Yeah actually after typing that response I realized what my issue was... I had both fan curve and thermal limit at stock... and apparently even if you come within 15C of thermal limit, the card already starts throttling just in case...
My best score with stock fan curve and case closed: 5997
Fan pegged at 100%, top of case removed: 6252
100% power limit (seemed to be fine in 3DMark at least): 6292
And then I got another idea... what if the memory is eating too much of the power budget? My VRAM is stable at +1500 (Matching the RTX 3060) but is this really needed considering the other limitations of this chip?
Reduce Memory OC from +1500 to +1250: 6328
Reduced Mem OC to +1000: 6330
Reduced Mem OC to +500: 6378
Important lesson learned: at these low TDPs, don't exaggerate on the memory OC. Memory also "steals" power budget! I assume this effect is even bigger on a 12GB A2000 like mine.
I still don't really understand why the GPU keeps downclocking during the run though... I would expect it to stay pegged at 1560MHz or so but by the end of Graphics Test 2 it is hovering around 1200 MHz.
Even better news: Also Superposition doesn't trip the PSU anymore at 100% power limit (and my score is up from 3300 ish to 3719). I assume the issue was transient spikes, and the flat freq/voltage curve completely irons those out (the chip can simply not go above 750mV anymore, ever.)
I was about to chime in with very similar advice - power and temp limit are both crucial when you're aiming for those sweet sweet 3DMark graphics pts.
Yes, GDDR6 will definitely use a good chunk of your power budget, so you'll often want LOWER memory OC for a HIGHER core OC (and it's not even the same for Graphics1 and Graphics2! I even wonder if the FAN is considered within the power budget, so I don't set fan to 100%, usually opting for about 55-70% depending on how spicy the previous runs were (higher at higher voltages, 45% fan is
plenty for 750mV runs)
.
FYI my A2000 can boost happily to a higher speed on test1 but crash in test2). Test 2, similar to Furmark will have a lower core clock than you'll see in Test 1.
You won't see it hitting the actual max frequency you select on your curve much, usually 200-300MHz below. From memory when I was selecting 1590MHz stock it was actually doing more like 1250MHz. I think maybe I led you astray there, I'd have to go back and look as I don't think I was doing 1650MHz curve when it was stock... Hmm.
By having the flat curve, you'll also eliminate those moments when the card tries to boost to some stupid frequency by shooting up to 1V or more, it would be nice if Afterburner could monitor the voltage too, but at least (IIRC) GPU-Z still can.
I did a whoooole lot of trial-and-error with my card, and suspect it's the only way to get the best out of each individual A2000 sample. I would typically follow this process (or reversed if trial/error testing voltage with a target frequency):
1) set power/temp limit at max with
TEMP preference;
2) set fan at fixed point (e.g. 55%) - reduce if previous run stayed under 75C, increase if hitting 80C;
3) choose a target voltage (e.g. 893mV);
4) set it as the 'max' by dropping all V/F points after it
-- shift+click the next V/F point up,
-- shift+click the highest voltage point on the graph (highlighting
-- click+drag any of the selected points below your target voltage point;
-- click APPLY back in the main Afterburner window;
-- validate if the V/F curve shifts up/down a little bit by closing & reopening the curve window and checking your target points;
5) drag the whole curve up (alt + click/drag) to an estimated stable frequency;
-- click APPLY back in the main Afterburner window;
-- validate if the V/F curve shifts up/down a little bit by closing & reopening the curve window and checking your target points;
6) run your preferred test (I always use TimeSpy);
- if test completes, review the frequency range and temperature range
- if it crashed, consider higher voltage (only if shunt-modded) or lower frequency
Go back and start again.
GOOD LUCK!
-Nathan