Article by Federico Barutto
The way Vega GPUs work is very particular, thanks also to a bad description of a fundamental WattMan setting. “HBM2 voltage” isn’t actually the voltage of the HBM2 chips (fixed at 1.2V on RX Vega 56s, and 1.35V on RX Vega 64s and Frontier Edition), but is more like a GPU voltage/frequency threshold.
By lowering this “HBM2 voltage” (Vbase, shortly) to 850mV you have the opportunity to undervolt nearly as much as you want the GPU, heavily reducing power consumption without losing too much performance if the card is water cooled, or actually enhancing them if the card is air-cooled.
If the power limit (PL) is kept at 0%, every GPU voltage (Vcore) over 1130mV is useless, since the card reaches its 220W (=~295W to the PCIe connectors) PL. On a watercooled card, GPU frequency is around 1580MHz.
By lowering Vcore to 950mV, GPU frequency lowers a bit to ~1530-1550MHz, but power consumption lowers dramatically to ~145W during rendering and around 115-165W during gaming.
And by lowering Vcore to 800MHz (lowest permitted voltage in WattMan/OverdriveNTool) frequency lowers to 1420MHz and consumption lowers around 125W during rendering and around 85-145W during gaming!
By doing that there’s a (non-) problem: the lower is the Vcore the higher is the lowest HBM frequency which makes possible to set the custom Vcore and Vbase. For example, with Vcore at 950mV and Vbase at 850mV this frequency is 965MHz, and with Vcore 800 and Vbase 850 it becomes 1000MHz… This isn’t bad, since Vega GPUs (as nearly every recent Radeon) love memory bandwidth. Raising the bandwidth from 484GB/s (945MHz) to 563GB/s (1100MHz) performance increases without any power increase. Think about that: rendered frames by FC5 go from 4750 to 4900 (+3%) by simply overclocking the HBM2 to 1100MHz, and LuxMark scores go up from ~31000 to ~33600 (+8.5%)!