AMD Vega Hardware Reviews

I thought there was some outline of the methods used, but when I tried to find the thread to refresh my memory I did not find posts that corresponded to a null or informative outcome.
I guess the next step is hopefully the open variant of the tool that was mentioned in the other Vega thread being made available at some point.
 
Has there been any reviews Vega FE where they have manged to keep the card running at 1600MHz, and stop it throttling? Via adjusting power limit/ voltage limit/ fan speed?

It's very easy to accomplish. The power limit needs to be increased to 30-40%, and max fan speed needs to be set to ~3500. I've done a bunch of benchmarking with mine and have never had a problem getting 1600mhz and much higher. I don't know why people are having trouble with wattman, it's been great for me.

Here's a random firestrike: http://www.3dmark.com/3dm/21034813

I get anywhere from 29k-31k with my overclocked 1080Ti in the same system, so Vega isn't as hopelessly far away from the Ti as some people think, at least in the ultimate e-peen measuring contest.
 
It's very easy to accomplish. The power limit needs to be increased to 30-40%, and max fan speed needs to be set to ~3500. I've done a bunch of benchmarking with mine and have never had a problem getting 1600mhz and much higher. I don't know why people are having trouble with wattman, it's been great for me.

Here's a random firestrike: http://www.3dmark.com/3dm/21034813

I get anywhere from 29k-31k with my overclocked 1080Ti in the same system, so Vega isn't as hopelessly far away from the Ti as some people think, at least in the ultimate e-peen measuring contest.
I would consider a card 22% slower than 1080 Ti using 400W to be pretty hopelessly behind it :neutral:
 
The power draw is still crazy with the WC version. I was hoping better results (because less leakage ?), and 1600mhz even within the 300W settings. It's sad.
 
Temperature does have an effect on leakage, but leakage was particularly problematic at 28nm planar. FinFETs do still leak, but it seems like the amount is much less than at the last planar nodes. Those troubles likely played a role in why 20nm planar was skipped by so many.
The better control over the channel that FinFETs have pushed back the level of leakage sensitivity to temperature, and have improved a number of other types like static and (within more modest clock ranges) sub-threshold leakage.

What was rather noticeable for Fury X probably should not be as significant for Vega.
 
Clock range does increase by 100MHz, despite the same 300W limit and power draw. So leakage reduction ia still there.
 
So in the PcPer review they don't include water cooled FE out-of-the-box settings, but use overclocked FE settings for all game benchmarks?
 
Well, it's out of the box, just after a BIOS switch. Both BIOS(i?) came out of the box. :p

Is it?
PCPer Review said:
The progression from air, to liquid, to liquid at the higher TDP is clearly visible in this graph. At stock settings and running near 350 watts, the Vega Frontier Edition card hits its target 1600 MHz reliably, though bouncing between that and 1528 MHz.
pretty much at the bottom of the page:
https://www.pcper.com/reviews/Graph...-16GB-Liquid-Cooled-Review/Power-Consumption-

I am a bit unclear on what they are presenting on the following pages. AC is clear, WC might be misleading because of the same clock speeds given as for AC - typo?
 
It's quite clearly the BIOS switch tested. 350W is the stock setting for the BIOS switch.
While we are all eager to see how the gaming performance of the Vega Frontier Edition liquid-cooled option compares to the air-cooled, I first want to dive into the different power settings and how the clock speeds of the Vega GPU are affected by the better cooling. As I mentioned above, the out-of-box setting on the card switch is toward the display connectors, in the 300-watt configuration. Moving the switch to the right, and then rebooting, puts it in the ~350-watt configuration, though there is no software indication that the change has been made, which would be a welcome addition.
https://www.pcper.com/reviews/Graph...-16GB-Liquid-Cooled-Review/Power-Consumption-
 
After reading through the overclocking section, they clearly state that 1712 MHz was the limit they could reach stably - with occasional dips back down to 1637 MHz.
In the graphics, there's two Vega FE WC - one at 1382 MHz and one at 1712 MHz. Since PCPer is giving base clocks for the other cards as well, it seems like they tested the 1382-WC-version at stock settings (no mention of BIOS switching) and the 1712-WC-version ist their manually OC'ed variant, since they do not mention anywhere that clock speeds go up with the 350W BIOS. On the contrary, they state that clocks stay at the same ceiling height with the 350W BIOS here:
https://www.pcper.com/image/view/84108?return=node/68126
 
Clock range does increase by 100MHz, despite the same 300W limit and power draw. So leakage reduction ia still there.

It would be nice if more factors were controlled for, or measured.
When Anandtech did a rough check for the temperature-related effect on power consumption, it was between separate runs of the same Crysis workload, one when the CLC was at 40C and one after the coolant loop had reached a steady state of 65C.
In that modest range, there was a difference of about 18W.

It's not a precise experiment, but it does keep silicon and board variability confined to one card. The Vega comparison is between two different chips and different treatment of the on-board components in terms of cooling. Not getting to the thermal ceiling that the air-cooled chip gets to may also reduce effects like the HBM2 stacks overheating or localized hot spots causing the GPU to scale back.

Getting a figure for how much power consumption changes between the cooled loop and the fully warmed loop can give a picture of leakage scaling, with the same caveat with Fury X that the scaling for the 65-85 range would be unmeasured. It may be a more important omission this time because FinFETs are at least expected to handle this better.

Going with the 100 MHz difference at an assumed fixed power target, something like 7% of an upclock. However, let's assume the base voltage isn't higher or that the GPU didn't select a higher clock and voltage point, which may be a shaky assumption, and that the step is modest enough to assume the difference comes down wholly to leakage savings translating into a linear increase from the clock. It's a roughly 7% increase of clock in the range between 65C and 85C if ASIC power is in the 250W range, which Fury X experienced in the 40C and 65C range.
 
I would consider a card 22% slower than 1080 Ti using 400W to be pretty hopelessly behind it :neutral:

It's a 1080Ti with one of the best aftermarket coolers vs a blower with a heatsink from a laptop though. Power draw...yeah it's a problem. But heat isn't and clocks should be much better and more stable with a competent cooler.

One odd thing I noticed is that setting voltage control to manual in wattman sets the max voltage to 1.2V by default, which results in better power usage without actually changing anything. But the card is fully stable at 1.2 even when I overclock it. Also, looked at the drivers there aren't any new registry settings related to power, I would've expected some new keys or something.
 
Wiki has the answers:

"Vega is only about a tenth of the age of the Sun, but since it is 2.1 times as massive its expected lifetime is also one tenth of that of the Sun.
...
Most of the energy produced at Vega's core is generated by the carbon–nitrogen–oxygen cycle (CNO cycle), a nuclear fusion process that combines protons to form helium nuclei through intermediary nuclei of carbon, nitrogen, and oxygen. This process requires a temperature of about 15 million K, which is higher than the core temperature of the Sun, but is less efficient than the Sun's proton-proton chain reaction fusion reaction. "
 
Vega is huge and uses HBM.. why does it suck so badly compared to GP104?

AMD have devoted too much silicon for uses that don't do much for gaming? HBCC and the FP16 doubling. Perhaps the HBM 2 controller isn't as small as for hbm in furyx.

Memory bandwidth is likely the major culprit for performance. The ln2 fury run was easily 10%ahead of 1080 while being clocked at 1400Mhz.

On the power side it looks like AMD need higher voltage for same clocks. I think Pascal can do 1600Mhz easily with 0.9V, maybe even less while AMD have to put it at 1.2V to be sure.
 
Back
Top