RX Vega efficiency

[wowsers]

So I was bored a few days ago and decided to try and find some info on how the e-pene factor of the currently known RX Vega cards we have info on stack up against the various consumer/enthusiast GTX 10XX cards.

I wanna make it clear that these values really mean nothing when it comes to gaming, drivers and optimizations plays a way larger role than pure "performance" metrics like GFLOPs used here.
Anyways, onto efficiency!

Spoiler

 

 

These numbers puts the RX Vega right in line with how well the GTX 10XX cards "performs" per watt (and I threw in the GTX 9XX cards for fun), again this is at the moment just e-pene talk, but it sure gives us some interesting data to go by. For example the performance seems to scale perfectly between the amount of stream processors and speed of the graphics card, which would put a hypothetical RX Vega 48 at ~8600 GFLOPs and a hypothetical RX Vega 32 at ~5500 GFLOPs. This would make the RX Vega a nice in-between if real gaming performs similarly to GTX 10XX cards per watt.

Here's the price/e-pene performance graph:

Spoiler

These prices might not represent the exact reality thanks to some (GTX 9XX series) cards having already been out for a while and are being phased out, this graph is also the reason I choose to remove the GTX Titan cards, they would screw up the more consumer/enthusiast oriented pricing that the Ti models have. Anyways, as you can see RX Vega is almost guaranteed to give more e-pene for your hard earned money.

Hope you enjoyed looking at these graphs I made while learning to use google sheets...

[Mira Yurizaki]

If you really want to trigger someone, do FP64 FLOPS instead.

[wowsers]

Just now, M.Yurizaki said:

If you really want to trigger someone, do FP64 FLOPS instead.

Nah that would require me to put more than a couple minutes of effort into this Think FP16 FLOPs would be more relevant considering fragment shaders (visual filters and such) don't normally need full precision.

[wowsers]

I've now looked at a few more statistics, and they hold both good and bad information... first though, some more metrics to figure out the actual Vega performance.

When looking at the FPS average and 99'th percentile we can see that RX VEGA 64 performs about 12 or so percent more than a GTX 1080:

Spoiler

This leads me to believe at least the RX Vega 64 performs about 21% worse than what the e-pene value of GFLOPs would lead you to believe.
So I decided to look if that could be observed someplace else, and found the performance metric on adoredtv's "RX Vega Unveiled" video he posted yesterday or so.
The image in question shows power efficiency amongst other things, 220W RX Vega, 150W RX Vega, and a GTX 1080 FE

Spoiler

With this you can start trying to extract some meaningful data, in my case it was figuring out if there was a correlation between the perlin noise efficiency and the gaming performance, and as it turns out if we apply the 220W RX Vega efficiency compared to the Pascal efficiency curve we have a winner. Since the Vega cards seems to be roughly in line with Pascal on theoretical performances we can just use the closest card, the RX Vega 56. The equation becomes 2.4/3.2*10544 (and the theoretical max is just used as a reference point, Pascal cards are assumed to be the baseline), and this yields 7908 or 25% less performance compared to the theoretical max. This felt way too close to the 21% worse than theoretical max that the doom benchmark had. So to test if this actually is a connection with performance I then decided I should calculate the efficiency of the still unknown 150W RX Vega card.

 

A quick guess in what the 150W card might be is an RX Vega 48, this should follow the naming scheme of the other two RX cards (that are air cooled). To calculate the theoretical max performance of this card I just extrapolate the performance delta between an RX Vega 64 and an RX Vega 56 (negatively), this gives me 8415 GFLOPs, and assuming 150W that should in theory be in line with the other RX Vega cards in terms of performance/watt. Then we take a card of the Pascal line with a similar wattage, the 1070 (which should in theory perform the same per watt as a 1080), then I took the calculation (3.1/3.2)*6463 (6463 being the Pascal baseline) and got 6261. Calculating the delta in performance between the theoretical performance and the perlin efficiency benchmark gave me (drumroll)... 26%.

What this could mean is that whatever RX Vega 48 is it's gonna be right on the cusp of Pascal level performance per watt, RX Vega in general seems to perform roughly somewhere in between 20 and 25% worse than it's theoretical max (when compared relatively to Pascal cards). And of course I made a nice graph outlining the more relevant real life performance of RX Vega cards:

Spoiler

 

Potentially we're gonna see some really attractive consumer graphics cards later on, especially if RX Vega cards actually do follow roughly the same performance curve I've tried to calculate, whatever the case may be the future is looking bright for us consumers (especially entry level consumers).

[Darkseth]

The true potencial of Performance per Watt shows when you Undervolt

Compared to that, running Pascal at Stock has a VERY bad Efficiency tbh ^^"

 

My GTX 1080 MSI Gaming X @ stock consumes around 200-220 Watt alone. Slightly overclocked from the Founder's edition, LEDs, higher Powerlimit etc.

When i undervolt it down to 0.8 Volt (Stock is around 1.025 - 1.031 Volt), Clock down to 1823 (from 1963), the Card consumes only 130~ Watt or so, even if i Overclock the Memory from 5000 to 5500 Mhz.

performance loss is around 5-8%.

Power Draw reduced by 40%

 

My GTX 1080 this way is STILL faster than every Founder's edition running @ stock. But consumes a few single Watt more than a GTX 1060. 

 

 

I do hope, Vega shows similar potencial with Undervolting there.

Bringing the Vega 64 down to 200-220 Watt, instead 300 Watt would be pretty good, when it means sacraficing 2-4% performance.

[wowsers]

1 hour ago, Darkseth said:

The true potencial of Performance per Watt shows when you Undervolt

Compared to that, running Pascal at Stock has a VERY bad Efficiency tbh ^^"

 

My GTX 1080 MSI Gaming X @ stock consumes around 200-220 Watt alone. Slightly overclocked from the Founder's edition, LEDs, higher Powerlimit etc.

When i undervolt it down to 0.8 Volt (Stock is around 1.025 - 1.031 Volt), Clock down to 1823 (from 1963), the Card consumes only 130~ Watt or so, even if i Overclock the Memory from 5000 to 5500 Mhz.

performance loss is around 5-8%.

Power Draw reduced by 40%

 

My GTX 1080 this way is STILL faster than every Founder's edition running @ stock. But consumes a few single Watt more than a GTX 1060. 

 

 

I do hope, Vega shows similar potencial with Undervolting there.

Bringing the Vega 64 down to 200-220 Watt, instead 300 Watt would be pretty good, when it means sacraficing 2-4% performance.

I'm not sure how well it'll undervolt (as in I really honestly don't even know half of the factors), but it should behave similarly to how Ryzen behaves considering they use the same manufacturing for the actual dies.

If we take a Ryzen 1800X, it's 95 watts TDP, the die size of a Zeppelin (Ryzen) die is 192 mm², the die size of a Vega 10 die is 484 mm², 484/192*95 = ~240W. Then when you look at the RX Vega 64 it has a slightly higher TDP (probably due to a higher density of working transistors), so AMD might have done the same thing as Ryzen and already pushed the safe limits of what you can extract from the chip (until 14nm LPP comes around or something).

[Nena Trinity]

That Vega 56 looks like a worthy replacement for my GTX 970.