Just last week, during its Q1 2018 earnings call, AMD announced that it already had 7nm GPUs already up and running in its labs. Thanks to a trusted source, the folks over at WCCFtech have stumbled upon (also corroborated by boffins over at VideoCardz) a 3DMark database entry which could correspond to a 32GB HBM2 variant of Vega 20. The fact that this entry exists isn’t really as exciting as the numbers themselves.
As with most pre-release silicon, the entry shows a generic GPU clocked at 1,000MHz, and the 32GB HBM2 memory clocked at 1.25GHz. 3DMark2011 often misreports GPU and memory clocks on early engineering samples, but in that case, they’re typically reported as 0 (which isn’t the case here), lending at least some credibility to this result.
Now, if the numbers are to be believed, AMD could be sitting on a beast of a GPU here, largely due to a combination of microarchitecture-level optimizations, coupled with efficiencies of moving to TSMC’s N7 process node. Compared to a Vega 64 Liquid running at 1750MHz, Vega 20 appears to be at least 70% faster, clock-for-clock, which is great for performance, but even better from a TDP standpoint, as you're essentially besting last-gen performance in a significantly reduced thermal envelope. In fact, Vega 10 fared quite poorly when it came to power consumption (both idle and under load) and overall heat output, especially compared to Nvidia’s 10 series GPUs based on its Pascal microarchitecture, which offered far better performance-per-watt.
To put things in perspective, the Radeon Technologies Group has recently seen some high-profile departures, including former Head and Chief GPU Architect Raja Koduri and more recently, former Senior Director for Global Product Marketing Chris Hook, both of whom have joined Intel to head up their efforts on finally developing discrete GPUs.
In any case, we’d urge you to take this result with a grain of salt - but, if such performance were to really materialize, things certainly seem to be looking up for AMD.