Fellow members I would like to discuss HBM clocking
So basically when I got my Fiji cards I noted AMD Matt had stated on a few forums that HBM clocks in discreet steps, 500.00/545.45/600.00/666.66MHz. Now bear in mind this guy is involved with AMD and also that he has no reason to share misinformation. I was surprised by this information, as when HBM overclocking became available in MSI AB, the slider would not increment in such steps as AMD Matt had posted about.
Basically HBM overclocking via MSI AB (other OC tools) will not increment in steps as it has no knowledge how HBM increments.
Next let's look at OverDrive page, the slider for HBM clock became available through bios mod, again OverDrive has no knowledge how HBM increments. If you install older Catalyst driver it will increment in 1MHz steps, with Crimison drivers it will increment in 5MHz steps.
Next VRAM_Info, this section sets up the HBM RAM. In that section we can see there are 100MHz, 400MHz, 500MHz and 600MHz straps/timings, this information is not again solid proof how HBM steps.
So basically for months viewing at all of above I was confused what is going with HBM
. 3DM13 had been the most sensitive benchmark on Hawaii for RAM tweaks, so I used it on Fiji as well. Now as HBM performance scaling was so small, it become difficult to use this bench data to categorically say how HBM clocks.
Then I thought stability testing is the answer, like highlighted in my previous posts I had assessed 1135/535 was stable for:-
i) lengthy [email protected]
ii) 3DM13/Heaven/Valley each looped for over 1hrs each.
iii) general gaming.
Where as 540MHz or 540MHz was creating issues.
Then one day whilst running [email protected]
my 1135 (+31mV) / 535 (+0mV) was failing in [email protected]
, I lowered HBM to 525MHz, which still did not solve the "bad state" in GPU slot. Finally only using 500MHz HBM clock solved the issue, I tested then ~ 60hrs [email protected]
with 1135 (+31mV) / 500 (+0mV), to me this meant GPU OC was fine and HBM OC is the issue. This also meant my idea to test HBM clocking steps by stability testing was a flop
Due to all the wasted time testing HBM OC and lack of performance gain from it I thought just tweak lowering MVDDC below stock whilst being at 500MHz. This exercise showed that 1.2V @ 500MHz is not stable for all uses I had for GPU, ~1.263V was. This again was confusing and a setback IMO
. SK Hynix state 1.2V, AMD state 1.3V, I can only conclude from this testing that HBM 1 due to perhaps production variation requires 1.3V.
So after failing to successfully OC HBM with full stability and under volting at stock clock I was no closer to knowing how the steps of HBM clocking were
. Whilst going over past data I came across an AIDA64 GPGPU benchmark which I had done on Fiji, this to me seemed like the EUREKA
.HBM Clock/stepping testing (Click to show)
Basically 100MHz vs 400MHz huge difference, then there is large difference between 400MHz vs 500MHz. 520MHz is clocking at 500MHz. 535MHz is clocking at 545MHz and so did 565MHz. 575MHz is clocking to 600MHz and so did 630MHz (due to time limit ran 1 run, thus no screenie). I increased HBM voltage upto 1.35V to test 666MHz but as soon as bench started I got artifact on screen.
So viewing the data above I can see now why 535MHz and 525MHz was failing stability testing recently, I can only conclude why 545MHz and 540MHz artifact quicker in stability testing is due to possibly HBM clock getting close to correct step thus aspects we're unaware of concerning HBM/MC coming into play.
On the i5+Fury X rig I have been testing for just over 61.5hrs continuously using [email protected]
to get 1135/545 stable. Initially I set HBM voltage stock and over a 22.5hr run GPU entered "bad state" twice and did not lose a work unit. So then I upped HBM voltage by +6.25mV, within 12hrs I had 1x bad state. I then increased HBM voltage to +12.5mV, within 5hrs 1x bad state, so then I went +18.75mV so far passed 22hrs [email protected]
. I'm gonna let this card fold until at least another 9hrs
I hope members will share data on AIDA64 GPGPU benchmark for HBM clocks testing.