TensorDock vs ThunderCompute

TensorDock's per-second billing with spot instances enables granular cost control, ideal for bursty or short workloads, offering discounts up to 80% off on-demand rates via marketplace bidding. Post-acquisition stabilization reduces out-of-stock issues, but preemptions require resilient job queuing. No public details on reserved instances. ThunderCompute uses per-minute billing, likely on-demand focused, providing predictable costs without spot volatility but potentially higher baselines. This suits steady usage patterns like development sessions, avoiding per-second minimums that penalize idle time in spot models. Implications: TensorDock favors high-utilization (>70%) long runs or experiments with autoscaling; ThunderCompute benefits interactive, variable-duration tasks where setup time is minimized, though per-minute granularity may inflate costs for sub-minute bursts compared to per-second precision.

TensorDock delivers superior value for small experiments and large training runs, where spot prices (e.g., $0.20/hr A100 equivalents) slash costs for 100+ GPU hours, assuming <10% preemption downtime via fault-tolerant frameworks like Ray. Less ideal for production inference needing 99.9% uptime. ThunderCompute offers better value for fine-tuning/experimentation and real-time inference in dev environments, as VS Code integration cuts ramp-up time by hours, justifying 20-50% higher pricing for teams valuing velocity over pennies-per-second. For batch inference, TensorDock edges out on cost if volumes exceed 10k inferences; ThunderCompute wins for low-volume, interactive batches. Overall, TensorDock maximizes ROI for compute-heavy solos; ThunderCompute for collaborative dev teams.

TensorDock's marketplace aggregates bare-metal and virtualized GPUs from diverse datacenters, offering flexible storage (e.g., NVMe) and basic networking; Kubernetes support via user configs, but no native orchestration noted. Post-Voltage Park acquisition, inventory includes consumer-grade to enterprise H100s with improved uptime. ThunderCompute emphasizes virtualized instances optimized for remote dev, with VS Code extension handling SSH/Port forwarding; storage and networking details limited, likely EBS-like volumes. No confirmed Kubernetes or bare-metal options, focusing on single/multi-GPU VMs for dev workflows.

TensorDock provides high GPU availability via marketplace (spot fills in seconds), strong multi-GPU scaling on NVLink clusters, but spot variability impacts sustained perf (e.g., 5-10% downtime). Benchmarks show near-native speeds on A100/H100. ThunderCompute offers consistent on-demand perf with quick provisioning, but multi-GPU scaling and interconnects (e.g., InfiniBand) undocumented; VS Code UX aids debugging without perf overhead. Limited data suggests reliable single-GPU for dev, potentially trailing TensorDock in raw cluster throughput.