deepspeed zero stage 0 vs pytorch ddp

[asdfry]

Hello. I'm currently experimenting with multi-GPU training using Hugging Face Accelerate and DeepSpeed.
As mentioned in the title, I compared DeepSpeed ZeRO Stage 0 with PyTorch DDP and found that DeepSpeed ZeRO Stage 0 had significantly lower memory usage.
Does anyone know why this is?
I will also leave the Accelerate config I used below.

---

config (deepspeed zero stage 0)

compute_environment: LOCAL_MACHINE                                                                                                                    
debug: false                                                                                                                                          
deepspeed_config:
  gradient_accumulation_steps: 1
  zero3_init_flag: false
  zero_stage: 0
distributed_type: DEEPSPEED
downcast_bf16: 'no'
machine_rank: 0
main_training_function: main
mixed_precision: bf16
num_machines: 1
num_processes: 4
rdzv_backend: static
same_network: true
tpu_env: []
tpu_use_cluster: false
tpu_use_sudo: false
use_cpu: false

---

config (pytorch ddp)

compute_environment: LOCAL_MACHINE                                                                                                                    
debug: false                                                                                                                                          
distributed_type: MULTI_GPU
downcast_bf16: 'no'
gpu_ids: all
machine_rank: 0
main_training_function: main
mixed_precision: bf16
num_machines: 1
num_processes: 4
rdzv_backend: static
same_network: true
tpu_env: []
tpu_use_cluster: false
tpu_use_sudo: false
use_cpu: false