Even if I didn't look up this guide until after I'd undervolted my CPU ( ::) ), it was still great to confirm my assumptions and have everything verified in a thorough manner. Thanks for a great guide!
Also, either your average UV numbers (-0,060 to -0,080V for KBL) are conservative, or there's really something to Intel's binning process: my i7-7600U happily runs both CPU, Cache and iGPU at -0,100V. 0,110 causes crashes, but 0,100 is stable after 6 hours of Prime95/3 hours of 3DMark looping. Good enough for me.
Also, the UV gave me a 10% increase in Cinebench scores, and generally allows for significantly higher boost clocks - around 3-400MHz in the XTU Stress Test, at no change or a minor reduction in temperatures. Not bad at all for maybe 20 minutes of actual effort.
Will: yes, it seems that core and cache voltages are either linked or XTU treats them as linked (makes sense as cache clocks are sort of linked to CPU clocks). Adjusting one adjusts the other, at least for me.
What kind of graphics stress test are you running? Something reasonable (3DMark, Unigine Heaven or Superposition), or a power virus like FurMark or MSI Kombustor? The latter will at least tend towards making things crash "unrealistically", i.e. inducing loads you'll never, ever see in real-world usage. On paper, undervolting the CPU shouldn't affect GPU tests at all, but of course the CPU will always be active feeding the GPU data, so there's probably some correlation there. I'd try stepping back the voltage a few steps and redoing the GPU stress test until stable - though unless you game on your ultrabook (or require absolute 100% stability), you're probably okay leaving it at -0,110V.