 70B models in FP8 with a long context or 120B-class MoE models comfortably; eight make the standard node for serving DeepSeek and Kimi-class models."
  },
  {
   "slug": "l4-24gb",
   "name": "NVIDIA L4 24GB",
   "brand": "NVIDIA",
   "type": "datacenter",
   "vram": 24,
   "bw": 300,
   "arch": "Ada Lovelace",
   "released": "2023-03",
   "notes": "A low-power 24 GB cloud GPU common on Google Cloud. It holds 27–32B models at Q4_K_M and supports FP8, but 300 GB/s of bandwidth makes it slower per token than consumer 24 GB cards."
  },
  {
   "slug": "l40s-48gb",
   "name": "L40S 48GB",
   "brand": "NVIDIA",
   "type": "datacenter",
   "vram": 48,
   "bw": 864,
   "arch": "Ada Lovelace",
   "released": "2023-08",
   "notes": "A common cloud inference card. Same memory as the RTX 6000 Ada, with passive cooling for servers."
  },
  {
   "slug": "tesla-p40-24gb",
   "name": "Tesla P40 24GB",
   "brand": "NVIDIA",
   "type": "datacenter",
   "vram": 24,
   "bw": 347,
   "arch": "Pascal",
   "released": "2016-09",
   "eff": 0.55,
   "notes": "A cheap used datacenter card with 24 GB. It runs 32B models at Q4_K_M in llama.cpp, but it is slow, has no display output or fan, and lacks fast FP16, so it is poor for image generation and unsupported by many newer tools."
  },
  {
   "slug": "tesla-t4-16gb",
   "name": "Tesla T4 16GB",
   "brand": "NVIDIA",
   "type": "datacenter",
   "vram": 16,
   "bw": 320,
   "arch": "Turing",
   "released": "2018-09",
   "notes": "The 70 W inference card behind many free and cheap cloud notebooks (Google Colab, Kaggle). 16 GB fits 14B models at high quality, but low bandwidth keeps generation modest."
  },
  {
   "slug": "mi300x-192gb",
   "name": "AMD Instinct MI300X 192GB",
   "brand": "AMD",
   "type": "datacenter",
   "vram": 192,
   "bw": 5300,
   "arch": "CDNA 3",
   "released": "2023-12",
   "notes": "AMD's datacenter GPU with 192 GB of HBM3, more than an H100 or H200. Single-card capacity is the main advantage: Llama 3.1 405B at Q2–Q3 or 200B-class MoE models at Q4 fit on a single card. vLLM and SGLang support it through ROCm."
  },
  {
   "slug": "apple-m1-max-64gb",
   "name": "Apple M1 Max 64GB",
   "brand": "Apple",
   "type": "apple",
   "vram": 64,
   "usable": 48,
   "bw": 400,
   "arch": "Apple Silicon",
   "released": "2021-10",
   "notes": "A popular used choice for local LLMs. About 48 GB is GPU-usable by default, enough for 70B models at Q4_K_M with a short context, at roughly three quarters of an M4 Max's speed."
  },
  {
   "slug": "apple-m2-ultra-192gb",
   "name": "Apple M2 Ultra 192GB",
   "brand": "Apple",
   "type": "apple",
   "vram": 192,
   "usable": 144,
   "bw": 800,
   "arch": "Apple Silicon",
   "released": "2023-06",
   "notes": "A Mac Studio or Mac Pro configuration with 800 GB/s and about 144 GB GPU-usable by default. It runs 100–230B MoE models and 70B models at Q8_0; raising the wired-memory limit gives about 180 GB."
  },
  {
   "slug": "apple-m3-max-128gb",
   "name": "Apple M3 Max 128GB",
   "brand": "Apple",
   "type": "apple",
   "vram": 128,
   "usable": 96,
   "bw": 400,
   "arch": "Apple Silicon",
   "released": "2023-11",
   "notes": "Holds the same models as an M4 Max 128GB, about 96 GB GPU-usable by default, but with 400 GB/s instead of 546 GB/s, so it generates about 25% slower. Good value used."
  },
  {
   "slug": "apple-m3-ultra-512gb",
   "name": "Apple M3 Ultra 512GB",
   "brand": "Apple",
   "type": "apple",
   "vram": 512,
   "usable": 384,
   "bw": 819,
   "arch": "Apple Silicon",
   "released": "2025-03",
   "ref": true,
   "notes": "The Mac Studio configuration that runs the largest open models in one box, including DeepSeek R1 at Q4_K_M once you raise the default GPU memory limit (about 75% of RAM). Speed is limited by its 819 GB/s bandwidth.",
   "tips": "Setup tips: with about 384 GB GPU-usable by default (more with a raised wired limit) it can hold Qwen 3.5 397B, MiniMax-M2 or DeepSeek V4 Flash. Long prompts process slowly compared with NVIDIA GPUs, so keep system prompts short and reuse prompt caches where your app supports it."
  },
  {
   "slug": "apple-m4-16gb",
   "name": "Apple M4 16GB",
   "brand": "Apple",
   "type": "apple",
   "vram": 16,
   "usable": 10.7,
   "bw": 120,
   "arch": "Apple Silicon",
   "released": "2024-10",
   "notes": "The base Mac mini and MacBook. macOS lets the GPU use about two thirds of 16 GB by default, so plan for about 10.7 GB. Good for 7B to 8B models at modest speed."
  },
  {
   "slug": "apple-m4-max-36gb",
   "name": "Apple M4 Max 36GB",
   "brand": "Apple",
   "type": "apple",
   "vram": 36,
   "usable": 24,
   "bw": 410,
   "arch": "Apple Silicon",
   "released": "2024-10",
   "notes": "The entry M4 Max (14-core CPU) with 410 GB/s. About 24 GB is GPU-usable by default, similar to a 24 GB GPU for model size but slower per token. Raising the wired-memory limit gives 32B models room for longer contexts."
  },
  {
   "slug": "apple-m4-max-64gb",
   "name": "Apple M4 Max 64GB",
   "brand": "Apple",
   "type": "apple",
   "vram": 64,
   "usable": 48,
   "bw": 546,
   "arch": "Apple Silicon",
   "released": "2024-10",
   "notes": "About 48 GB GPU-usable by default, enough for 70B models at Q4_K_M with a short context. The 546 GB/s bandwidth roughly matches an RTX 4070 Ti Super."
  },
  {
   "slug": "apple-m4-max-128gb",
   "name": "Apple M4 Max 128GB",
   "brand": "Apple",
   "type": "apple",
   "vram": 128,
   "usable": 96,
   "bw": 546,
   "arch": "Apple Silicon",
   "released": "2024-10",
   "ref": true,
   "notes": "The largest laptop memory pool. About 96 GB is GPU-usable by default (more with iogpu.wired_limit_mb), which runs 70B models at Q8_0 or gpt-oss-120b natively.",
   "tips": "Setup tips: raise the GPU memory limit with sudo sysctl iogpu.wired_limit_mb=114688 to give models about 112 GB. Prefer MLX builds in LM Studio and mixture-of-experts models, which make the best use of its bandwidth. Keep the laptop plugged in for long runs."
  },
  {
   "slug": "apple-m4-pro-24gb",
   "name": "Apple M4 Pro 24GB",
   "brand": "Apple",
   "type": "apple",
   "vram": 24,
   "usable": 16,
   "bw": 273,
   "arch": "Apple Silicon",
   "released": "2024-10",
   "notes": "The base Mac mini M4 Pro. Only about 16 GB is GPU-usable by default, so treat it like a 16 GB GPU: 14B models comfortably, 24B models at Q4 with short context."
  },
  {
   "slug": "apple-m4-pro-48gb",
   "name": "Apple M4 Pro 48GB",
   "brand": "Apple",
   "type": "apple",
   "vram": 48,
   "usable": 36,
   "bw": 273,
   "arch": "Apple Silicon",
   "released": "2024-10",
   "notes": "A Mac mini or MacBook Pro configuration that holds 32B models comfortably. About 36 GB is GPU-usable by default; raising the wired-memory limit gives more.",
   "tips": "Setup tips: about 36 GB is GPU-usable by default, so 32B models fit at Q4_K_M with room for context. Mixture-of-experts models such as Qwen 3.6 35B-A3B run especially well on its 273 GB/s."
  },
  {
   "slug": "apple-m5-32gb",
   "name": "Apple M5 32GB",
   "brand": "Apple",
   "type": "apple",
   "vram": 32,
   "usable": 21.3,
   "bw": 153,
   "arch": "Apple Silicon",
   "released": "2025-10",
   "notes": "The base M5 adds neural accelerators to each GPU core, which speeds up prompt processing, but 153 GB/s of bandwidth still limits token generation. About 21 GB is GPU-usable by default."
  },
  {
   "slug": "apple-m5-max-128gb",
   "name": "Apple M5 Max 128GB",
   "brand": "Apple",
   "type": "apple",
   "vram": 128,
   "usable": 96,
   "bw": 614,
   "arch": "Apple Silicon",
   "released": "2026-03",
   "ref": true,
   "notes": "The 40-core M5 Max with 614 GB/s, about 12% more bandwidth than the M4 Max and much faster prompt processing thanks to the GPU neural accelerators. About 96 GB is GPU-usable by default."
  },
  {
   "slug": "apple-m5-pro-64gb",
   "name": "Apple M5 Pro 64GB",
   "brand": "Apple",
   "type": "apple",
   "vram": 64,
   "usable": 48,
   "bw": 307,
   "arch": "Apple Silicon",
   "released": "2026-03",
   "notes": "The top M5 Pro configuration. About 48 GB is GPU-usable by default, enough for 70B models at Q4 with short context and, just, for 80B-class MoE models such as Qwen3-Next, which run fast on its 307 GB/s."
  },
  {
   "slug": "dgx-spark-128gb",
   "name": "NVIDIA DGX Spark 128GB",
   "brand": "NVIDIA",
   "type": "apu",
   "vram": 128,
   "usable": 112,
   "bw": 273,
   "arch": "Grace Blackwell",
   "released": "2025-10",
   "ref": true,
   "notes": "A compact Grace Blackwell system with 128 GB of unified memory and the full CUDA stack. Like Strix Halo, its 273 GB/s bandwidth favours MoE models over large dense ones.",
   "tips": "Setup tips: it runs NVIDIA's DGX OS with CUDA preinstalled, so vLLM, TensorRT-LLM and PyTorch work as on a datacenter GPU. Use it for MoE models and for fine-tuning experiments; dense 70B models run, but slowly, because of its 273 GB/s bandwidth."
  },
  {
   "slug": "ryzen-ai-max-395-128gb",
   "name": "Ryzen AI Max+ 395 128GB",
   "brand": "AMD",
   "type": "apu",
   "vram": 128,
   "usable": 96,
   "bw": 256,
   "arch": "Strix Halo",
   "released": "2025-02",
   "ref": true,
   "notes": "AMD's Strix Halo APU with up to 96 GB assignable to the GPU in firmware on Windows, and more on Linux. It holds models that need two or three discrete GPUs, but 256 GB/s of bandwidth makes large dense models slow. MoE models suit it best.",
   "tips": "Setup tips: set the GPU memory allocation (UMA frame buffer or variable graphics memory) in the BIOS or driver so the GPU can use up to 96 GB. Use llama.cpp's Vulkan backend or ROCm on Linux. Large mixture-of-experts models are the best match for its 256 GB/s."
  }
 ]
}