Option A is correct. GPU-accelerated workloads, such as LLM inference and model training, require instance families specifically designed with GPUs. The documentation lists instance family names starting with 'GPU' for this purpose, such as or 'GPU_GCP NV L4 Option B is incorrect. While 'MIN NODES and 'MAX NODES define the range, the size of compute clusters in Snowpark Container Services does ''not'' auto-scale dynamically based on workload demand. Users must manually alter the number of instances at runtime using commands like 'ALTER SERVICE MIN INSTANCES = s. Snowflake does handle load balancing across instances within the configured node counts. Option C is correct. The 'AUTO_RESUME = TRUE parameter, when specified during compute pool creation, enables the pool to automatically resume operation when a service or job is submitted, removing the need for explicit SALTER COMPUTE POOL RESUME commands. option D is incorrect. Setting = prevents the compute pool from automatically suspending, meaning it will continue to consume credits even when idle. This would generally lead to higher costs, not cost optimization, unless the pool is constantly active. The default is 3600 seconds (1 hour). SPCS Compute Nodes have a minimum charge of five minutes when started or resumed, making intelligent use of auto-suspend important for cost management. Option E is incorrect. Snowpark-optimized warehouses are a type of 'virtual warehouse' and are recommended for Snowpark workloads with large memory requirements or specific CPU architecture, typically for single-node ML training workloads 'within a warehouse'. SPCS compute pools, however, provide their own dedicated instance families (CPU, HighMemory, GPU) for containerized workloads, abstracting the underlying infrastructure and supporting distributed GPU clusters directly within SPCS, not Snowpark-optimized warehouses as a 'compute pool type' for SPCS.