FluidStack vs Salad

FluidStack bills per-minute for greater granularity than hourly models, supporting spot instances for up to 90% discounts on spare capacity alongside on-demand options. This suits variable-length jobs but incurs overhead for sub-minute tasks. Salad's per-second billing offers finer precision, ideal for bursty or short experiments, with spot instances leveraging its decentralized network for rock-bottom rates. Neither prominently features reserved instances, emphasizing flexibility. Implications: Salad minimizes costs for intermittent usage (e.g., <5min jobs save ~80% vs per-minute), while FluidStack favors sustained runs (hours+) where per-minute aligns with training epochs. Spot availability risks preemptions, demanding checkpointing; FluidStack's aggregation may yield steadier spot access via diverse facilities.

Salad provides superior value for small experiments and fine-tuning, where per-second spot pricing on consumer GPUs can drop below $0.05/GPU-hour, enabling 10x+ throughput on shoestring budgets despite variability. For large training runs, FluidStack delivers better value through reliable multi-GPU scaling and global capacity, avoiding Salad's fault-tolerance overhead. Production inference favors FluidStack's consistent data center performance for low-latency SLAs, while Salad shines in massive batch inference (e.g., 100k+ samples) with residential economics. Overall, Salad maximizes ROI for interruptible, high-volume workloads; FluidStack for predictable, high-stakes compute where downtime costs exceed savings.

FluidStack employs a supercloud aggregator model, virtualizing bare-metal GPUs from Tier 1-4 data centers with unified APIs, supporting Kubernetes, high-speed networking (InfiniBand/RoCE), and block/NVMe storage. Global footprint aids multi-region deployments. Salad uses a decentralized P2P network of consumer hardware, offering virtualized access via custom orchestration; Kubernetes-compatible but with residential internet limiting bandwidth. Storage is ephemeral/object-focused, lacking enterprise NVMe options.

FluidStack delivers consistent high-end performance (e.g., A100/H100 at full spec) with reliable multi-GPU scaling via NVLink/InfiniBand, though underlying facility differences may affect interconnect speeds. Salad provides abundant consumer GPUs (RTX 3090+) at low cost, excelling in parallel embarrassingly parallel tasks but with variable clocks, no NVLink, and 10-50% lower effective TFLOPS due to heterogeneity and preemptions. Availability favors FluidStack for urgent needs.