Add Arc Pro B70 results: Qwen3.6-27B and Qwen3.8-27B

results/shacortes/qwen3-6-27b-llamacpp-q4_k_m-sycl.json

@@ -0,0 +1,16 @@
+{
+  "$schema": "https://raw.githubusercontent.com/labscommunity/intelinside/main/results/schema.json",
+  "rig": "3",
+  "component": "intel-arc-pro-b70",
+  "componentQuantity": 1,
+  "model": "qwen3-6-27b",
+  "quant": "q4_k_m",
+  "runtime": "llamacpp",
+  "runtimeVersion": "b11026 (b49650adb)",
+  "runtimeFlags": "sycl backend, -ngl 99 -fa 1 -p 512 -n 256 -r 5, single request",
+  "decodeTps": 26.635976,
+  "promptTps": 312.574979,
+  "batchSize": 1,
+  "runDate": "2026-09-17",
+  "notes": "Stock upstream ggml-org/llama.cpp at b49650adb, clean tree, sycl backend. Weights unsloth/Qwen3.6-27B-GGUF. decodeTps is tg256 averaged over 5 runs, one request at a time."
+}

results/shacortes/qwen3-6-27b-llamacpp-q4_k_m-vulkan.json

@@ -0,0 +1,16 @@
+{
+  "$schema": "https://raw.githubusercontent.com/labscommunity/intelinside/main/results/schema.json",
+  "rig": "3",
+  "component": "intel-arc-pro-b70",
+  "componentQuantity": 1,
+  "model": "qwen3-6-27b",
+  "quant": "q4_k_m",
+  "runtime": "llamacpp",
+  "runtimeVersion": "b11026 (b49650adb)",
+  "runtimeFlags": "vulkan backend, -ngl 99 -fa 1 -p 512 -n 256 -r 5, single request",
+  "decodeTps": 11.492122,
+  "promptTps": 679.700179,
+  "batchSize": 1,
+  "runDate": "2026-09-17",
+  "notes": "Stock upstream ggml-org/llama.cpp at b49650adb, clean tree, vulkan backend. Weights unsloth/Qwen3.6-27B-GGUF. decodeTps is tg256 averaged over 5 runs, one request at a time."
+}

results/shacortes/qwen3-6-27b-vllm-int4.json

@@ -0,0 +1,17 @@
+{
+  "$schema": "https://raw.githubusercontent.com/labscommunity/intelinside/main/results/schema.json",
+  "rig": "3",
+  "component": "intel-arc-pro-b70",
+  "componentQuantity": 1,
+  "model": "qwen3-6-27b",
+  "quant": "int4",
+  "runtime": "vllm",
+  "runtimeVersion": "0.29.0 (v0.29.0 98dff2a81)",
+  "runtimeFlags": "FLASH_ATTN v2, chunked prefill on, prefix caching on, fp16, kv cache auto, --max-num-seqs 1 --gpu-memory-utilization 0.92 --max-model-len 4096, single request",
+  "decodeTps": 27.96,
+  "ttftMs": 98.07,
+  "contextLength": 4096,
+  "batchSize": 1,
+  "runDate": "2026-09-17",
+  "notes": "Stock upstream vllm-project/vllm v0.29.0 (98dff2a81), clean checkout in its own venv, vllm_xpu_kernels from PyPI, plugins disabled. Weights cyankiwi/Qwen3.6-27B-AWQ-INT4. 8 prompts of 64 tokens, 256 tokens out each, one request at a time. decodeTps = 1000/mean_tpot_ms (35.77 ms), matching llama.cpp tg256; vLLM output_throughput over total wall time was 27.77 tok/s."
+}

results/shacortes/qwen3-8-27b-llamacpp-q4_k_m-sycl.json

@@ -0,0 +1,16 @@
+{
+  "$schema": "https://raw.githubusercontent.com/labscommunity/intelinside/main/results/schema.json",
+  "rig": "3",
+  "component": "intel-arc-pro-b70",
+  "componentQuantity": 1,
+  "model": "qwen3-8-27b",
+  "quant": "q4_k_m",
+  "runtime": "llamacpp",
+  "runtimeVersion": "b11026 (b49650adb)",
+  "runtimeFlags": "sycl backend, -ngl 99 -fa 1 -p 512 -n 256 -r 5, single request",
+  "decodeTps": 20.354646,
+  "promptTps": 288.562413,
+  "batchSize": 1,
+  "runDate": "2026-09-17",
+  "notes": "Stock upstream ggml-org/llama.cpp at b49650adb, clean tree, sycl backend. Weights unsloth/Qwen3.8-27B-GGUF. decodeTps is tg256 averaged over 5 runs, one request at a time."
+}

results/shacortes/qwen3-8-27b-llamacpp-q4_k_m-vulkan.json

@@ -0,0 +1,16 @@
+{
+  "$schema": "https://raw.githubusercontent.com/labscommunity/intelinside/main/results/schema.json",
+  "rig": "3",
+  "component": "intel-arc-pro-b70",
+  "componentQuantity": 1,
+  "model": "qwen3-8-27b",
+  "quant": "q4_k_m",
+  "runtime": "llamacpp",
+  "runtimeVersion": "b11026 (b49650adb)",
+  "runtimeFlags": "vulkan backend, -ngl 99 -fa 1 -p 512 -n 256 -r 5, single request",
+  "decodeTps": 14.810784,
+  "promptTps": 679.554036,
+  "batchSize": 1,
+  "runDate": "2026-09-17",
+  "notes": "Stock upstream ggml-org/llama.cpp at b49650adb, clean tree, vulkan backend. Weights unsloth/Qwen3.8-27B-GGUF. decodeTps is tg256 averaged over 5 runs, one request at a time."
+}

results/shacortes/qwen3-8-27b-vllm-int4.json

@@ -0,0 +1,17 @@
+{
+  "$schema": "https://raw.githubusercontent.com/labscommunity/intelinside/main/results/schema.json",
+  "rig": "3",
+  "component": "intel-arc-pro-b70",
+  "componentQuantity": 1,
+  "model": "qwen3-8-27b",
+  "quant": "int4",
+  "runtime": "vllm",
+  "runtimeVersion": "0.29.0 (v0.29.0 98dff2a81)",
+  "runtimeFlags": "FLASH_ATTN v2, chunked prefill on, prefix caching on, bf16, kv cache auto, --max-num-seqs 1 --gpu-memory-utilization 0.92 --max-model-len 4096, single request",
+  "decodeTps": 27.01,
+  "ttftMs": 107.88,
+  "contextLength": 4096,
+  "batchSize": 1,
+  "runDate": "2026-09-17",
+  "notes": "Stock upstream vllm-project/vllm v0.29.0 (98dff2a81), clean checkout in its own venv, vllm_xpu_kernels from PyPI, plugins disabled. Weights cyankiwi/Qwen3.8-27B-AWQ-INT4. 8 prompts of 64 tokens, 256 tokens out each, one request at a time. decodeTps = 1000/mean_tpot_ms (37.02 ms), matching llama.cpp tg256; vLLM output_throughput over total wall time was 26.81 tok/s."
+}