{"product_id":"kentino-ai-punkstation-pro","title":"Home AI Server — Punkstation Pro - 88 GB VRAM - EPYC","description":"\u003cstyle\u003e\n  .kp{--acc:#334fb4;max-width:960px;margin:0 auto;line-height:1.6}\n  .kp h2{background:var(--acc);color:#fff !important;margin:2rem 0 0;padding:.7rem 1rem;border-radius:8px 8px 0 0;font-size:1.15rem}\n  .kp .sec{border:1px solid #e4e7ef;border-top:none;border-radius:0 0 8px 8px;padding:1.1rem 1.2rem;margin-bottom:.6rem}\n  .kp .lead{font-size:1.08rem;margin:0 0 1rem}\n  .kp .vram{display:flex;gap:1rem;flex-wrap:wrap;margin:1rem 0}\n  .kp .vcard{flex:1;min-width:210px;border:1px solid #e4e7ef;border-radius:8px;padding:1rem;text-align:center}\n  .kp .vcard .big{font-size:2rem;font-weight:700;color:var(--acc);display:block}\n  .kp table{width:100%;border-collapse:collapse;margin:.3rem 0}\n  .kp th{background:var(--acc);color:#fff;text-align:left;padding:.55rem .8rem;font-size:.95rem}\n  .kp td{padding:.55rem .8rem;border-bottom:1px solid #eef0f6;vertical-align:top}\n  .kp td:first-child{font-weight:600;width:34%}\n  .kp .note{background:#fff8e6;border:1px solid #f0dca0;border-radius:8px;padding:.9rem 1.1rem;margin:1rem 0;font-size:.95rem}\n  .kp .note b{color:#8a6d00}\n  .kp ul{margin:.4rem 0;padding-left:1.2rem}\n  .kp .tag{display:inline-block;background:#eef1fb;color:var(--acc);border-radius:20px;padding:.2rem .7rem;font-size:.8rem;font-weight:600;margin:.15rem .2rem 0 0}\n\u003c\/style\u003e\n\n\u003cdiv class=\"kp\"\u003e\n\n\u003cp class=\"lead\"\u003eThe Punkstation on a \u003cb\u003eserver-grade platform\u003c\/b\u003e: a consumer \u003cb\u003eRTX 4090\u003c\/b\u003e paired with a\ndriver-unlocked \u003cb\u003eNVIDIA CMP 170HX\u003c\/b\u003e (64 GB HBM2e) — \u003cb\u003e88 GB of aggregated VRAM\u003c\/b\u003e — on a\nsingle-socket \u003cb\u003eAMD EPYC\u003c\/b\u003e board with \u003cb\u003eECC memory, IPMI remote management\u003c\/b\u003e and four PCIe 4.0 ×16\nslots to grow into. Not the budget desktop — a robust, expandable machine, \u003cb\u003ebuilt and configured for\nmaximum performance\u003c\/b\u003e. Available as a quiet \u003cb\u003etower\u003c\/b\u003e or a rack-mount \u003cb\u003eserver\u003c\/b\u003e.\u003c\/p\u003e\n\n\u003cdiv class=\"vram\"\u003e\n  \u003cdiv class=\"vcard\"\u003e\n\u003cspan class=\"big\"\u003e88 GB\u003c\/span\u003eaggregated VRAM (24 + 64)\u003c\/div\u003e\n  \u003cdiv class=\"vcard\"\u003e\n\u003cspan class=\"big\"\u003eEPYC\u003c\/span\u003eserver platform · ECC · IPMI\u003c\/div\u003e\n  \u003cdiv class=\"vcard\"\u003e\n\u003cspan class=\"big\"\u003e2 → 4\u003c\/span\u003eGPU slots (room to expand)\u003c\/div\u003e\n\u003c\/div\u003e\n\n\u003ch2\u003eWhat it's for\u003c\/h2\u003e\n\u003cdiv class=\"sec\"\u003e\n\u003cp\u003eSame aggregated-VRAM strengths as the Punkstation, on hardware that lasts and scales. VRAM is\n\u003cb\u003eaggregated, not pooled\u003c\/b\u003e — the two cards are independent memories, so it's at its best on video and\nimage generation, large mixture-of-experts models that fit one card, and mixture-of-agents \/ multi-model\npipelines. The RTX 4090 drives the desktop and fast image generation; the 170HX's 64 GB carries the big\nmodel. The EPYC platform adds \u003cb\u003eECC memory, out-of-band IPMI management, more PCIe lanes and NVMe\u003c\/b\u003e,\nand a clear upgrade path to the 4× 170HX build.\u003c\/p\u003e\n\u003cp\u003e\u003cspan class=\"tag\"\u003eImage generation\u003c\/span\u003e\u003cspan class=\"tag\"\u003eMixture-of-agents\u003c\/span\u003e\u003cspan class=\"tag\"\u003eVideo generation\u003c\/span\u003e\u003cspan class=\"tag\"\u003eMoE reasoning\u003c\/span\u003e\u003c\/p\u003e\n\u003cul\u003e\n  \u003cli\u003e\n\u003cb\u003eValidated:\u003c\/b\u003e Qwen3 35B-A3B (mixture-of-experts, W8A8, compiled) runs entirely in the 170HX's\n  64 GB with no offload at \u003cb\u003e~140 tokens\/s single-stream\u003c\/b\u003e (and ~3,270 tok\/s across a 48-request batch) —\n  faster than a 7B dense model while reasoning coherently.\u003c\/li\u003e\n  \u003cli\u003e\n\u003cb\u003eVideo \/ image:\u003c\/b\u003e generation models can't be split across cards — each fits one GPU. The 170HX's\n  64 GB hosts large diffusion and video models at high resolution and batch; the 4090 gives fast SDXL\/Flux-class\n  image generation. Run both in parallel for concurrent jobs.\u003c\/li\u003e\n\u003c\/ul\u003e\n\u003c\/div\u003e\n\n\u003ch2\u003eSpecification\u003c\/h2\u003e\n\u003cdiv class=\"sec\"\u003e\n\u003ctable\u003e\n  \u003ctr\u003e\n\u003cth\u003eComponent\u003c\/th\u003e\n\u003cth\u003eDetail\u003c\/th\u003e\n\u003c\/tr\u003e\n  \u003ctr\u003e\n\u003ctd\u003eAccelerator 1\u003c\/td\u003e\n\u003ctd\u003eNVIDIA GeForce RTX 4090 · 24 GB GDDR6X · drives display + fast single-card work\u003c\/td\u003e\n\u003c\/tr\u003e\n  \u003ctr\u003e\n\u003ctd\u003eAccelerator 2\u003c\/td\u003e\n\u003ctd\u003eNVIDIA CMP 170HX · 64 GB HBM2e (modified, driver-unlocked, compute-only)\u003c\/td\u003e\n\u003c\/tr\u003e\n  \u003ctr\u003e\n\u003ctd\u003eTotal VRAM\u003c\/td\u003e\n\u003ctd\u003e88 GB aggregated (24 + 64), both fully accessible\u003c\/td\u003e\n\u003c\/tr\u003e\n  \u003ctr\u003e\n\u003ctd\u003eCPU\u003c\/td\u003e\n\u003ctd\u003eAMD EPYC (Rome \/ Milan, SP3) — model configured to order\u003c\/td\u003e\n\u003c\/tr\u003e\n  \u003ctr\u003e\n\u003ctd\u003eMainboard\u003c\/td\u003e\n\u003ctd\u003eSingle-socket EPYC server board · 8-channel DDR4 ECC · 4× PCIe 4.0 ×16 · onboard graphics + IPMI\/BMC remote management\u003c\/td\u003e\n\u003c\/tr\u003e\n  \u003ctr\u003e\n\u003ctd\u003eMemory\u003c\/td\u003e\n\u003ctd\u003eDDR4 ECC, octa-channel — configurable (128 GB typical)\u003c\/td\u003e\n\u003c\/tr\u003e\n  \u003ctr\u003e\n\u003ctd\u003eStorage\u003c\/td\u003e\n\u003ctd\u003eM.2 NVMe (up to 3× PCIe 4.0 ×4)\u003c\/td\u003e\n\u003c\/tr\u003e\n  \u003ctr\u003e\n\u003ctd\u003eNetworking\u003c\/td\u003e\n\u003ctd\u003eDual 2.5 GbE + dedicated IPMI remote-management port\u003c\/td\u003e\n\u003c\/tr\u003e\n  \u003ctr\u003e\n\u003ctd\u003ePower\u003c\/td\u003e\n\u003ctd\u003eUprated PSU sized to the build\u003c\/td\u003e\n\u003c\/tr\u003e\n  \u003ctr\u003e\n\u003ctd\u003eChassis\u003c\/td\u003e\n\u003ctd\u003eQuiet desktop tower — or rack-mount server (choose above)\u003c\/td\u003e\n\u003c\/tr\u003e\n  \u003ctr\u003e\n\u003ctd\u003eOperating system\u003c\/td\u003e\n\u003ctd\u003eUbuntu, NVIDIA drivers + CUDA pre-configured, compiled\/cudagraph runtime tuned\u003c\/td\u003e\n\u003c\/tr\u003e\n  \u003ctr\u003e\n\u003ctd\u003eAssembly\u003c\/td\u003e\n\u003ctd\u003eBuilt, burned-in and validated at maximum performance before dispatch\u003c\/td\u003e\n\u003c\/tr\u003e\n\u003c\/table\u003e\n\u003c\/div\u003e\n\n\u003ch2\u003ePlease read before ordering\u003c\/h2\u003e\n\u003cdiv class=\"sec\"\u003e\n\u003cdiv class=\"note\"\u003e\n\u003cb\u003e⚠ Driver-unlocked card.\u003c\/b\u003e The CMP 170HX is a compute card unlocked to full speed\n\u003cb\u003eoutside NVIDIA's specification\u003c\/b\u003e. The configuration is confirmed working and ships validated, but the\n170HX carries \u003cb\u003eno manufacturer warranty\u003c\/b\u003e — Kentino covers the build and the standard components; the\n170HX is supported on a best-effort basis.\u003c\/div\u003e\n\u003cul\u003e\n  \u003cli\u003eThe 170HX is \u003cb\u003ecompute-only\u003c\/b\u003e (no display) — the RTX 4090 (or onboard BMC graphics) drives the console.\u003c\/li\u003e\n  \u003cli\u003eThe 170HX links at \u003cb\u003ePCIe 2.0\u003c\/b\u003e (×4 standard, ×16 mod available) — limits host↔device transfer, not resident-model inference.\u003c\/li\u003e\n\u003c\/ul\u003e\n\u003c\/div\u003e\n\n\u003ch2\u003eOrdering\u003c\/h2\u003e\n\u003cdiv class=\"sec\"\u003e\n\u003cp\u003e\u003cb\u003ePrice in preparation\u003c\/b\u003e — request a quote. Choose \u003cb\u003eTower\u003c\/b\u003e or \u003cb\u003eServer (rack)\u003c\/b\u003e above.\nLead time \u003cb\u003e10–28 days\u003c\/b\u003e. Prices are \u003cb\u003eEUR ex-VAT\u003c\/b\u003e. Bespoke CPU \/ memory \/ storage configured to order.\u003c\/p\u003e\n\u003c\/div\u003e\n\n\u003c\/div\u003e\n","brand":"Kentino s.r.o.","offers":[{"title":"Tower","offer_id":53761591640392,"sku":null,"price":4599.0,"currency_code":"EUR","in_stock":true},{"title":"Server (rack)","offer_id":53761591673160,"sku":null,"price":4899.0,"currency_code":"EUR","in_stock":true}],"thumbnail_url":"\/\/cdn.shopify.com\/s\/files\/1\/0843\/5479\/3800\/files\/kentino-ai-punkstation-pro_e6bc4cbf-a8fa-4c20-97d7-99cf36c69730.jpg?v=1786626850","url":"https:\/\/kentino.com\/da\/products\/kentino-ai-punkstation-pro","provider":"Kentino","version":"1.0","type":"link"}