Three computational structure-prediction backends running in parallel: NIM, Dell, DGX Spark GB10
Three structure-prediction backends now running in parallel
The platform operates three independent GPU compute backends, each running structure-prediction and protein-design workloads continuously. All figures below are measured on our own runs, not vendor specifications.
Active backends (2026-04-27)
| Backend | Hardware | Workload | Status |
|---|---|---|---|
| NIM API (cloud) | NVIDIA hosted | Boltz-2 + ESMFold + MolMIM | 162 calls/5 min, 86% OK |
| Dell Demo Center | RTX Pro 6000 Blackwell 96 GB | Boltz-2 PPI prediction | 4,200+ pair predictions, 82% GPU utilization |
| NVIDIA DGX Spark GB10 | Grace Blackwell 128 GB unified, sm_121 | Chai-1 ligand prediction + local LLM | 21+ predictions, ~33/hour throughput |
Measured hardware benchmarks
Identical workloads run on each backend produced comparable timing data, published at /infrastructure/gpu-benchmark:
| Workload | Spark GB10 (48 SMs) | Dell RTX Pro 6000 (188 SMs) |
|---|---|---|
| LLM tokens/sec (Qwen 35B Q8) | 49.8 t/s | 186.6 t/s (3.7x faster) |
| Matmul BF16 8192² | 99.9 TFLOPS | 395.4 TFLOPS (4.0x) |
| Memory ceiling | 122 GB unified | 96 GB GDDR |
Trade-off: the Dell card wins compute throughput per second; the Spark wins memory capacity for larger models that exceed 96 GB.
Methodology notes
- Boltz-2 vs Chai-1 dual deployment: Boltz-2 hangs on Spark sm_121 with PyTorch 2.11+cu130 (Lightning Predict phase blocked); Chai-1 0.6.1 runs natively on the same hardware. Spark therefore runs Chai-1, Dell runs Boltz-2. Both are independently published methods and provide complementary scoring.
- BindCraft dual-mode: Spark runs design + AF2-confidence filtering (Bennett 2023 methodology); Dell runs full BindCraft including PyRosetta scoring (no aarch64 PyRosetta wheel exists).
- We use BindCraft 1.5 rather than RFdiffusion for binder design, following the higher reported design-success rates in Pacesa et al. (Nature 2025). These are computational design-success rates from the source publications, not our own experimental confirmations.
Storage architecture
The canonical research-data root was migrated to Spark /data/research-data/ (3.6 TB capacity, 215 GB currently hosted), with an automated cloud mirror and a daily migrator that pulls staged fleet results into the canonical tree.
All predictions on this platform are computational. None have been experimentally validated. Source code and raw benchmark data: /api/v2/infrastructure/gpu-roi.