AMD Vega Hardware Reviews

Been tired and under the weather all day, so not had energy to do much with my new Vegas. I unpacked them, then plugged one of them into my now ancient Nehalem i7-920 because the chassis is easy to work in (and has requisite power connectors available), just to see what she could do when folding with the adrenaline drivers. Seems like 350k, up to half a million PPD; even a bit more at its peak! :LOL: Even at the low end, that's nearly as much as both my 390Xes typically do together and far more at the top end... :runaway:

Poor Nehalem btw... Ugh. It gets less than 10k PPD when CPU folding (at one point it only hit about 6.6k PPD). My 7900X can do 120k no probs. :p (Even though it runs really stinking hot then, ugh. I really should get it decapped, if I can get someone to do it for me and dope it up with conductonaut...) I suppose lack of more modern AVX implementations is holding it back.

Card fans are very understated too. Now, the fans in that PC are NOT silent, it's an old chassis with old fans so they muffle the video card's sound profile a bit, but even so it's a very quiet card compared to anything else I've owned so far even when folding all-out; the noisiest fan of them all is the 120mm noctua front intake fan which has a really annoying high-pitched whine to it when it runs at full speed. In this poorly ventilated chassis GPU core temps hover around 75-80C and hotspot hit 88C with occasional blips up to 89. Clocks land around 1554MHz after hitting thermal equilibrium, sometimes a little more, up into the mid 1560s or 70s most of the time. Balanced or turbo mode in Wattman made absolutely ziplock diff as far as performance is concerned from what I could tell once card got hot.

I'm entertaining the idea of replacing that Noctua with a more modern 4-pin fan, but not sure if it's actually worth the bother fecking around with this old box. Top fan is an Antech "big boy" 200mm fan which can't even be bought anymore, which attaches using 2 pins only to a hardwired three-speed fan controller built into the chassis, so there's no genuine way to run this box entirely temp controlled anyway. Not without some potentially ugly modding, and the hardware in it is mostly from around 2008, so again, bad value proposition... :p Still, it keeps chugging, so I'm reluctant to let it go.
 
Ok, played a little bit more with these cards. Quite often they actually seem to do about 530-550k PPD. Peak was a hair over 600k on one particular work unit. Some work units are around 450k PPD, so average is well over what my 390s could do as a pair. Switching them down to power saving profile in Wattman knocks about 30MHz and 5 degrees C off core speed and temp respectively, as well as 40-60W of power draw (according to GPU-Z.) 30MHz means the card hovers around 1525MHz or so, IE basically no difference whatsoever in performance, but quite a noticeable reduction in power/temps...

Both cards have quite similar core temp behavior. One is slightly cooler in the hotspot by a degree or two and on average seems to pull a little less volts and power, but it's a marginal difference. HBM sits at about 74-75C on the warmer card and around 73-74 on the other, slightly more frugal card.

It also does over 21k graphics score in 3dmark Firestrike, on an ancient 3.3GHz Nehalem CPU. Not terribly shabby, IMO... Tried bumping HBM clock to 1000; card completed fine but no appreciable increase in performance. I suppose either this PC is too CPU limited, or maybe regular ole Firestrike is simply too undemanding, I dunno. Still, with a better CPU I should see some sort of difference I'd think, since there's quite a lot of heat haze, DoF, smoke effects and whatnot in this bench...
 
Auhum, that's interesting... Time Spy stock clock/"balanced": 5708, and Time Spy with HBM@1000: 5660; 5741

Ok, so diff is literally less than +/- 1%; I dunno what error margin is on this bench but it looks to be inside it. (3 data points is all you ever need! :LOL:) Still, is Time Spy entirely either CPU or memory bandwidth unlimited at stock Vega HBM clock...? Could be this tired old 920 nag of a CPU holding the card back again. I'm going to move them both into my main rig in just a little while to maybe gain a clearer picture. :p

Or, does HBM have some similar type of link CRC/ECC as GDDR5, and running it too fast might actually be slowing it down? I'm not seeing any graphics corruption, so transmission errors aren't overwhelming any error correcting features which might be there. Unlike what happened if I tried to raise GDDR clock on my 390Xes even just an itty-bitty... Screen got garbled pretty bad, sadly.
 
The HBM2 standard provides for ECC, although the DRAM and GPU would need the hardware and firmware present to support it. AMD's mostly mentioned it as an Instinct or other high-end feature, and the enablement of ECC support even then has only started to come out recently.

AMD does have the ability to throttle HBM2 clocks, and there are other measures such as thermally-activated refresh rate increases or stall cycles that would reduce bandwidth without affecting clock. Not knowing the sampling rate or how the GPU is reporting its frequencies, it's possible some of these are kicking in without it being made fully clear.
 
Auhum, that's interesting... Time Spy stock clock/"balanced": 5708, and Time Spy with HBM@1000: 5660; 5741

Ok, so diff is literally less than +/- 1%; I dunno what error margin is on this bench but it looks to be inside it. (3 data points is all you ever need! :LOL:) Still, is Time Spy entirely either CPU or memory bandwidth unlimited at stock Vega HBM clock...? Could be this tired old 920 nag of a CPU holding the card back again. I'm going to move them both into my main rig in just a little while to maybe gain a clearer picture. :p

Or, does HBM have some similar type of link CRC/ECC as GDDR5, and running it too fast might actually be slowing it down? I'm not seeing any graphics corruption, so transmission errors aren't overwhelming any error correcting features which might be there. Unlike what happened if I tried to raise GDDR clock on my 390Xes even just an itty-bitty... Screen got garbled pretty bad, sadly.

It's more to do with HBM2 throttling threshold. If your stacks are at or above 85C, you're loosing a lot of performance, and getting HBM2 on Vega to 85C is easy with memory bandwidth hungry compute workload.
Also, if you find some time, play with undervolting as manual tweaking can get you either good sustained clock gains or lower temperatures and power. For example, my Vega 56 stock would draw 250W mining Etherium at 37MH/s and tweaked with manual clock and voltages I get 43.8MH/s with 140W draw. Can't go lower on power with Vega 64 BIOS mod, but stock 56 BIOS allowed me to go down to 100W as I could lower vGPU to 0.8V instead of 0.925V.

Have fun Grall!
 
Well...that was...interesting!

AMD driver clearly gets really confused by switching out one model card with another. Lost video after first bootup with the vega boards in my new rig, before I had a chance to log in; with no way to recover I had to push reset button. On bootup I got video back again, but windows refused to recognize my login password for some super weird reason I can't figure out... Rebooted AGAIN, was able to log in. Did a "clear" install of 17.12.2 from within the Radeon Settings app. Took for-fucking EVER to complete, not sure why the process was so slow, and lost video again after the main driver was installed, right when it asks if you want to install ReLive. Put computer into sleep with ATX power button, woke it up again, video returned. (Lucky I set power button to sleep instead of shutdown...) Was able to complete driver install even though it complained at some point that VulcanRT.exe or whatever could not be started properly, so I did another restart just to be sure.

Now, to be fair I've had that error on this rig every driver install/update since I first bought it. Not sure why it happens, but vulcan seems to work anyway, so it's just some temporary glitch while installing it appears. Anyhow, AMD's driver/install quality is still somewhat lacking and unstable I would say. Like, every startup I get a bad sound device error in device manager if I have two GPUs slotted in. It's one of the HDMI audio devices which wonks out, gods know why or which one it is; I'm not even using HDMI, I run DVI right now... I was hoping I'd be rid of the problem once and for all after switching over to my Vegas, but no. It's still there.

Not knowing the sampling rate or how the GPU is reporting its frequencies, it's possible some of these are kicking in without it being made fully clear.
Hm, the HBM only hit 75 tops with just one board installed in my oldest PC (uATX chassis, so these 2.5-slot cooler cards won't fit both at the same time), so surely they won't be temp throttling...? :p

Also, if you find some time, play with undervolting as manual tweaking can get you either good sustained clock gains or lower temperatures and power.
Yeah, that was the plan... What's a good crash course on Vega undervolting with Wattman? :D What's like, a good volt target to aim for with stock core clocks? I don't think I'll bother OCing these cards as they thermal throttle so easy anyway, and I have two of them too right up next to each other baking themselves, so not going to make things any easier.

Dropping the cards down to power saving profile put vCore to a tad over 1V for one on these boards when folding and a hair under 1V on the other, most of the time. They sort of wiggle up and down all the time. :D Didn't check volts while running 3DMark.

Have fun Grall!
Thanks! I hope so. :D Even if I get nothing but problems I must say everything LOOKS pretty damn boss installed in this chassis. Just about room enough for everything to fit, pretty crazy...!
 
Well...that was...interesting!

AMD driver clearly gets really confused by switching out one model card with another. Lost video after first bootup with the vega boards in my new rig, before I had a chance to log in; with no way to recover I had to push reset button. On bootup I got video back again, but windows refused to recognize my login password for some super weird reason I can't figure out... Rebooted AGAIN, was able to log in. Did a "clear" install of 17.12.2 from within the Radeon Settings app. Took for-fucking EVER to complete, not sure why the process was so slow, and lost video again after the main driver was installed, right when it asks if you want to install ReLive. Put computer into sleep with ATX power button, woke it up again, video returned. (Lucky I set power button to sleep instead of shutdown...) Was able to complete driver install even though it complained at some point that VulcanRT.exe or whatever could not be started properly, so I did another restart just to be sure.

Now, to be fair I've had that error on this rig every driver install/update since I first bought it. Not sure why it happens, but vulcan seems to work anyway, so it's just some temporary glitch while installing it appears. Anyhow, AMD's driver/install quality is still somewhat lacking and unstable I would say. Like, every startup I get a bad sound device error in device manager if I have two GPUs slotted in. It's one of the HDMI audio devices which wonks out, gods know why or which one it is; I'm not even using HDMI, I run DVI right now... I was hoping I'd be rid of the problem once and for all after switching over to my Vegas, but no. It's still there.


Hm, the HBM only hit 75 tops with just one board installed in my oldest PC (uATX chassis, so these 2.5-slot cooler cards won't fit both at the same time), so surely they won't be temp throttling...? [emoji14]


Yeah, that was the plan... What's a good crash course on Vega undervolting with Wattman? :D What's like, a good volt target to aim for with stock core clocks? I don't think I'll bother OCing these cards as they thermal throttle so easy anyway, and I have two of them too right up next to each other baking themselves, so not going to make things any easier.

Dropping the cards down to power saving profile put vCore to a tad over 1V for one on these boards when folding and a hair under 1V on the other, most of the time. They sort of wiggle up and down all the time. :D Didn't check volts while running 3DMark.


Thanks! I hope so. :D Even if I get nothing but problems I must say everything LOOKS pretty damn boss installed in this chassis. Just about room enough for everything to fit, pretty crazy...!
Pics please...

Sent from my SM-G950U using Tapatalk
 
Maybe a bit too tight.
Technically, well, yes. But I really don't like dead space, that's why I picked this case. So many drivebay-less cases these days are standard cases that have simply had the drivebays removed, leaving a giant void at the front of the chassis. Sometimes there's even mounting spaces for drives to sit flat on the backplate. Jeesh! What a waste!

The Define C chops off all the dead space, leaving a really neat and small chassis. And it comes with dust filters preinstalled, which cuts down on the amount of crud that gets in, but not everything stops at the filter of course...

Downside: it's hell to work in with honkin' big components like I had to go and choose, but everything DOES fit, I checked in advance to make sure of that. :)

One of the vegas run up to 85C core, but that's cooler than the R390s could hit in the same chassis (would sometimes start to throttle at 94C).

The vegas dial down to about 1350MHz while running compute load (Folding@Home) and stay there. Haven't tried to do any undervolting yet, so stock volts.
 
One of the vegas run up to 85C core, but that's cooler than the R390s could hit in the same chassis (would sometimes start to throttle at 94C).

The vegas dial down to about 1350MHz while running compute load (Folding@Home) and stay there. Haven't tried to do any undervolting yet, so stock volts.

Dead space is not all that bad, heat/temps wise. Especially with 2 high-end graphic cards installed.
Are NF A14s worth it's money (30 euros here), and how long is their cable?

I'd refrain from undervolting at first and run everything stock/default until I get the feel for the entire system. But that's me...
Congrats on the new rig :)
 
Dead space is not all that bad, heat/temps wise. Especially with 2 high-end graphic cards installed.
Yes, I know. I still don't like it though. My ideal PC case would be something like a pre-trashcan Mac Pro, with cross-airflow front to back throughout the entire case, heatsinks, fans and so on all matching the case, fitting like a glove inside it nice and efficiently. Impossible to do in the PC world of course because the standards are too lax (no way to integrate everything with that level of precision), but that is what I would desire anyhow.

The trashcan Mac Pro is an incredible design in of itself, but it's 100% totally proprietary, so not an ideal design for a computer. Even the twin GPUs in it are each a unique design PCB. Plus the cooling capacity is limited.

Are NF A14s worth it's money (30 euros here), and how long is their cable?
Worth it, well I dunno. It depends on how crazy you are I guess. :D Noctuas are very high quality and run relatively quiet compared to their capacity, plus these black ones look friggin boss if you ask me, so yeah, I'd say yes. I wish I hadn't bought two sets of them though, because now I have a surplus of fans... :p

The cables for the standard pig&chocolate color noctuas are very nicely sized, you get a black little puppydog's tail of a cord on the fan which usually is enough to reach a motherboard header (maybe 6cm or so?), plus an additional black extender in the box which is like 15cm. The black fans do away with the tail on the fans, they have just a little stump which is like 1.5-2cm or so, and then the black extender. I assume this is for color-matching OCD freaks, because you can now buy color-matched extenders. It means you usually have a bit more cable than you really need to reach the mobo header which can be a bit of a boggle, but I solved it sufficiently.

I'd refrain from undervolting at first and run everything stock/default until I get the feel for the entire system. But that's me...
Congrats on the new rig :)
Yeah, I'm not going to go and undervolt and overclock first thing I do, that's for sure. :p

And thansk! :D
 
So I just grabbed a Vega FE for $749 from NewEgg's eBay store. That was the best price I've seen the card in a long time, and it should do my video editing and mechanical design duties justice for some time to come.

Sent from my SM-G950U using Tapatalk

Tip: If you want to do super fast raytracing renders on CAD software, download the Radeon ProRender plugin. It's really fast!
 
It's certainly a good showing for the Vega 56. I half expected an article that only had results for 1070ti and "sorry we couldn't get a Vega 56"
 
Back
Top