This post has been de-listed
It is no longer included in search results and normal feeds (front page, hot posts, subreddit posts, etc). It remains visible only via the author's post history.
Tl;dr at the bottom.
Alot has been said about how undervolting a Vega 56/64 is pretty much mandatory and how bad the cards perform out of the box. But how bad is it really and where's the sweetspot?
I set out to collect some data on undervolting and overclocking my card. The evidence is obviously purely anecdotal but maybe others can use this as guidance or adapt my methods to find the sweetspot on their specific cards.
Equipment
- Sapphire Nitro RX Vega 64
- BIOS 240W
- Ryzen 5 2600x@stock
- Software: AMD Wattman, AIDA64, HWinfo, Unigine Heaven, Unigine Superposition, 3DMark Firestrike, The Division
Method
First i set out to find the highest stable clocks for core voltages from 900 to 1200mV. I ran Unigine Heaven 1080p in the background and set the core voltages and a very conservative target clock for p7. I raised core clocks fairly quickly in 5Mhz increments until a crash occured and then stepped down 10Mhz.
This resulted in stable overclocks for all voltages up to 1050mV. At 1100 and 1150mV I had to step down an additional 5Mhz to get stable benchmarks and 1200mV would not be stable at frequencies higher than those at 1150mV (although AIDA64 stress test was running fine at 1740Mhz effective). Instead I added tests for the stock settings with Power Level 50%.
Next I used AIDA64 to put synthetic load on the card and let it reach peak thermals. Then I used HWinfo to measure average effective clocks and Chip power consumption.
Afterwards I first ran 3DMark Firestrike on all settings since I found this to be the most likely test to crash the system then Unigine Superposition and lastly The Division benchmark.
All three benchmarks were onlx run once (sorry) and the results averaged and expressed as percentage of stock settings.
All test were done with HBM2 at 1050Mhz and voltage floor at 900mV.
Thermals were never a problem with the highest core temps at 76° C throughout all test.
Observations
- Raising the core voltage above 1150mV and frequency target over 1800 did not lead to stable settings. I guess there's the limit of my silicon.
- Under 1000mV the HBM2 will throttle down to 800Mhz, under 950 it will throttle to 500Mhz!
Results
This graph shows the target clocks for p7 and the stable frequencies achieved under AIDA64 synthetic load for each undervolt. Note the bad performance for stock settings.
Power consumption in AIDA64 shows that at stock PWL 50% the card is not using the available power. This is due to the fact that the 1630Mhz frequency target for p7 only results in 1570Mhz effective clocks thus not nearly using the available power.
Here's the important numbers. At 1100mV my card will consume as much power as stock PWL 50% with 10% more performance and at 950mV will perform as good as stock with a hefty 38% less power uptake.
But at 950mV the HBM2 will throttle to 800Mhz and thats just leaving performance on the table. So the sweetspot appears to be 1000mV with 7% improved performance and more than 25% reduced power.
Could AMD have done better? I suspect the very conservative settings were put in place to achieve higher silicon yields. This is underlined by the fact that there are Vegas out there that don't undervolt well or at all. Those are sadly the loosers of the silicon lottery.
Maybe AMD could have binned the chips and released a "locked" version of Vega that has conservative settings and an "unlocked" version that comes factory undervolted/overclocked and lets you play around with Wattman even more? That sounds alot like intel marketing strategy...
Or is my chip a "golden chip"? Can other people weigh in with what their cards achieve when undervolted?
Tl;dr
On my specific card, I can either get 10% extra performance at roughly the same power consumption or the same performance at a third less power uptake. Sweetspot is at 1000mV.
Post Details
- Posted
- 6 years ago
- Reddit URL
- View post on reddit.com
- External URL
- reddit.com/r/Amd/comment...