hardware//GPU//discrete

A discrete GPU is a standalone GPU on its own PCB with dedicated GDDR or HBM memory, connected to the host CPU through PCIe or NVLink. It delivers maximum compute throughput for ML training, rendering and HPC.

A discrete GPU is a standalone GPU on its own PCB with dedicated GDDR or HBM memory, connected to the host CPU through PCIe or NVLink. It delivers maximum compute throughput for ML training, rendering and HPC.

Data must be transferred explicitly between the host RAM and the device memory, which adds latency but gives massive bandwidth on the card. Power draw ranges from 75 W for entry cards to 700 W and more in the datacenter; a typical hand-sized card at 450 W serves car cameras, heavier data streams and datacenter workloads.