A Look at RX Vega 64 Efficiency

[Bosco]

Tonight Jim takes a look at RX Vega 64 Efficiency.

 

Check out the review here:

 

http://www.overclockersclub.com/reviews/vega_efficiency/

[Guest_Jim_*]

Just noticed there is some odd formatting issue on page 6, where all of the graphs and P-State tables are supposed to be, causing the results for only one game (Wolfenstein 2: The New Colossus) to be present. So, for anyone interested, here is a link to a zip with everything: graphs; P-State tables (in the Clocks.txt files); original data; and R-scripts.

Vega Efficiency - Data and Graphs.zip (176 MB)

[Molecular]

Guest Jim, thank you for the hard work you put into this comprehensive article and also thanks to ANTITHAT for the ZIP. Love the graphs, what application did you use to generate them?

[Guest_Jim_*]

I am Guest_Jim_*; ANTITHAT is just my member title. Anyway, I used R to generate all of the graphs, and the scripts for doing so are in the ZIP. (The scripts will actually generate PDFs and not PNGs because I found converting the PDFs to PNGs with another application reduced aliasing compared to R directly creating the PNGs.)

[Guest_Jim_*]

Just noticed last night I screwed up putting in the graphs for the Killing Floor 2 page, putting in some Wolfenstein 2 graphs instead. If you were to look at the specific file names you will see they are only different by one digit, so I must have missed changing them when I was pasting the embed code into that page text. Luckily I made it so the game title and set title were in the graphs, so we can easily see which are wrong.

Anyway, I don't think the incorrect frame time and frame rate graphs are so interesting, and you can find all the correct ones in the ZIP linked above, but the three ASIC power usage graphs are also wrong on that page, so here are the correct ones.

My only defense is it is a lot of graphs I was working with, and the way I was composing the article (combining the formatted text and the graph embed code in an offline text editor) was perhaps not the best way to ensure I avoided these mistakes.

[Guest_Jim_*]

For anyone interested in looking at the scripts I used in this article and in the reviews, and who do not want to look at the ZIP I shared above, here is a link to a GitHub repository I have for them.

GuestJim/OCC-OCAT

 

There are three scripts involved with making each final script: a Batch file for easy drag-and-drop use; a Python script the Batch file calls and passes information to; and an R source script that the Python script makes a copy of and then replaces the appropriate references within the copy. Technically I also used another script to convert the output PDFs to PNGs, but there are many ways of doing that a flag to just output PNGs from R directly, so I am not including that one. (It just calls ImageMagick to do the conversion.)

I cannot promise they are ideal but there is a Fully-Commented branch on this GitHub that contains copies of these files but filled with comments explaining the different things. (Technically they are not fully commented, but most of it is done really just skipping the ggplot stuff, which is well documented elsewhere online.)

[Guest_Jim_*]

Just want to give a quick update on something. I was hoping to do a follow-up article, looking more closely at some of this, especially the odd behavior at stock seen with Killing Floor 2, but after starting to collect some data for it, I discovered I overlooked a detail in Wattman. For the Stock runs I did, I did indeed run my RX Vega 64 at stock settings, which for voltage is leaving it on Auto. Undervolting necessitates using a Manual voltage setting. If you try running at the stock voltage value (1100 mV) with the Manual voltage setting, you do not get the same behavior. This means the Auto voltage control is required for this drop in power usage to occur, and not just some weird quirk to how undervolting impacts the switching of P-states.

With this revelation it means such a follow-up article is in limbo. I do not want to abandon this, but for now, that behavior cannot be more closely examined than I already have, short of collecting data at manually set stock voltages. (I can already tell you the results will almost certainly be more power being used compared to the undervolted runs, across the board.) On the bright side though, this does suggest someone not touching Wattman but using a feature like FRTC, or Radeon Chill, may benefit from this significantly reduced power usage, when playing the appropriate game and using appropriate frame rate targets/limits.