A few readers have been asking about performance, particularly when it comes to power usage and i'd imagine there is some serious inquisitiveness as to how CF scales too. Today i'm just compiling numbers from synthetic benchmarks rather than games purely because they are consistent and repeatable. The chosen few are Firestrike (Graphics!), Heaven and Superposition all in their most extreme 1080p form.
So the first set of data is taking a look at Crossfire scaling, the 'perfect' result here would be 400%. In order to control as many variables as possible all runs were at the same settings: 4.8ghz, +100% power target and 1095MHz rather than the highest achievable score. The CPU was clocked high to mitigate it's impact on scaling just as you would when comparing two different cards.
Firestrike: CF off - 18174, CF on - 63826,
350%
Heaven: CF off - 1838, CF on - 6620,
360%
Superposition: CF off - 3557, CF on - 13795,
388%
To help interpret the CF scaling results I ran another string of tests with the CPU frequency as the controlled variable. If these revealed a significant disparity between stock and overclocked CPU speeds that would account for some of the losses which may be falsely attributed to CF scaling. To save me making yet another run of all 3 benchmarks the GPU's are stock for stock this time.
Firestrike: 3.5GHz - 42881, 4.8GHz - 57797,
+34.8%
Heaven: 3.5GHz - 5512, 4.8GHz - 6127,
+11.2%
Superposition 3.5GHz - 11656, 4.8GHz - 11740,
+0.7%
Now for the fun variables! Just like recent Nvidia cards raising the power target lets the GPU throttle up to an elevated TDP. However unlike Nvidia, AMD didn't mess about with +20% and instead +100% is a welcomed option, so obviously I went for that. The cores can run right up to 67°C without deviating from the desired clock speed so with the EK blocks that issue was completely removed. So in this set of data we see the difference between +0% and +100%, the clock speed remains the same at 1000MHz however with +0% the card is restrained to 350W rather than being free to take as much as it likes.
Firestrike: +0% - 57797, +100% - 60436,
+4.6%
Heaven: +0% - 6127, +100% - 6254,
+2.15%
Superposition: +0% - 11740, +100% - 12784,
+8.9%
And exactly how much extra power (peak draw from the wall) did the cards take to achieve that performance gain...
Firestrike: +0% - 1065W, +100% - 1368W,
+28.5%
Heaven: +0% - 1023W, +100% - 1138W,
+11.2%
Superposition: +0% - 1010W, +100% - 1285W,
+27.2%
For the final set we look at the same thing again only this time raising the voltage, core frequency and power target simultaneously. So not only is the card free to take as much power as it needs but the demand has also increased in a true reflection of stock vs. overclock. Because the GPU's are now allowed adequate power and are sufficiently cooled this is a completely static overclock, locked to 1267mV @1095MHz as soon as 3d clocks are enabled. As in the previous set the CPU remains at 4.8GHz for both runs.
Firestrike: stock - 57797, +100% - 63826,
+10.4%
Heaven: stock - 6127, +100% - 6620,
+8.0%
Superposition: stock - 11740, +100% - 13795,
+17.5%
Firestrike: stock - 1065W, overclock - 1509W,
+41.7%
Heaven: stock - 1023W, overclock - 1266W,
+23.8%
Superposition: stock - 1010W, overclock - 1452W,
+43.8%
Right then, some form of conclusion is in order. Personally I think it's fair to say that the Crossfire scaling is exceptional, to achieve 366% on average is more than anyone would've expected at such a low resolution. To do that with PLX and PCIe bandwidth alone is impressive, but then when you consider Firestrike is still CPU limited it becomes something else.
When it comes to power though it's pretty clear that the Fiji cores will quite happily take all that's going and not give much back in return. Perhaps that was a harsh perspective to take but as a percentage the gains come across somewhat small at the side of the power thrown at them. It's clear that they very rarely operate at peak efficiency, infact the 1000MHz stock clock is stretching them out considerably but that does make them very fun to play with in 2018.
Once furnished with a waterblock (or a very speedy fan on the aio) when you type in a core frequency and voltage then it either will or won't run at that exact frequency. There are no offsets, no dynamics at play it just does as asked and even more ridiculously the 200% TDP 1267mV setting is literally in the drivers.
TLDR; big numbers right here!
4.8GHz, 4.6GHz cache, 3200MHz C15, 1100/500MHz, 1525W
4.8GHz, 4.5GHz cache, 3200MHz C15, 1120/500MHz, 1286W
4.8GHz, 4.5GHz cache, 3200MHz C15, 1115/500MHz, 1472W
Running flawlessly all day, 51°C was the absolute hottest the cards ever got, PSU didn't trip. Might run some games tomorrow, check out how GTAV scales.
JR
