接受arm解决方案的话可以Nvidia GB10 理论上统一内存架构128G是能支撑的。当然还有一些奇葩解决方案比如用85xx/84xx的AVX512完成推理 也是可行的,这模型小 内存宽带压力不会特别大,推理本身不是特别吃算力
Supermicro AS-3015MR-H10TNR MicroCloud 3U Barebone Single AMD EPYC 4004 Series Processor
• Cloud Computing
• Web/Collocation Services
• Web Cache, CDN, Video Streaming
• Social Networking, Downloads
• Corporate - WINS, DNS, Print, Login
Barebone 3U Rackmount SuperServer, Per Node: Support Single Socket AM5 (LGA-1718), Supporting AMD EPYC 4004 Series Processor (Up to 16C/32T, Up to 32MB Cache). Max GPU Count: Up to 1 single-width GPU, Supported GPU: NVIDIA PCIe: T4, L4, A2. Up to 192GB in 4 DIMM Slots, UDIMM, Chipset: System on Chip, On-Board: NVMe (Native). Drive Bays: 2 Internal Fixed 2.5in PCIe 5.0 NVMe Drive Bays, 2 Internal Fixed 2.5in SATA Drive Bays, 1 M.2 NVMe slot (M-key 2280/22110), Network Connectivity: No NIC Option Supported, 2000W Power Supply - (CSE-MC301-000NFP + H13SRE-F) - AS-3015MR-H10TNR
超微10节点EPYC 4004/Ryzen 7000/9000系列准系统有兴趣可以私聊 报价7700USD W/o Shipping,原厂发
• Cloud Computing
• Web/Collocation Services
• Web Cache, CDN, Video Streaming
• Social Networking, Downloads
• Corporate - WINS, DNS, Print, Login
Barebone 3U Rackmount SuperServer, Per Node: Support Single Socket AM5 (LGA-1718), Supporting AMD EPYC 4004 Series Processor (Up to 16C/32T, Up to 32MB Cache). Max GPU Count: Up to 1 single-width GPU, Supported GPU: NVIDIA PCIe: T4, L4, A2. Up to 192GB in 4 DIMM Slots, UDIMM, Chipset: System on Chip, On-Board: NVMe (Native). Drive Bays: 2 Internal Fixed 2.5in PCIe 5.0 NVMe Drive Bays, 2 Internal Fixed 2.5in SATA Drive Bays, 1 M.2 NVMe slot (M-key 2280/22110), Network Connectivity: No NIC Option Supported, 2000W Power Supply - (CSE-MC301-000NFP + H13SRE-F) - AS-3015MR-H10TNR
超微10节点EPYC 4004/Ryzen 7000/9000系列准系统有兴趣可以私聊 报价7700USD W/o Shipping,原厂发