
NVIDIA H20 96GB (China-spec)
An oddball with a niche: modest compute but full-fat HBM3 bandwidth and 900 GB/s NVLink. For memory-bound serving of large models, H20 delivers Hopper-class token throughput at a fraction of H100 money.
MOQ 1 unit
| Architecture | NVIDIA Hopper (GH100, export-compliant) |
| Memory | 96 GB HBM3 |
| Memory bandwidth | 4.0 TB/s |
| FP8 performance | 296 TFLOPS |
| FP16 performance | 148 TFLOPS |
| Interconnect | NVLink 900 GB/s |
| MIG support | Up to 7 instances |
| TDP | 400 W |
| Form factor | SXM module / 8-GPU HGX-style platforms |
| Positioning | Memory-bound inference, KV-cache-heavy serving |
| Warranty | 12 months, SYNDROME AG |
An oddball with a niche: modest compute but full-fat HBM3 bandwidth and 900 GB/s NVLink. For memory-bound serving of large models, H20 delivers Hopper-class token throughput at a fraction of H100 money.
Current inventory: 7 units across two free-port hubs. Quoted FOB Singapore / Dubai; ask the desk for landed-cost calculation to your port. Units ship with serial-number records; refurbished stock comes with test reports.
Air and sea freight from Singapore or Dubai under EXW, FOB, CIF or DDP (Incoterms 2020). Export clearance and shipping documents are handled by our hubs. Commercial terms are structured individually for each deal and jurisdiction — your manager will walk you through the options.
New hardware carries a 12-month warranty (refurbished — 6 months). RMA is handled through the nearest hub with advance replacement from stock where available. Serial numbers are recorded for every unit that leaves the warehouse.
