otprilike kao i kod rx4xx/5xx serije, undervolt+power offset...
Citiraj:
HBM2 vs. Core Overclocking on RX Vega 56
Writing at the tail-end of back-to-back 22-hour work days – one eye now closed to block the monitor’s light – makes recollection of overclocking steps a challenge, but here’s what we’ve got:
The behavioral outcome of HBM2 vs. core overclocking is going to change based on the application, so these results will not apply evenly across all games. Some are more memory-intensive than others. We are using FireStrike for rapid prototyping. Impact in real-world games will vary heavily, though we do find that most games somewhat follow the behavior of FireStrike, even if the scaling isn’t identical.
http://media.gamersnexus.net/images/...bm2-v-core.png
Here’s a look at FireStrike 1080 graphics score increments as we overclock the card. Stock, we’re at 18816 points, for an AVG FPS of 90 in GT1. Increasing power target by 50% boosts us to 21188 – no other changes – and our power consumption goes from about 196W to about 300W at the PCIe cables. We’ll talk about power more in a moment. Anyway, that’s a gain of 12.6% performance from the power target offset. That’s not linear to all games, of course, but is significant here.
If we overclock HBM2 and offset the power target, we end up about 3.6% boosted at 950MHz over just the power target offset. That’s not a bad gain from HBM2 only. Overclocking to 980MHz HBM2 and with a 10% offset on core – because manual input didn’t hold – boosts us 6.4% over the power offset with no HBM2 overclock, or nearly 19.6% over stock. The memory overclock and power offset alone get us to 19.2% over stock, and the power offset gets us 12.6% over stock.
The take-away here is that HBM2 overclocking and the power offset are far more important than core overclocking on this very limited version of Vega. Core overclocking isn’t worth it for most users, though we can recommend considering the power offset approach. Naturally, this massively increases power consumption (by 50% -- go figure), but we have a solution for that, too – undervolting. We’ll look at that shortly. We’re completely starved for power up until the 50% offset, and even thereafter. This card is capable of more, but unless an AIB partner makes a custom card without the security check, we’re choked at 300W when the board and GPU can both easily handle 360W.
Still, a 180MHz OC on HBM2 gets it past VFE stock HBM2 speeds (945MHz), so not bad on that front. The performance uplift from an HBM-only approach is noteworthy, and more worth the effort. Core overclocking, in our still-limited and expanding experience, does not seem to impact tests quite as much as HBM+power offset. That said, we’re still experimenting with this card, so it’s not a closed book yet. There may be more to it. We are having trouble pushing beyond 9-10% on the core, though. We managed 10.5% for synthetic tests, but couldn’t get it to hold in games. For this reason, we’re revisiting overclocking at a later date for game benchmarking. Need more time to stabilize the OC.
AMD RX Vega 56 Undervolting
So, then – overclocking was disappointingly limited, might as well try the opposite. Let’s undervolt this thing.
http://media.gamersnexus.net/images/...56-power_2.png
Here’s a chart showing Vega 56 power as measured at the PCIe cables with a current clamp, running 12.3V. In absolute stock configuration and during a half-hour burn-in with FireStrike, the V56 consumes on average about 180W at the PCIe cables, not accounting for the other 28W or so at the PCIe slot. The PCIe slot is primarily responsible for fan power, at 2.4A and 12V – though this isn’t always true, it is on V56.
As for the line denoting a +50% offset, remember what we learned earlier: increasing power target alone can stabilize clocks and improve performance by upwards of 12% in FireStrike. Of course, doing this also increases power and lands us at 270W sustained power consumption at the cables, so it’s a bit of a trade.
When undervolting, we managed to stabilize at 1025mv (allegedly -- software is untrustworthy, and we haven’t probed during undervolting yet) with DPM states 6 & 7 set to a sort of fake “1652MHz.” Note that this creates a clock output of about 1524MHz, because AMD now has a dynamic clock, so 1652 becomes 1524 in reality. Anyway, we’ve dropped from 1200mv to 1025mv, increased the power target, and increased the clock speed overall. The result is here:
http://media.gamersnexus.net/images/...56-power_3.png
The first 300 seconds or so was when we were trying to work with AMD’s software issues, at which point we gave up on WattMan and resorted to Wattool. Doing this made undervolting work better, as seen when the line levels-out at around 210W. Note also that the line is nearly perfectly flat now, meaning that we’ve controlled for fluctuations in power delivery and clock frequency. The result is smoothed-out performance on the whole, while drawing 55W less power than the offset V56, but 30W more than the stock card. Ultimately, we get way better performance overall. This could be tuned to a finer degree, we think, though we stopped there for now.
RX Vega 56 Undervolting Frequency Plot
Let’s look at a frequency chart.
http://media.gamersnexus.net/images/...6-freq_3_2.png
We’re at a higher clock than just the 50% offset – 1524MHz, steady as a sniper, versus 1475MHz. The stock card reports “1300MHz” here, resulting in a 224MHz boost for 30W more power – not a bad trade at all, and there’s plenty more room to play if getting serious about it.
Of course, the frequency numbers partly rely on the software being accurate – but validating with performance seems to align things, and our power numbers are from external tools.
RX Vega 56 Undervolting Temperatures
http://media.gamersnexus.net/images/...6-temp_3_1.png
Finally, here’s a look at temperatures. The fan was left alone for all these tests, so there’s obviously maneuverability for users willing to move speeds around.
With the V56 at +50%, we were at 84C by the end of the test. The stock card and undervolted card both fluctuated around the 74-75C mark. Undervolting gave us more clock speed, a middle-of-the-road power metric, and didn’t impact thermals negatively. That’s a big improvement.
|
|
