No reason other than "I was curious". This is not perfectly professional (as is obvious from Vega's inclusion); I took pains to make the graph true to the original data, but the data sources are not uniform. My two were captured by running prime95, adjusting max CPU frequency in OS power management tools, and watching Vcore (at stock settings, no OC involved). Vega's also appears to be the stock power management's curve, though of course it doesn't quite fit here (and I doubled its frequencies to put it on a more comparable part of the graph). The Zen curve is from The Stilt's measurements here
. I don't know how he captured the data, but he clearly knows what he's doing.
More data would be great! I wish I had more hardware around to test. If you post it, I'll add it to the graph. Sparse datasets (to make it quicker to measure) are no problem, especially at low voltages where the line should be straight anyway.Here's
the original image - the one below is updated.
.... And of course I've got a pile of unanswered (sometimes unanswerable) questions to go along with it:
What parts of this does chip-to-chip variance mostly affect? I thought thresholds were pretty consistent on a given process (not that I'd know), and the two Intel chips here both look to be on threshold at 800 MHz (voltage drifts oddly at that speed and no other, possibly due to temperature). The slope of the line (where straight) must be pretty consistent if it's precisely the same across two chips that different. Zen and Vega both cornering at 1000mV suggests that point may also be fairly constant. That would mainly leave the shape of the line after the corner.
Extrapolating my G3258's line out with no corner results in 4.8 at 1.36, which is much closer to reality than I would have guessed (for Haswell in general). At least some Haswell chips must either corner very late or not do much of anything after the corner. Mine isn't so golden - it used to do 4.3 at 1.33 and has degraded substantially since. That's still a lot better than it looks like Skylake should be able to do by that line. At that rate, a 6700K would take nearly 1.5V just to turbo. This i3-6100 does seem to corner strangely low; 2.2 is actually the last sample on the straight line I got. (Windows ramps it up mostly in 200 MHz increments, but jumped 2.2 -> 2.5, which seems like it might be hiding a spike.)
How do Broadwell and Kaby compare, and how will Cannonlake? It feels like Broadwell and Cannon have similar things going on, but I'm curious what that thing looks like on a graph like this. Does Intel still use different process variants for mobile and desktop? If so, how does their mobile process differ?
Zen and Vega have fairly similar shapes, but for a given voltage bump, Vega gains a lot more speed (as a percentage). That seems odd. Could it maybe be something about transmission limitations in the metal layers versus limitations in the transistors themselves? On that line of thought, does Skylake-X have anything similar going on?
GloFo's 7LP is purported to be rather HP. How much of this wondrous low voltage performance we see Zen turning in will it be able to maintain?
14LPP's marketing slides say 800mV nominal, 945mV overdrive. This doesn't at first glance appear to have much bearing on Zen and Vega's characteristics. Does it actually? If so, can we learn anything from seeing that the equivalent marketing slide for 7LP says 750mV nominal and 850mV overdrive?
I've so far been operating under the assumption that chip wear and power use are affected by voltage changes in about the same way regardless of what the starting voltage is. Is that true? The corner in these lines seems like it could have some other discontinuities associated with it.