GT3 desktop workstation (compact, 3.5L)

Specifications

• Size: 3.5L 212*190*86mm mmSupplier claim
• CPU: Intel 12-14 DesktopSupplier claim
• RAM: Up to 96GB SO-DIMM DDR5*2 GBSupplier claim
• ROM: M.2 2280 NVMe*1; Supports PCIe 4.0Supplier claim
• I/O: HDMI*2;DP*1;USB2.0*1;USB-C*2; USB-A 3.2*4; RJ45*1(2.5G);WiFi 6ESupplier claim
• GPU: NVIDIA Pro 4000 SFF/Pro 2000 (Total Compute 770/545 TOPS); NVIDIA RTX 2000 Ada (Total Compute 191.9 TOPS) TOPSSupplier claim
• Series tagline: Compact & PowerfulSupplier claim
• GPU compatibility: Compatible with NVIDIA Pro 4000 SFF / Pro 2000 (also RTX 2000 Ada listed in TOPS box)Supplier claim
• Max VRAM: Up to 24GB VRAMSupplier claim
• Compute: 770 TOPS (Theoretical Peak) TOPSSupplier claim
• TOPS by GPU: NVIDIA Pro 4000 SFF: 770 TOPS; Pro 2000: 545 TOPS; RTX 2000 Ada: 191.9 TOPS TOPSSupplier claim
• RAM: SO -DIMM DDR5*2，(Max. 96GB)Supplier claim
• ROM: 2*M.2 SSD 2280 NVME; Supports PCIe 4.0Supplier claim
• I/O: 2*HDMI®；1*DP；1*USB2.0 ；2*USB-C；4*USB-A 3.2 ；RJ45*1（2.5G）；WiFi 6ESupplier claim
• Chassis Volume: 3.5L 212*190*86mm mmSupplier claim
• Test platform: NVIDIA RTX PRO 4000 Blackwell SFF Edition 24GBSupplier claim
• LLM inference summary (single/multi concurrent): Model | Framework | Quantization | Single-Concurrent Peak (tok/s) | Best Config | Multi-Concurrent Peak (tok/s) | Max Context | ITL Latency ;; Qwen3.5-4B-AWQ-4bit / vLLM / AWQ-4bit / 87.27 / input=256, output=32 / 1,401 / 262K / 9~15ms ; Qwen3.6-35B-A3B-UD-Q4_K_M / Llama.cpp / Q4_K_M / 70.66 / input=1024, output=32 / — / 65K / 10~13ms ; Qwen3.5-9B-GPTQ-4bit / vLLM / GPTQ-4bit / 53.32 / input=256, output=32 / 958 / 262K / 16~18ms ; Qwen3.6-27B-Q4_K_M / Llama.cpp / Q4_K_M / 16.83 / input=1024, output=32 / — / 131K / 48~62msSupplier claim
• Maximum Context Generation Throughput (Single Concurrent, output=32, tok/s) by context length: Model | 256 | 512 | 1K | 2K | 4K | 8K | 16K | 32K | 65K | 131K | 262K ;; Qwen3.5-4B-AWQ-4bit / 87.3 / 84.9 / 79 / 68.3 / 51.9 / 35 / 20.8 / 11 / 5.3 / 2.3 / 0.9 ; Qwen3.5-9B-GPTQ-4bit / 53.3 / 51.5 / 48.1 / 39 / 30.5 / 20.4 / 12.2 / 6.5 / 3.2 / 1.5 / 0.6 ; Qwen3.6-35B-A3B-Q4_K_M / — / 60.9 / 70.7 / 67.8 / 66 / 60.8 / 52.8 / 42.3 / 25 / 失败 / — ; Qwen3.6-27B-Q4_K_M / 15.2 / 15.7 / 16.8 / 16.5 / 16.1 / 14.9 / 13.6 / 11.3 / 7.7 / 4.3 / 失败Supplier claim
• Multi-Concurrent Throughput (input=512, output=32, tok/s): Model | conc=1 | conc=2 | conc=4 | conc=8 | conc=16 | conc=32 | conc=64 | 峰值 (peak) ;; Qwen3.5-4B-AWQ-4bit / 84.9 / 145.9 / 251.6 / 376.9 / 456.8 / 548.3 / 527.0 / 548.3 ; Qwen3.5-9B-GPTQ-4bit / 51.5 / 85.5 / 151.0 / 229.6 / 282.1 / 321.8 / 325.9 / 325.9Supplier claim