{
  "benchmarks": [
    {
      "id": "rx-6800m-qwen-27b-ridge",
      "gpu": "RX 6800M",
      "gpu_vram": "12 GB",
      "gpu_bw": "~384 GB/s",
      "model": "Qwen 27B Ridge",
      "params": "27B (dense)",
      "quant": "3.7bpw",
      "context": "16K",
      "kv_quant": "q8_0",
      "offload": true,
      "tested": true,
      "source_article": "qwen-27b-ridge-rocm-vs-vulkan.html",
      "backends": [
        {
          "name": "ROCm",
          "decode": 18.12,
          "prompt": 217.6
        },
        {
          "name": "Vulkan",
          "decode": 21.84,
          "prompt": 149.6
        }
      ],
      "notes": "Not fully resident on 12 GB (~15.6 GB needed). Partial offload to system RAM explains the ~18-21 tok/s decode. Vulkan wins decode (+20%), ROCm wins prompt eval (+45%)."
    },
    {
      "id": "rx-6800m-ornith-35b-moe",
      "gpu": "RX 6800M",
      "gpu_vram": "12 GB",
      "gpu_bw": "~384 GB/s",
      "model": "Ornith 35B-A3B MoE",
      "params": "35B (3B active)",
      "quant": "Q5_K/Q4_K",
      "context": "262K",
      "kv_quant": "q8_0",
      "offload": true,
      "tested": true,
      "source_article": "ornith-35b-moe-262k-rocm-vs-vulkan.html",
      "backends": [
        {
          "name": "ROCm",
          "decode": 25.62,
          "prompt": 83.8
        },
        {
          "name": "Vulkan",
          "decode": 19.66,
          "prompt": 83.1
        }
      ],
      "notes": "28 CPU experts, MoE-sparse + q8 KV + CPU offload. ROCm wins decode (+30%). Decodes ~3B active per token, so speed holds even at 262K. Needs --load-mode none, no -ngl 999."
    },
    {
      "id": "rx-6800m-tiel-coder-35b-mtp",
      "gpu": "RX 6800M",
      "gpu_vram": "12 GB",
      "gpu_bw": "~384 GB/s",
      "model": "Tiel-Coder 35B-A3B",
      "params": "35B (3B active)",
      "quant": "Q4_K_XL",
      "context": "262K",
      "kv_quant": "q8_0",
      "offload": true,
      "tested": true,
      "source_article": "tiel-coder-35b-mtp-rx6800m.html",
      "backends": [
        {
          "name": "ROCm (no MTP)",
          "decode": 25.39,
          "prompt": 118.96
        },
        {
          "name": "ROCm (MTP)",
          "decode": 29.09,
          "prompt": 111.4
        }
      ],
      "notes": "MTP speculative decoding on MoE (28 vs 32 CPU experts). MTP +14.6% decode (25.39 \u2192 29.09 tok/s), prompt -6.4%, 54.7% acceptance (2.64 mean). 35B Q4 ~17.5GB + q8 KV 262K needs --n-cpu-moe offload on 12GB."
    },
    {
      "id": "rx-6800m-qwen38-27b-gsq-rco",
      "gpu": "RX 6800M",
      "gpu_vram": "12 GB",
      "gpu_bw": "~384 GB/s",
      "model": "Qwen3.8 27B GSQ-RCO",
      "params": "27B (GDN hybrid)",
      "quant": "IQ2_XS / IQ3_S",
      "context": "52K / 4K",
      "kv_quant": "q8_0",
      "offload": false,
      "tested": true,
      "source_article": "qwen38-27b-gsq-rco-rx6800m.html",
      "backends": [
        {
          "name": "ROCm (IQ2_XS-mtp)",
          "decode": 20.77,
          "prompt": 113.51
        },
        {
          "name": "ROCm (IQ2_XS-mtp + MTP)",
          "decode": 23.07,
          "prompt": 104.54
        },
        {
          "name": "ROCm (IQ3_S)",
          "decode": 18.97,
          "prompt": 178.32
        },
        {
          "name": "ROCm (IQ3_S, auto-fit)",
          "decode": 7.3,
          "prompt": 112.03
        }
      ],
      "notes": "IQ2_XS wins decode (+9%) and fits 52K ctx with q8_0 KV under --fit. Enabling the MTP head (--spec-type draft-mtp) lifts IQ2_XS to 23.07 t/s (48% acceptance, mean len 2.44) but ROCm lacks TOP_K for the draft sampler and the draft context makes --fit bail to 4K. IQ3_S at -ngl 999 wins prompt eval (+57%) yet only fits a 4K ctx at 97% VRAM; letting --fit choose (no -ngl) is worse, not better: it offloads layer 0 to CPU and the fused Gated Delta Net kernel is unsupported on ROCm, collapsing decode to 7.3 t/s. Manual -ngl 999 is the correct config for IQ3_S on this card."
    },
    {
      "id": "rx-6800m-ternary-bonsai-2-27b",
      "gpu": "RX 6800M",
      "gpu_vram": "12 GB",
      "gpu_bw": "~384 GB/s",
      "model": "Ternary-Bonsai-2-27B",
      "params": "27B (ternary hybrid)",
      "quant": "PQ2_0 / PTQ1_0",
      "context": "32K",
      "kv_quant": "f16",
      "offload": false,
      "tested": true,
      "source_article": "ternary-bonsai-2-27b-rx6800m.html",
      "backends": [
        {
          "name": "ROCm (PQ2_0, gfx1031 build)",
          "decode": 32.8,
          "prompt": 64.2
        },
        {
          "name": "ROCm (PTQ1_0, gfx1031 build)",
          "decode": 18.8,
          "prompt": 22.5
        }
      ],
      "notes": "Stock llama.cpp rejects type 142/143; Prism fork b10709 built with -DCMAKE_HIP_ARCHITECTURES=gfx1031 required (prebuilt ROCm 7.2 gives 'device kernel image is invalid' on gfx1031). PQ2_0 wins prompt +185% and decode +74% over PTQ1_0 on RDNA2; PTQ1_0 saves 1.3GB (5.95 vs 7.21GB)."
    },
    {
      "id": "rtx-4090-qwen-7b-q4",
      "gpu": "RTX 4090",
      "gpu_vram": "24 GB",
      "gpu_bw": "1008 GB/s",
      "model": "Qwen 7B",
      "params": "7B",
      "quant": "Q4",
      "context": "8K",
      "kv_quant": "Q4",
      "offload": false,
      "tested": false,
      "source_article": "best-gpu-for-local-llm.html",
      "backends": [
        {
          "name": "CUDA",
          "decode": 160,
          "prompt": null
        }
      ],
      "notes": "Estimated from the GPU AI Performance Calculator. Fully resident."
    },
    {
      "id": "rtx-4090-qwen-14b-q4",
      "gpu": "RTX 4090",
      "gpu_vram": "24 GB",
      "gpu_bw": "1008 GB/s",
      "model": "Qwen 14B",
      "params": "14B",
      "quant": "Q4",
      "context": "8K",
      "kv_quant": "Q4",
      "offload": false,
      "tested": false,
      "source_article": "best-gpu-for-local-llm.html",
      "backends": [
        {
          "name": "CUDA",
          "decode": 120,
          "prompt": null
        }
      ],
      "notes": "Estimated from the GPU AI Performance Calculator. Fully resident."
    },
    {
      "id": "rtx-4090-qwen-27b-q4",
      "gpu": "RTX 4090",
      "gpu_vram": "24 GB",
      "gpu_bw": "1008 GB/s",
      "model": "Qwen 27B",
      "params": "27B",
      "quant": "Q4",
      "context": "8K",
      "kv_quant": "Q4",
      "offload": false,
      "tested": false,
      "source_article": "best-gpu-for-local-llm.html",
      "backends": [
        {
          "name": "CUDA",
          "decode": 60,
          "prompt": null
        }
      ],
      "notes": "Estimated. Tight but resident on 24 GB."
    },
    {
      "id": "rx-7900-xtx-qwen-27b-q4",
      "gpu": "RX 7900 XTX",
      "gpu_vram": "24 GB",
      "gpu_bw": "960 GB/s",
      "model": "Qwen 27B",
      "params": "27B",
      "quant": "Q4",
      "context": "8K",
      "kv_quant": "Q4",
      "offload": false,
      "tested": false,
      "source_article": "best-gpu-for-local-llm.html",
      "backends": [
        {
          "name": "ROCm",
          "decode": 58,
          "prompt": null
        }
      ],
      "notes": "Estimated. Best VRAM-per-rupee on the AMD side."
    },
    {
      "id": "rtx-5070-ti-qwen-7b-q4",
      "gpu": "RTX 5070 Ti",
      "gpu_vram": "16 GB",
      "gpu_bw": "896 GB/s",
      "model": "Qwen 7B",
      "params": "7B",
      "quant": "Q4",
      "context": "8K",
      "kv_quant": "Q4",
      "offload": false,
      "tested": false,
      "source_article": "best-gpu-for-local-llm.html",
      "backends": [
        {
          "name": "CUDA",
          "decode": 95,
          "prompt": null
        }
      ],
      "notes": "Estimated from the GPU AI Performance Calculator."
    },
    {
      "id": "rtx-5070-ti-qwen-14b-q4",
      "gpu": "RTX 5070 Ti",
      "gpu_vram": "16 GB",
      "gpu_bw": "896 GB/s",
      "model": "Qwen 14B",
      "params": "14B",
      "quant": "Q4",
      "context": "8K",
      "kv_quant": "Q4",
      "offload": false,
      "tested": false,
      "source_article": "best-gpu-for-local-llm.html",
      "backends": [
        {
          "name": "CUDA",
          "decode": 90,
          "prompt": null
        }
      ],
      "notes": "Estimated. Fully resident on 16 GB."
    },
    {
      "id": "rx-9070-xt-qwen-7b-q4",
      "gpu": "RX 9070 XT",
      "gpu_vram": "16 GB",
      "gpu_bw": "644 GB/s",
      "model": "Qwen 7B",
      "params": "7B",
      "quant": "Q4",
      "context": "8K",
      "kv_quant": "Q4",
      "offload": false,
      "tested": false,
      "source_article": "best-gpu-for-local-llm.html",
      "backends": [
        {
          "name": "ROCm",
          "decode": 72,
          "prompt": null
        }
      ],
      "notes": "Estimated from the GPU AI Performance Calculator."
    },
    {
      "id": "rx-9070-xt-qwen-14b-q4",
      "gpu": "RX 9070 XT",
      "gpu_vram": "16 GB",
      "gpu_bw": "644 GB/s",
      "model": "Qwen 14B",
      "params": "14B",
      "quant": "Q4",
      "context": "8K",
      "kv_quant": "Q4",
      "offload": false,
      "tested": false,
      "source_article": "best-gpu-for-local-llm.html",
      "backends": [
        {
          "name": "ROCm",
          "decode": 68,
          "prompt": null
        }
      ],
      "notes": "Estimated. Fully resident on 16 GB."
    },
    {
      "id": "rtx-4060-ti-qwen-7b-q4",
      "gpu": "RTX 4060 Ti",
      "gpu_vram": "16 GB",
      "gpu_bw": "288 GB/s",
      "model": "Qwen 7B",
      "params": "7B",
      "quant": "Q4",
      "context": "8K",
      "kv_quant": "Q4",
      "offload": false,
      "tested": false,
      "source_article": "best-gpu-for-local-llm.html",
      "backends": [
        {
          "name": "CUDA",
          "decode": 42,
          "prompt": null
        }
      ],
      "notes": "Estimated. Low bandwidth limits decode."
    },
    {
      "id": "rx-7600-xt-qwen-7b-q4",
      "gpu": "RX 7600 XT",
      "gpu_vram": "16 GB",
      "gpu_bw": "288 GB/s",
      "model": "Qwen 7B",
      "params": "7B",
      "quant": "Q4",
      "context": "8K",
      "kv_quant": "Q4",
      "offload": false,
      "tested": false,
      "source_article": "best-gpu-for-local-llm.html",
      "backends": [
        {
          "name": "ROCm",
          "decode": 42,
          "prompt": null
        }
      ],
      "notes": "Estimated from the GPU AI Performance Calculator."
    },
    {
      "id": "rx-6800m-spark-4b-bf16",
      "gpu": "RX 6800M",
      "gpu_vram": "12 GB",
      "gpu_bw": "~384 GB/s",
      "model": "Spark-X2.5 4B",
      "params": "4.1B",
      "quant": "BF16",
      "context": "101K",
      "kv_quant": "f16",
      "offload": true,
      "tested": true,
      "source_article": "spark-x2-5-quants-rx6800m.html",
      "backends": [
        {
          "name": "ROCm",
          "decode": 39.11,
          "prompt": 474.36
        }
      ],
      "notes": "Measured -p 512 -n 128 -r 3, per-test isolated telemetry (benchscope). Slowest and hottest (62 C); the only quant that failed the JSON quality task. Full weights: 7.66 GiB."
    },
    {
      "id": "rx-6800m-spark-4b-q8-0",
      "gpu": "RX 6800M",
      "gpu_vram": "12 GB",
      "gpu_bw": "~384 GB/s",
      "model": "Spark-X2.5 4B",
      "params": "4.1B",
      "quant": "Q8_0",
      "context": "206K",
      "kv_quant": "f16",
      "offload": true,
      "tested": true,
      "source_article": "spark-x2-5-quants-rx6800m.html",
      "backends": [
        {
          "name": "ROCm",
          "decode": 67.39,
          "prompt": 2148.7
        }
      ],
      "notes": "Measured -p 512 -n 128 -r 3, per-test isolated telemetry (benchscope). Prompt-eval king (4.5x BF16) but high run variance (+/-215 tok/s); decode trails Q4_K_M. Full weights: 4.07 GiB."
    },
    {
      "id": "rx-6800m-spark-4b-q4-k-m",
      "gpu": "RX 6800M",
      "gpu_vram": "12 GB",
      "gpu_bw": "~384 GB/s",
      "model": "Spark-X2.5 4B",
      "params": "4.1B",
      "quant": "Q4_K_M",
      "context": "254K",
      "kv_quant": "f16",
      "offload": true,
      "tested": true,
      "source_article": "spark-x2-5-quants-rx6800m.html",
      "backends": [
        {
          "name": "ROCm",
          "decode": 98.47,
          "prompt": 1510.64
        }
      ],
      "notes": "Measured -p 512 -n 128 -r 3, per-test isolated telemetry (benchscope). Best decode (2.5x BF16), biggest context budget, coolest card (53 C); matched BF16 on every quality task. Full weights: 2.42 GiB."
    }
  ]
}