{ "meta": { "title": "pc.l3.ai — UK Local LLM Hardware Observatory", "subtitle": "Building or buying >=128 GB unified/system memory under £3,000 in the UK during the October 2026 3x DRAM surge.", "updatedAt": "2026-10-23", "budgetCapGbp": 3000, "minMemoryGb": 128 }, "dramCrisis": { "headline": "Between Oct 2025 and Oct 2026, consumer DDR5 128GB kits jumped +478% YoY (£3.17/GB → £18.31/GB). Buying 2-channel consumer DDR5 now consumes 73% of a £3,000 budget before buying a GPU.", "ddr5MedianGbpPerGb": 17.45, "ddr5YoYPercent": 272, "ddr5_128gbMedianGbpPerGb": 18.31, "ddr5_128gbYoYPercent": 478, "ddr4MedianGbpPerGb": 8.22, "ddr4YoYPercent": 134, "ddr4EccGbpPerGb": 3.75, "ddr4EccYoYPercent": 103, "monthlySeries": [ { "month": "2025-10", "ddr5_128gb": 3.17, "ddr5_median": 4.69, "ddr4_median": 3.51, "ddr4_ecc_rdimm": 1.85 }, { "month": "2025-11", "ddr5_128gb": 3.45, "ddr5_median": 4.95, "ddr4_median": 3.68, "ddr4_ecc_rdimm": 1.92 }, { "month": "2025-12", "ddr5_128gb": 4.1, "ddr5_median": 5.4, "ddr4_median": 3.9, "ddr4_ecc_rdimm": 2.05 }, { "month": "2026-01", "ddr5_128gb": 5.25, "ddr5_median": 6.3, "ddr4_median": 4.25, "ddr4_ecc_rdimm": 2.2 }, { "month": "2026-02", "ddr5_128gb": 6.8, "ddr5_median": 7.55, "ddr4_median": 4.7, "ddr4_ecc_rdimm": 2.4 }, { "month": "2026-03", "ddr5_128gb": 8.9, "ddr5_median": 9.1, "ddr4_median": 5.2, "ddr4_ecc_rdimm": 2.65 }, { "month": "2026-04", "ddr5_128gb": 11.2, "ddr5_median": 10.85, "ddr4_median": 5.8, "ddr4_ecc_rdimm": 2.85 }, { "month": "2026-05", "ddr5_128gb": 13.6, "ddr5_median": 12.9, "ddr4_median": 6.45, "ddr4_ecc_rdimm": 3.1 }, { "month": "2026-06", "ddr5_128gb": 15.1, "ddr5_median": 14.4, "ddr4_median": 6.95, "ddr4_ecc_rdimm": 3.25 }, { "month": "2026-07", "ddr5_128gb": 16.05, "ddr5_median": 15.17, "ddr4_median": 7.27, "ddr4_ecc_rdimm": 3.4 }, { "month": "2026-08", "ddr5_128gb": 16.9, "ddr5_median": 16.05, "ddr4_median": 7.6, "ddr4_ecc_rdimm": 3.55 }, { "month": "2026-09", "ddr5_128gb": 17.65, "ddr5_median": 16.8, "ddr4_median": 7.95, "ddr4_ecc_rdimm": 3.65 }, { "month": "2026-10", "ddr5_128gb": 18.31, "ddr5_median": 17.45, "ddr4_median": 8.22, "ddr4_ecc_rdimm": 3.75 } ] }, "systems": [ { "id": "epyc-milan-3090", "name": "Custom EPYC 7443 Milan + RTX 3090 24GB + 128GB 8-Ch DDR4-3200", "shortName": "EPYC 7443 + RTX 3090 (128GB 8-Ch)", "category": "epyc-server", "categoryLabel": "Enterprise Hybrid MoE DIY", "condition": "Hybrid (New + Used)", "priceGbp": 2834, "underBudget": true, "totalRamGb": 152, "usableAiMemoryGb": 144, "vramGb": 24, "primaryBandwidthGbs": 936, "secondaryBandwidthGbs": 205, "bandwidthLabel": "936 GB/s VRAM + 205 GB/s 8-Ch RAM", "powerWatts": 520, "noiseDb": "34 dBA (Noctua Tower)", "ukSource": "eBay UK + CeX UK + Scan UK", "stockStatus": "Parts in stock across UK sellers", "verdictBadge": "Best DIY Hybrid MoE Overall", "summary": "Combines 24GB GDDR6X (936 GB/s) for attention/KV cache & fast prompt prefill with 128GB 8-channel DDR4-3200 ECC RDIMM (205 GB/s) for sparse MoE routed experts at 1/4th the cost of consumer DDR5.", "pros": [ "128GB 8-channel DDR4-3200 ECC RDIMM costs £480 vs £2,199 for 2-channel DDR5", "RTX 3090 24GB delivers 936 GB/s VRAM bandwidth & fast CUDA FlashAttention prefill", "Runs DeepSeek V3/R1, Qwen3.5-122B-A10B & gpt-oss-120B via KTransformers / ik_llama", "Full PCIe 4.0 x16 lanes across 7 slots on Gigabyte MZ32-AR0 for future 2nd/3rd 3090" ], "caveats": [ "Verify EPYC 7443 CPU is unlocked (not Dell/Lenovo PSB fuse-locked)", "Dense 70B models (>24GB) that spill to CPU RAM decode at ~11–13 tok/s (MoE models hit 26–42 tok/s)" ], "costSplit": { "ramGbp": 480, "gpuGbp": 975, "cpuBoardGbp": 814, "chassisPsuStorageGbp": 565 }, "benchmarks": { "qwen35bMoeDecode": 78, "qwen35bMoePrefill": 1850, "llama70bDenseDecode": 13, "llama70bDensePrefill": 540, "moe120bDecode": 32, "moe120bPrefill": 680 } }, { "id": "mac-studio-m1-ultra-128", "name": "Apple Mac Studio M1 Ultra (20C CPU / 48–64C GPU / 128GB Unified)", "shortName": "Mac Studio M1 Ultra 128GB (Used)", "category": "apple-silicon", "categoryLabel": "Apple Silicon Unified", "condition": "Refurbished / Used", "priceGbp": 2450, "underBudget": true, "totalRamGb": 128, "usableAiMemoryGb": 114, "vramGb": 114, "primaryBandwidthGbs": 800, "secondaryBandwidthGbs": 800, "bandwidthLabel": "800 GB/s Unified LPDDR5", "powerWatts": 85, "noiseDb": "22 dBA (Near-silent)", "ukSource": "CeX UK (5-Yr Warranty) / Hoxton Macs", "stockStatus": "High turnover — £1,895 Hoxton / £2,450–£2,925 CeX", "verdictBadge": "800 GB/s Dense 70B King", "summary": "Unmatched 800 GB/s unified memory bandwidth under £3,000. Beats Strix Halo by 2.8x on 70B dense token generation while drawing under 90W at whisper-quiet acoustics.", "pros": [ "800 GB/s memory bandwidth — fastest 70B dense decode (~19 tok/s) under £3,000", "114GB allocatable to Metal/MLX via sysctl iogpu.wired_limit_mb", "CeX UK includes mandatory 5-Year Warranty (£2,450–£2,925 Grade B/C)", "Draws <90W under full LLM load; tiny footprint" ], "caveats": [ "Slow prompt prefill (150–300 tok/s) on long >16k contexts due to M1 GPU compute & no SME", "No hardware bfloat16 (use Q4/Q5/Q8 MLX or FP16 weights to avoid bf16 emulation slowdown)" ], "costSplit": { "ramGbp": 0, "gpuGbp": 0, "cpuBoardGbp": 2450, "chassisPsuStorageGbp": 0 }, "benchmarks": { "qwen35bMoeDecode": 56, "qwen35bMoePrefill": 420, "llama70bDenseDecode": 19, "llama70bDensePrefill": 185, "moe120bDecode": 29, "moe120bPrefill": 240 } }, { "id": "strix-halo-uk-retail", "name": "Bosgame M5 / GMKtec EVO-X2 (Ryzen AI Max+ 395, 128GB LPDDR5X-8000, 2TB)", "shortName": "Strix Halo 128GB (UK Retail)", "category": "strix-halo", "categoryLabel": "Unified Mini PC (Strix Halo)", "condition": "New", "priceGbp": 2914, "underBudget": true, "totalRamGb": 128, "usableAiMemoryGb": 120, "vramGb": 120, "primaryBandwidthGbs": 256, "secondaryBandwidthGbs": 256, "bandwidthLabel": "256 GB/s 256-bit LPDDR5X-8000", "powerWatts": 120, "noiseDb": "38 dBA", "ukSource": "Amazon UK (Bosgame £3,099 - 6% voucher / GMKtec £2,999)", "stockStatus": "In Stock on Amazon UK (Next-day delivery)", "verdictBadge": "Best New Turnkey 128GB (UK Stock)", "summary": "Brand-new x86/Linux Mini PC with 16 Zen 5 cores, 40-CU Radeon 8060S GPU, and 128GB LPDDR5X-8000. Unlocks up to 120GB usable VRAM on Linux 6.16+ via TTM page limits.", "pros": [ "Up to 120GB usable GPU memory on Linux (options ttm pages_limit=31457280)", "81–85 tok/s decode & 750 tok/s prefill on Qwen 3 30B-A3B MoE (Vulkan RADV / ROCm)", "Runs 120B MoE (gpt-oss-120b / DeepSeek V4 Flash) entirely in unified memory at 24–27 tok/s", "Standard x86_64 Fedora/Ubuntu compatibility + 120W wall power" ], "caveats": [ "256 GB/s bandwidth limits 70B dense models to ~14–16 tok/s (vs 19 tok/s on M1 Ultra)", "Price doubled in 2026 from £1,499 launch due to LPDDR5X shortage" ], "costSplit": { "ramGbp": 0, "gpuGbp": 0, "cpuBoardGbp": 2914, "chassisPsuStorageGbp": 0 }, "benchmarks": { "qwen35bMoeDecode": 82, "qwen35bMoePrefill": 755, "llama70bDenseDecode": 15, "llama70bDensePrefill": 290, "moe120bDecode": 25, "moe120bPrefill": 380 } }, { "id": "strix-halo-direct-import", "name": "GMKtec EVO-X2 / Bosgame M5 128GB (Direct EU/CN Hub Import)", "shortName": "Strix Halo 128GB (Direct Import)", "category": "strix-halo", "categoryLabel": "Unified Mini PC (Strix Halo)", "condition": "New", "priceGbp": 1950, "underBudget": true, "totalRamGb": 128, "usableAiMemoryGb": 120, "vramGb": 120, "primaryBandwidthGbs": 256, "secondaryBandwidthGbs": 256, "bandwidthLabel": "256 GB/s 256-bit LPDDR5X-8000", "powerWatts": 120, "noiseDb": "38 dBA", "ukSource": "GMKtec / Bosgame Direct via EU Warehouse", "stockStatus": "In Stock (5–9 day shipping)", "verdictBadge": "Best Value New 128GB (£1,050 Under Cap)", "summary": "Identical Ryzen AI Max+ 395 128GB hardware sourced directly via vendor EU/CN logistics hubs, bypassing the ~£1,000 UK retail distributor markup.", "pros": [ "Saves ~£960–£1,150 compared to Amazon UK / UK retail listings", "Leaves £1,050 budget headroom for 4TB NVMe upgrades or USB4/Oculink networking", "Full 120GB usable VRAM under Linux 6.16+ with Mesa RADV Vulkan" ], "caveats": [ "Return-to-base vendor warranty rather than Amazon UK / CeX local returns", "Same 256 GB/s memory bandwidth ceiling on dense models" ], "costSplit": { "ramGbp": 0, "gpuGbp": 0, "cpuBoardGbp": 1950, "chassisPsuStorageGbp": 0 }, "benchmarks": { "qwen35bMoeDecode": 82, "qwen35bMoePrefill": 755, "llama70bDenseDecode": 15, "llama70bDensePrefill": 290, "moe120bDecode": 25, "moe120bPrefill": 380 } }, { "id": "dell-precision-7920-3090", "name": "Dell Precision 7920 Tower + RTX 3090 24GB + 192GB 12-Ch DDR4 ECC", "shortName": "Dell 7920 (192GB 12-Ch) + RTX 3090", "category": "epyc-server", "categoryLabel": "Enterprise Hybrid MoE DIY", "condition": "Refurbished / Used", "priceGbp": 2280, "underBudget": true, "totalRamGb": 216, "usableAiMemoryGb": 200, "vramGb": 24, "primaryBandwidthGbs": 936, "secondaryBandwidthGbs": 210, "bandwidthLabel": "936 GB/s VRAM + 210 GB/s 12-Ch RAM", "powerWatts": 560, "noiseDb": "36 dBA (Baffled Chassis)", "ukSource": "Bargain Hardware UK + CeX UK", "stockStatus": "In Stock at Bargain Hardware UK", "verdictBadge": "Best Sub-£2,300 Capacity (216GB Total)", "summary": "Refurbished dual-socket Dell enterprise tower with 1400W Platinum PSU, 192GB 12-channel DDR4 ECC RDIMM, and a used RTX 3090 24GB — £720 below the £3,000 ceiling.", "pros": [ "192GB 12-channel DDR4 ECC + 24GB VRAM = 216GB total memory for £2,280", "Includes 1400W 80+ Platinum hot-swap PSU and acoustic-baffled workstation chassis", "Room and PSU headroom to drop in a second RTX 3090 later for 48GB VRAM" ], "caveats": [ "Dual-socket NUMA requires pinning threads or running --numa distribute in llama.cpp", "Requires genuine Dell 8-pin GPU power harness (part D92C9 / GYF0C, ~£19)" ], "costSplit": { "ramGbp": 520, "gpuGbp": 975, "cpuBoardGbp": 570, "chassisPsuStorageGbp": 215 }, "benchmarks": { "qwen35bMoeDecode": 72, "qwen35bMoePrefill": 1750, "llama70bDenseDecode": 12, "llama70bDensePrefill": 490, "moe120bDecode": 30, "moe120bPrefill": 620 } }, { "id": "dual-5060ti-ddr4-diy", "name": "Dual RTX 5060 Ti 16GB (32GB GDDR7) + Ryzen 9 5900X + 128GB DDR4", "shortName": "2x RTX 5060 Ti (32GB) + 128GB DDR4", "category": "consumer-diy", "categoryLabel": "Consumer GPU + DDR4 DIY", "condition": "New", "priceGbp": 2583, "underBudget": true, "totalRamGb": 160, "usableAiMemoryGb": 150, "vramGb": 32, "primaryBandwidthGbs": 896, "secondaryBandwidthGbs": 51, "bandwidthLabel": "2x 448 GB/s GDDR7 (32GB) + 51 GB/s DDR4", "powerWatts": 430, "noiseDb": "33 dBA", "ukSource": "PCPartPicker UK / Scan / Amazon UK", "stockStatus": "All parts in stock (New w/ 3-yr warranty)", "verdictBadge": "Best All-New GPU Build (32GB GDDR7)", "summary": "Pairs two Blackwell RTX 5060 Ti 16GB cards (32GB GDDR7 total, native FP4/FP8 Tensor Cores) with 128GB DDR4 (£814 vs £2,199 for DDR5) on a proven AM4 platform.", "pros": [ "100% brand-new retail components with full UK manufacturer warranties", "32GB fast GDDR7 VRAM holds 30B–35B dense/MoE models & 36k+ KV cache entirely in VRAM", "128GB DDR4 saves £1,385 vs DDR5, funding both Blackwell GPUs" ], "caveats": [ "Models exceeding 32GB VRAM spill into 2-channel DDR4 (51 GB/s), slowing >70B decode", "Best for models <=35B in pure VRAM or sparse MoE with small active expert footprints" ], "costSplit": { "ramGbp": 814, "gpuGbp": 900, "cpuBoardGbp": 499, "chassisPsuStorageGbp": 370 }, "benchmarks": { "qwen35bMoeDecode": 112, "qwen35bMoePrefill": 2600, "llama70bDenseDecode": 8, "llama70bDensePrefill": 380, "moe120bDecode": 16, "moe120bPrefill": 450 } }, { "id": "lenovo-p620-threadripper-3090", "name": "Lenovo ThinkStation P620 (Threadripper PRO 3955WX) + RTX 3090 + 128GB 8-Ch", "shortName": "Lenovo P620 TR PRO + RTX 3090", "category": "epyc-server", "categoryLabel": "Enterprise Hybrid MoE DIY", "condition": "Refurbished / Used", "priceGbp": 2840, "underBudget": true, "totalRamGb": 152, "usableAiMemoryGb": 144, "vramGb": 24, "primaryBandwidthGbs": 936, "secondaryBandwidthGbs": 205, "bandwidthLabel": "936 GB/s VRAM + 205 GB/s 8-Ch RAM", "powerWatts": 500, "noiseDb": "35 dBA", "ukSource": "eBay UK + CeX UK", "stockStatus": "In Stock (Refurbished UK Workstations)", "verdictBadge": "Best High-Clock 8-Channel Workstation", "summary": "Single-socket 8-channel DDR4-3200 workstation with 4.3 GHz Zen 2 Threadripper PRO clocks and 128 PCIe 4.0 lanes, paired with a 24GB RTX 3090.", "pros": [ "Single NUMA domain with 8-channel DDR4-3200 (~205 GB/s) and high 4.3 GHz boost clocks", "1000W standard ATX-pinout Platinum PSU included in Lenovo P620 chassis", "Zero motherboard assembly required — slot in 8 RDIMMs and the RTX 3090" ], "caveats": [ "CPU is Lenovo PSB-locked (stays paired with the P620 motherboard)" ], "costSplit": { "ramGbp": 460, "gpuGbp": 975, "cpuBoardGbp": 1250, "chassisPsuStorageGbp": 155 }, "benchmarks": { "qwen35bMoeDecode": 76, "qwen35bMoePrefill": 1820, "llama70bDenseDecode": 12.5, "llama70bDensePrefill": 520, "moe120bDecode": 31, "moe120bPrefill": 650 } }, { "id": "mbp-m3-max-96gb-cex", "name": "Apple MacBook Pro 16\" M3 Max (16C CPU / 40C GPU / 96GB Unified)", "shortName": "MacBook Pro M3 Max 96GB (CeX)", "category": "apple-silicon", "categoryLabel": "Apple Silicon Unified", "condition": "Refurbished / Used", "priceGbp": 2840, "underBudget": true, "totalRamGb": 96, "usableAiMemoryGb": 86, "vramGb": 86, "primaryBandwidthGbs": 400, "secondaryBandwidthGbs": 400, "bandwidthLabel": "400 GB/s Unified LPDDR5", "powerWatts": 65, "noiseDb": "25 dBA", "ukSource": "CeX UK (Grade B In Stock, 5-Yr Warranty)", "stockStatus": "In Stock Now at CeX UK (£2,840)", "verdictBadge": "Best In-Stock Apple Silicon (<£3k)", "summary": "While 96GB (32GB shy of 128GB), it is the only high-memory Apple Silicon unit physically in stock under £3,000 at CeX UK today, adding native hardware bfloat16 and 2x faster prefill than M1 Ultra.", "pros": [ "Physically in stock at CeX UK for £2,840 with 5-Year UK Warranty", "Native hardware bfloat16 + dynamic caching + 2x faster prompt prefill than M1 Ultra", "Includes 16\" Liquid Retina XDR display & battery UPS built in" ], "caveats": [ "96GB total unified memory (~86GB usable VRAM) rather than 128GB", "400 GB/s bandwidth vs 800 GB/s on M1/M2 Ultra" ], "costSplit": { "ramGbp": 0, "gpuGbp": 0, "cpuBoardGbp": 2840, "chassisPsuStorageGbp": 0 }, "benchmarks": { "qwen35bMoeDecode": 44, "qwen35bMoePrefill": 920, "llama70bDenseDecode": 11.5, "llama70bDensePrefill": 390, "moe120bDecode": 19, "moe120bPrefill": 480 } }, { "id": "radeon-r9700-32gb-ddr4", "name": "Gigabyte Radeon AI PRO R9700 32GB + Ryzen 7 5700X + 128GB DDR4", "shortName": "Radeon AI PRO R9700 32GB + 128GB DDR4", "category": "consumer-diy", "categoryLabel": "Consumer GPU + DDR4 DIY", "condition": "New", "priceGbp": 2965, "underBudget": true, "totalRamGb": 160, "usableAiMemoryGb": 150, "vramGb": 32, "primaryBandwidthGbs": 640, "secondaryBandwidthGbs": 51, "bandwidthLabel": "640 GB/s VRAM (32GB) + 51 GB/s DDR4", "powerWatts": 390, "noiseDb": "37 dBA (Blower)", "ukSource": "PCPartPicker UK / Scan UK", "stockStatus": "In Stock (£1,719.99 GPU + £814 RAM)", "verdictBadge": "Contiguous 32GB Single-GPU Build", "summary": "Uses AMD's 32GB RDNA 4 workstation blower card with 128GB DDR4 system RAM to fit 32GB models on a single card without multi-GPU tensor split overhead.", "pros": [ "Single contiguous 32GB VRAM pool (640 GB/s) — no PCIe peer-to-peer split needed", "Blower cooler exhausts heat directly out of the case" ], "caveats": [ "ROCm software ecosystem requires more tuning than CUDA on new architectures", "2x RTX 5060 Ti 16GB (£900) delivers the same 32GB VRAM for £820 less" ], "costSplit": { "ramGbp": 814, "gpuGbp": 1720, "cpuBoardGbp": 245, "chassisPsuStorageGbp": 186 }, "benchmarks": { "qwen35bMoeDecode": 94, "qwen35bMoePrefill": 1950, "llama70bDenseDecode": 7.5, "llama70bDensePrefill": 310, "moe120bDecode": 15, "moe120bPrefill": 390 } }, { "id": "quad-v100-128gb-epyc", "name": "4x NVIDIA Tesla V100 32GB (128GB HBM2) + EPYC 7402P + 128GB DDR4", "shortName": "4x Tesla V100 (128GB HBM2) + EPYC", "category": "epyc-server", "categoryLabel": "Enterprise Multi-GPU DIY", "condition": "Refurbished / Used", "priceGbp": 2950, "underBudget": true, "totalRamGb": 256, "usableAiMemoryGb": 240, "vramGb": 128, "primaryBandwidthGbs": 900, "secondaryBandwidthGbs": 205, "bandwidthLabel": "4x 900 GB/s HBM2 (128GB) + 205 GB/s RAM", "powerWatts": 1150, "noiseDb": "52 dBA (High Airflow Required)", "ukSource": "eBay UK", "stockStatus": "Used Parts Available", "verdictBadge": "128GB Pure HBM2 VRAM (Enthusiast)", "summary": "Four enterprise 32GB HBM2 GPUs on a 7-slot EPYC SP3 board give 128GB of 900 GB/s HBM2 VRAM plus 128GB 8-channel DDR4 RAM under £3,000.", "pros": [ "128GB of pure 900 GB/s HBM2 GPU VRAM + 128GB 8-channel system RAM (256GB total)", "Runs 70B Q8 and 120B Q4 models 100% inside HBM2 VRAM at 28–38 tok/s" ], "caveats": [ "Volta architecture (SM70) lacks native FlashAttention-2, bfloat16, and INT4/FP8 Tensor Cores", "1,150W peak power draw and requires 3D-printed fan shrouds for passive server cards" ], "costSplit": { "ramGbp": 480, "gpuGbp": 1650, "cpuBoardGbp": 520, "chassisPsuStorageGbp": 300 }, "benchmarks": { "qwen35bMoeDecode": 68, "qwen35bMoePrefill": 1100, "llama70bDenseDecode": 26, "llama70bDensePrefill": 460, "moe120bDecode": 34, "moe120bPrefill": 510 } }, { "id": "ddr5-trap-diy", "name": "Consumer AM5 Ryzen 5 7600 + 128GB DDR5-5600 + Intel Arc B580 12GB", "shortName": "Consumer AM5 + 128GB DDR5 (Trap)", "category": "consumer-diy", "categoryLabel": "Consumer GPU + DDR5 DIY", "condition": "New", "priceGbp": 2953, "underBudget": true, "totalRamGb": 140, "usableAiMemoryGb": 128, "vramGb": 12, "primaryBandwidthGbs": 456, "secondaryBandwidthGbs": 75, "bandwidthLabel": "456 GB/s (12GB) + 75 GB/s 2-Ch DDR5", "powerWatts": 280, "noiseDb": "31 dBA", "ukSource": "PCPartPicker UK", "stockStatus": "In Stock", "verdictBadge": "Avoid — The 2026 DDR5 Trap", "summary": "Illustrates why standard consumer DDR5 DIY fails in Oct 2026: 128GB of 2-channel DDR5 costs £2,199 alone, leaving only £345 for a 12GB GPU and bottlenecking large models to 75 GB/s.", "pros": [ "Modern AM5 socket" ], "caveats": [ "£2,199 spent on 128GB DDR5 that only delivers ~75 GB/s dual-channel bandwidth", "Only 12GB GPU VRAM; 70B models crawl at ~4.5 tok/s over 2-channel DDR5" ], "costSplit": { "ramGbp": 2199, "gpuGbp": 345, "cpuBoardGbp": 269, "chassisPsuStorageGbp": 140 }, "benchmarks": { "qwen35bMoeDecode": 29, "qwen35bMoePrefill": 520, "llama70bDenseDecode": 4.5, "llama70bDensePrefill": 110, "moe120bDecode": 9.5, "moe120bPrefill": 160 } }, { "id": "framework-desktop-128", "name": "Framework Desktop DIY (Ryzen AI Max+ 395, 128GB LPDDR5X-8000)", "shortName": "Framework Desktop 128GB", "category": "strix-halo", "categoryLabel": "Unified Mini PC (Strix Halo)", "condition": "New", "priceGbp": 3538, "underBudget": false, "totalRamGb": 128, "usableAiMemoryGb": 120, "vramGb": 120, "primaryBandwidthGbs": 256, "secondaryBandwidthGbs": 256, "bandwidthLabel": "256 GB/s 256-bit LPDDR5X-8000", "powerWatts": 120, "noiseDb": "32 dBA (120mm Noctua/CoolerMaster)", "ukSource": "Framework UK Store (£3,149 Board / £3,538 DIY)", "stockStatus": "In Stock (Framework UK)", "verdictBadge": "Over £3k Cap (Board Alone £3,149)", "summary": "Open-source modular 4.5L Mini-ITX Strix Halo workstation. After Framework's May 2026 +11% UK DRAM price hike, even the bare mainboard (£3,149) exceeds the £3,000 cap.", "pros": [ "Standard Mini-ITX mainboard + 120mm fan + PCIe x4 slot for 50GbE RoCE clustering", "Official Fedora support and reproducible Framework LLM benchmark suite" ], "caveats": [ "Exceeds £3,000 budget (£3,149 mainboard only; £3,538 with case/PSU/SSD)" ], "costSplit": { "ramGbp": 0, "gpuGbp": 0, "cpuBoardGbp": 3149, "chassisPsuStorageGbp": 389 }, "benchmarks": { "qwen35bMoeDecode": 82, "qwen35bMoePrefill": 755, "llama70bDenseDecode": 15, "llama70bDensePrefill": 290, "moe120bDecode": 25, "moe120bPrefill": 380 } }, { "id": "mac-studio-m2-ultra-128", "name": "Apple Mac Studio M2 Ultra (24C CPU / 60C GPU / 128GB Unified)", "shortName": "Mac Studio M2 Ultra 128GB (Used)", "category": "apple-silicon", "categoryLabel": "Apple Silicon Unified", "condition": "Refurbished / Used", "priceGbp": 3350, "underBudget": false, "totalRamGb": 128, "usableAiMemoryGb": 114, "vramGb": 114, "primaryBandwidthGbs": 800, "secondaryBandwidthGbs": 800, "bandwidthLabel": "800 GB/s Unified LPDDR5", "powerWatts": 95, "noiseDb": "22 dBA", "ukSource": "CeX UK (£3,350 Grade B) / Hoxton Macs (£4,895)", "stockStatus": "Out of Stock / £350 Over Budget", "verdictBadge": "£350 Over Cap at CeX (£3,350)", "summary": "Adds 20–30% faster prompt prefill and native bfloat16 over M1 Ultra while keeping 800 GB/s bandwidth, but starts at £3,350 used in the UK.", "pros": [ "800 GB/s unified memory bandwidth + native hardware bfloat16" ], "caveats": [ "Starts at £3,350 at CeX UK (£350 above the £3,000 ceiling)" ], "costSplit": { "ramGbp": 0, "gpuGbp": 0, "cpuBoardGbp": 3350, "chassisPsuStorageGbp": 0 }, "benchmarks": { "qwen35bMoeDecode": 62, "qwen35bMoePrefill": 610, "llama70bDenseDecode": 20, "llama70bDensePrefill": 260, "moe120bDecode": 32, "moe120bPrefill": 340 } }, { "id": "dgx-spark-gb10", "name": "NVIDIA DGX Spark (GB10 Grace Blackwell Superchip, 128GB Unified)", "shortName": "NVIDIA DGX Spark 128GB", "category": "nvidia-gb10", "categoryLabel": "NVIDIA Grace Blackwell Mini", "condition": "New", "priceGbp": 4780, "underBudget": false, "totalRamGb": 128, "usableAiMemoryGb": 118, "vramGb": 118, "primaryBandwidthGbs": 273, "secondaryBandwidthGbs": 273, "bandwidthLabel": "273 GB/s LPDDR5X + NVFP4 Blackwell", "powerWatts": 170, "noiseDb": "36 dBA", "ukSource": "PCSpecialist UK / Scan 3XS", "stockStatus": "Available (£4,780 inc VAT)", "verdictBadge": "£1,780 Over Cap After 2026 Price Hike", "summary": "1 PFLOP FP4 Blackwell tensor compute with 128GB unified memory and ConnectX networking, but 2026 DRAM price hikes pushed UK pricing to £4,780.", "pros": [ "Full NVIDIA CUDA stack + native NVFP4 Tensor Cores (102 tok/s on Qwen3.5-35B-A3B)", "Fast prompt prefill (1,800+ pp/s) despite 273 GB/s memory bus" ], "caveats": [ "£4,780 UK retail price is £1,780 over the £3,000 budget", "273 GB/s bus still bottlenecks dense 70B Q4 decode to ~15–17 tok/s" ], "costSplit": { "ramGbp": 0, "gpuGbp": 0, "cpuBoardGbp": 4780, "chassisPsuStorageGbp": 0 }, "benchmarks": { "qwen35bMoeDecode": 102, "qwen35bMoePrefill": 2100, "llama70bDenseDecode": 16, "llama70bDensePrefill": 820, "moe120bDecode": 28, "moe120bPrefill": 950 } }, { "id": "mac-studio-m5-max-128", "name": "Apple Mac Studio M5 Max (18C CPU / 40C GPU / 128GB Unified, 1TB)", "shortName": "Mac Studio M5 Max 128GB (New)", "category": "apple-silicon", "categoryLabel": "Apple Silicon Unified", "condition": "New", "priceGbp": 5399, "underBudget": false, "totalRamGb": 128, "usableAiMemoryGb": 114, "vramGb": 114, "primaryBandwidthGbs": 546, "secondaryBandwidthGbs": 546, "bandwidthLabel": "546 GB/s Unified LPDDR5X", "powerWatts": 105, "noiseDb": "22 dBA", "ukSource": "Apple Store UK (apple.com/uk)", "stockStatus": "In Stock (£5,399 New)", "verdictBadge": "£2,399 Over Cap (Apple RAM Tax)", "summary": "Apple's late-2026 M5 Max Mac Studio starts at £2,499 (36GB), but upgrading to 40-core GPU + 128GB unified memory adds £2,900, bringing the total to £5,399.", "pros": [ "Massive prompt prefill speedup (>2,200 pp/s) via M5 Neural Accelerators & SME", "546 GB/s unified memory in a silent 2.7kg desktop" ], "caveats": [ "£5,399 UK price is nearly double the £3,000 budget ceiling" ], "costSplit": { "ramGbp": 2600, "gpuGbp": 300, "cpuBoardGbp": 2499, "chassisPsuStorageGbp": 0 }, "benchmarks": { "qwen35bMoeDecode": 48, "qwen35bMoePrefill": 2350, "llama70bDenseDecode": 14.5, "llama70bDensePrefill": 780, "moe120bDecode": 23, "moe120bPrefill": 910 } } ], "blueprints": [ { "id": "blueprint-epyc-milan", "title": "Blueprint A — The EPYC Milan 8-Channel + RTX 3090 Hybrid MoE Rig", "subtitle": "152GB Combined Memory (24GB GDDR6X @ 936 GB/s + 128GB 8-Ch DDR4-3200 @ 205 GB/s)", "totalGbp": 2834, "targetUse": "Best for running massive Sparse MoE models (DeepSeek V3/R1/V4, Qwen3.5-122B-A10B, gpt-oss-120B) with fast CUDA prompt prefill and quiet home-office acoustics.", "architectureNote": "KTransformers and ik_llama.cpp pin shared attention layers + KV cache onto the 24GB RTX 3090 while streaming active routed experts from 8-channel DDR4-3200 ECC RDIMMs at 205 GB/s.", "parts": [ { "component": "CPU", "item": "AMD EPYC 7443 Milan (24C/48T, 2.85–4.0 GHz, Zen 3, 128 PCIe 4.0 lanes, Unlocked)", "priceGbp": 234, "source": "eBay UK Verified Server Pull", "notes": "Confirm 'Unlocked / Not Dell or Lenovo PSB locked' in listing before purchase." }, { "component": "Motherboard", "item": "Gigabyte MZ32-AR0 Rev 3.0 (SP3, E-ATX, 16x DDR4 RDIMM, 7x PCIe 4.0 slots, BMC IPMI)", "priceGbp": 580, "source": "eBay UK / Server Factory UK", "notes": "Rev 3.0 supports EPYC 7003 Milan out of the box and has 8 remaining DIMM slots for 256GB." }, { "component": "Memory (128GB 8-Ch)", "item": "8x 16GB Samsung / Micron DDR4-3200AA ECC Registered RDIMM (2Rx8, PC4-25600R)", "priceGbp": 480, "source": "Bargain Hardware UK / eBay UK", "notes": "Must populate 8 identical sticks (1 DPC across channels A–H) to unlock full 204.8 GB/s." }, { "component": "GPU (24GB VRAM)", "item": "NVIDIA GeForce RTX 3090 24GB GDDR6X (Founders Edition / Gigabyte OC / Palit)", "priceGbp": 975, "source": "CeX UK (5-Yr Warranty) / eBay UK", "notes": "936 GB/s bandwidth handles prompt prefill, KV cache, and shared attention tensors." }, { "component": "CPU Cooler", "item": "Noctua NH-U14S TR4-SP3 (140mm Quiet Tower Cooler for Socket SP3)", "priceGbp": 85, "source": "Scan UK / Amazon UK", "notes": "Replaces screaming 2U server fans with <34 dBA desktop acoustics." }, { "component": "Chassis", "item": "Phanteks Enthoo Pro 2 Server Edition (Full Tower E-ATX/SSI-EEB)", "priceGbp": 140, "source": "Overclockers UK / Scan UK", "notes": "Fits SSI-EEB/E-ATX server boards and up to three dual-slot GPUs." }, { "component": "Power Supply", "item": "Corsair RM1000x / MSI MEG Ai1300P 1200W 80+ Gold/Platinum ATX 3.1 Modular", "priceGbp": 180, "source": "Scan UK / Amazon UK", "notes": "Includes dual 8-pin EPS12V cables required by SP3 server motherboards." }, { "component": "Storage & Fans", "item": "2TB Crucial P3 Plus PCIe 4.0 NVMe (£130) + 5-Pack Arctic P14 PWM PST Fans (£30)", "priceGbp": 160, "source": "Amazon UK / Crucial UK", "notes": "Direct fan airflow over the RDIMM banks and BMC heatsink." } ], "tuningTips": [ "In Gigabyte BIOS: set NUMA Nodes Per Socket (NPS) = 1 (or NPS4 with 8-channel interleaving enabled) and PCIe Speed = Gen4.", "Run ik_llama.cpp or llama.cpp with `-ngl 99 -ot \".*ffn_.*_exps.*=CPU\" -fa 1 --threads 24` to keep attention + KV cache on the RTX 3090 and routed MoE experts in 8-channel RAM." ] }, { "id": "blueprint-strix-halo", "title": "Blueprint B — Turnkey AMD Strix Halo 128GB Unified Mini Workstation", "subtitle": "128GB LPDDR5X-8000 (120GB Usable GPU VRAM via Linux TTM @ 256 GB/s, 120W Wall Power)", "totalGbp": 1950, "targetUse": "Best for developers wanting a tiny, low-power, brand-new Linux box that loads 120B MoE or 70B Q4 models entirely into unified memory with zero PCIe riser/server assembly.", "architectureNote": "AMD Ryzen AI Max+ 395 integrates 16 Zen 5 cores and a 40-CU RDNA 3.5 GPU (`gfx1151`) on a 256-bit LPDDR5X-8000 bus (~256 GB/s). Linux 6.16+ TTM parameters allow 120GB to be mapped directly as GPU VRAM.", "parts": [ { "component": "System (Option 1 — Direct Hub)", "item": "GMKtec EVO-X2 or Bosgame M5 (Ryzen AI Max+ 395, 128GB LPDDR5X-8000, 2TB NVMe)", "priceGbp": 1950, "source": "Direct Vendor EU/CN Warehouse", "notes": "Saves ~£960 vs UK retail by avoiding distributor DRAM markups." }, { "component": "System (Option 2 — UK Prime)", "item": "Bosgame M5 128GB (£3,099.99 minus 6% voucher = £2,914) or GMKtec EVO-X2 (£2,999)", "priceGbp": 2914, "source": "Amazon UK (ASIN B0H6XG8VTZ / B0F6X332N6)", "notes": "Next-day UK delivery and Amazon UK domestic return protection under £3,000." } ], "tuningTips": [ "In BIOS: set UMA Frame Buffer Size (GART reservation) to 512MB so the OS does not lock static RAM away from dynamic TTM.", "Create `/etc/modprobe.d/amdgpu_llm.conf` with `options ttm pages_limit=31457280` and `options ttm page_pool_size=15728640` to unlock 120GB usable VRAM.", "Use Linux kernel 6.16+ with Mesa RADV Vulkan (`llama-server --device Vulkan0 -fa 1 --mmap 0`) for stable 128k-context inference." ] }, { "id": "blueprint-mac-studio-m1u", "title": "Blueprint C — Used Mac Studio M1 Ultra 128GB (800 GB/s Bandwidth Champion)", "subtitle": "128GB Unified Memory (~114GB Wired GPU Limit @ 800 GB/s, <90W Power, 5-Year CeX Warranty)", "totalGbp": 2450, "targetUse": "Best for running dense 70B–100B models (Llama 3.3 70B Q4/Q6, Qwen 72B, Command-R+ 104B) at 18–20 tok/s with silent desktop operation.", "architectureNote": "M1 Ultra joins two M1 Max dies over UltraFusion to deliver an 800 GB/s unified memory bus—more than 3x the memory bandwidth of Strix Halo (256 GB/s) and DGX Spark (273 GB/s).", "parts": [ { "component": "Workstation (CeX UK)", "item": "Apple Mac Studio 13,2 / M1 Ultra (20-CPU 48-GPU or 64-GPU) / 128GB RAM / 1TB–2TB SSD (Grade B/C)", "priceGbp": 2450, "source": "CeX UK (uk.webuy.com) — £2,450 to £2,925", "notes": "Includes mandatory 5-Year CeX UK Hardware Warranty." }, { "component": "Workstation (Hoxton Macs)", "item": "Refurbished Apple Mac Studio M1 Ultra 128GB / 1TB–2TB SSD (Fair / Good / Excellent)", "priceGbp": 2095, "source": "Hoxton Macs UK (hoxtonmacs.co.uk) — £1,895 to £2,495", "notes": "Includes 1-Year Hardware Warranty; set restock alert as units sell out within hours." } ], "tuningTips": [ "Raise macOS Metal wired memory ceiling to 114GB: `sudo sysctl iogpu.wired_limit_mb=116736`.", "Prefer MLX (`mlx_lm`) 4-bit/6-bit/8-bit quantized models or FP16 GGUFs; avoid raw unquantized `bfloat16` weights on M1 Ultra.", "Enable prompt caching (`--prompt-cache-file`) in llama.cpp / MLX to avoid re-prefilling long system prompts." ] }, { "id": "blueprint-dual-5060ti", "title": "Blueprint D — All-New Dual RTX 5060 Ti 16GB (32GB GDDR7) + 128GB DDR4 Build", "subtitle": "32GB GDDR7 VRAM (@ 448 GB/s per GPU) + 128GB Dual-Channel DDR4-3200 (@ 51 GB/s)", "totalGbp": 2583, "targetUse": "Best for buyers who strictly require 100% brand-new UK retail parts with 3-year warranties, fast CUDA FP4/FP8 coding agents (27B–35B in pure VRAM), and 128GB system RAM.", "architectureNote": "Uses AM4 + DDR4 (£814 for 128GB) to dodge the £2,199 DDR5 tax, freeing budget for two 16GB GDDR7 Blackwell GPUs (32GB total VRAM).", "parts": [ { "component": "GPUs (32GB GDDR7 Total)", "item": "2x Asus Dual OC GeForce RTX 5060 Ti 16GB GDDR7 (£449.99 each)", "priceGbp": 900, "source": "PCPartPicker UK / Scan / Overclockers UK", "notes": "32GB combined GDDR7 VRAM with Blackwell FP4/FP8 Tensor Cores." }, { "component": "Memory (128GB DDR4)", "item": "2x 64GB Kits (4x32GB) DDR4-3200 CL16 Desktop UDIMM", "priceGbp": 814, "source": "Amazon UK (via ramprices.uk)", "notes": "Saves £1,385 compared to 128GB DDR5 (£2,199)." }, { "component": "CPU & Cooler", "item": "AMD Ryzen 9 5900X (12C/24T) (£276) + Thermalright Peerless Assassin 120 SE (£35)", "priceGbp": 311, "source": "Amazon UK / Scan UK", "notes": "12 cores for fast tokenization, compilation, and hybrid CPU offload." }, { "component": "Motherboard", "item": "MSI MPG B550 GAMING PLUS or Asus ProArt B550-CREATOR (Dual PCIe x8/x8 bifurcation)", "priceGbp": 188, "source": "Scan UK / Amazon UK", "notes": "ProArt B550-CREATOR supports clean x8/x8 PCIe 4.0 across two dual-slot GPUs." }, { "component": "Storage, PSU & Case", "item": "2TB Crucial P3 Plus NVMe (£130) + Corsair RM850e ATX 3.1 PSU (£120) + Lian Li Lancool 216 (£120)", "priceGbp": 370, "source": "Scan UK / Amazon UK", "notes": "430W peak load runs cool and quiet on an 850W Gold PSU." } ], "tuningTips": [ "Run 27B–35B models (Qwen3.8-27B Q6_K or Qwen3.5-35B-A3B Q5_K_M) 100% inside the 32GB GDDR7 VRAM with `--split-mode layer` or vLLM tensor parallel (`-tp 2`) for 100+ tok/s." ] } ], "ramKits": [ { "tier": "Server DDR4 ECC RDIMM (Best Value)", "kitName": "8x 16GB Samsung / Micron DDR4-3200AA Registered ECC (PC4-25600R)", "capacityGb": 128, "config": "8 x 16GB (8-Channel)", "speed": "DDR4-3200 CL22", "priceGbp": 480, "gbpPerGb": 3.75, "bandwidthGbs": 204.8, "retailer": "Bargain Hardware UK / eBay UK" }, { "tier": "Server DDR4 ECC RDIMM (256GB Upgrade)", "kitName": "8x 32GB Hynix / Samsung DDR4-3200AA Registered ECC", "capacityGb": 256, "config": "8 x 32GB (8-Channel)", "speed": "DDR4-3200 CL22", "priceGbp": 1020, "gbpPerGb": 3.98, "bandwidthGbs": 204.8, "retailer": "Bargain Hardware UK / eBay UK" }, { "tier": "Desktop DDR4 UDIMM", "kitName": "2x Gigastone Game PRO 64GB / NVTEK 128GB DDR4-3200 UDIMM", "capacityGb": 128, "config": "4 x 32GB (2-Channel)", "speed": "DDR4-3200 CL16", "priceGbp": 814, "gbpPerGb": 6.36, "bandwidthGbs": 51.2, "retailer": "Amazon UK (ramprices.uk)" }, { "tier": "Laptop / Mini PC DDR5 SO-DIMM", "kitName": "Crucial 128GB Kit (2x64GB) DDR5-5600 SO-DIMM (CT2K64G56C46S5)", "capacityGb": 128, "config": "2 x 64GB (2-Channel)", "speed": "DDR5-5600 CL46", "priceGbp": 1990, "gbpPerGb": 15.55, "bandwidthGbs": 89.6, "retailer": "Amazon UK" }, { "tier": "Desktop DDR5 UDIMM (3x Surge)", "kitName": "Crucial Pro 128GB Kit (2x64GB) DDR5-5600 UDIMM (CP2K64G56C46U5)", "capacityGb": 128, "config": "2 x 64GB (2-Channel)", "speed": "DDR5-5600 CL46", "priceGbp": 2199, "gbpPerGb": 17.18, "bandwidthGbs": 89.6, "retailer": "Amazon UK / PCPartPicker UK" }, { "tier": "Desktop DDR5 UDIMM (6000MT/s)", "kitName": "Corsair Vengeance RGB 128GB (2x64GB) DDR5-6000 CL40", "capacityGb": 128, "config": "2 x 64GB (2-Channel)", "speed": "DDR5-6000 CL40", "priceGbp": 2214, "gbpPerGb": 17.3, "bandwidthGbs": 96, "retailer": "Scan UK / Amazon UK" } ], "gpus": [ { "model": "NVIDIA Tesla V100 PCIe 32GB", "vramGb": 32, "memType": "HBM2", "bandwidthGbs": 900, "priceGbp": 415, "gbpPerGbVram": 12.97, "condition": "Used (Server Pull)", "retailer": "eBay UK", "notes": "Cheap 32GB HBM2 (4x = 128GB for £1,660), but older Volta SM70 (no bf16/FA2) & needs fan shroud." }, { "model": "Asus Dual OC GeForce RTX 5060 Ti 16GB", "vramGb": 16, "memType": "GDDR7", "bandwidthGbs": 448, "priceGbp": 450, "gbpPerGbVram": 28.13, "condition": "New", "retailer": "Scan UK / Overclockers UK", "notes": "Best new VRAM/£ card in Oct 2026; 2x gives 32GB GDDR7 for £900, 4x gives 64GB for £1,800." }, { "model": "Sparkle Intel Arc Pro B60 24GB", "vramGb": 24, "memType": "GDDR6", "bandwidthGbs": 456, "priceGbp": 900, "gbpPerGbVram": 37.5, "condition": "New", "retailer": "PCPartPicker UK", "notes": "Cheapest new single 24GB card, though SYCL/Vulkan software lags behind CUDA." }, { "model": "NVIDIA GeForce RTX 3090 24GB", "vramGb": 24, "memType": "GDDR6X", "bandwidthGbs": 936, "priceGbp": 975, "gbpPerGbVram": 40.63, "condition": "Used (CeX 5-Yr Warranty / eBay)", "retailer": "CeX UK (£845–£1,145) / eBay UK", "notes": "Still the undisputed king for Hybrid MoE prefill + 936 GB/s CUDA bandwidth." }, { "model": "Sapphire NITRO+ Radeon RX 7900 XTX 24GB", "vramGb": 24, "memType": "GDDR6", "bandwidthGbs": 960, "priceGbp": 1699, "gbpPerGbVram": 70.79, "condition": "New", "retailer": "Amazon UK / Overclockers UK", "notes": "960 GB/s bandwidth, though new UK pricing in late 2026 (£1,699) makes used 3090 or R9700 better value." }, { "model": "Gigabyte Radeon AI PRO R9700 32GB", "vramGb": 32, "memType": "GDDR6", "bandwidthGbs": 640, "priceGbp": 1720, "gbpPerGbVram": 53.75, "condition": "New", "retailer": "Scan UK / PCPartPicker UK", "notes": "Contiguous 32GB VRAM in a single blower slot for workstation multi-GPU builds." }, { "model": "NVIDIA GeForce RTX 5090 32GB", "vramGb": 32, "memType": "GDDR7", "bandwidthGbs": 1792, "priceGbp": 6200, "gbpPerGbVram": 193.75, "condition": "New", "retailer": "PCPartPicker UK / Marketplace", "notes": "1,792 GB/s monster, but Oct 2026 UK street prices (£6,199+) are >2x the entire £3,000 budget." } ] }