GPU VEGA NANO

[grumpyrobin]

My comments on power consumption are just based on what allegedly good sources are tweeting about the RX, not from the FE.
Performance, I am not expecting anything good or bad, I was just hoping that VEGA would be able to bring down the price on nvidia cards.
But at the moment it looks like that zotac 1080ti mini is going to stay in the same price range even after vega.

 

[grumpyrobin]

what would make me strongly consider vega at this point are SFF features from founders cards.

 

[Phuncz]

True, but the Vega FE is closes to a Titan than a Quadro.

Sure, it does look a lot like a $1,000 RX Vega which compared to Nvidia would lead you to a Titan card, but Raja Koduri warned about a month ago we shouldn't be looking at the Vega FE's performance for an indication on RX Vega [for gaming performance]. I've never heard Nvidia say the Titan isn't meant for gaming, quite the opposite. Maybe it's just the drivers that differ, I wouldn't be surprised. Though we all know what a difference a driver can make.

 

[K888D]

My guess would be that Vega will offer very competitive fps per $, but at the expense of lower fps per watt.

 

[Phuncz]

Someone on Reddit claims it's performing like an overclocked Fury because it's running like a Fury, due to Tile-Based Rendering not yet being enabled in the Vega FE:
https://www.reddit.com/r/Amd/comments/6kdwea/vega_fe_doesnt_seem_to_be_doing_tiled/djld98t/

I hope we'll find out soon either way.

 

[EdZ]

Tiled rasterisation improves memory utilisation at very large rendering resolutions making a GPU more efficient (mostly due to less hammering of the external bus), rather than directly improving performance (though using less power can allow you to squeeze out some performance headroom if you're power limited). If anything it is a slight performance hit clock-for-clock, as some work needs to be 're-done' per-tile where with Intermediate Mode rasterisation it's done once per frame (plus the overhead of determining which geometry belongs in which time to start with).
That AMD have supposedly implemented two different rendering paths for Vega is... odd. It's a bit of a waste of die area, and Maxwell managed to make the translation to tiled rasterisation completely transparent to legacy applications (hence why nobody even noticed it for 2 years).

 

[Therandomness]

PCPer have some benchmarks of the Vega FE, as well as photos of the board.
-Images snipped-
Looks like there's enough 'wasted' space to just about cram into a slightly shorter PCB, but the VRMs required to power Vega's 270W-300W hunger are going to take up more room than the R9 Nanos (and we all know the power spike penalty the Nano paid for reduced phases). That high power consumption nets performance as expected, usually around the 1070, sometimes between the 1070 and 1080, so dropping power to move to a smaller PCB is likely to net performance not far off of the R9 Nano's.

Actually, now that Buildzoid's VRM analysis is out.. that 12 phase core VRM is almost completely unnecessary, it could've easily been a 6 phase and there wouldn't have been any issues. Also, HBM2 only requires 12-24W of power most of the time (according to its single phase VRM), so that could probably be shoved somewhere small on the board of a Vega Nano card.

 

[EdZ]

Actually, now that Buildzoid's VRM analysis is out.. that 12 phase core VRM is almost completely unnecessary, it could've easily been a 6 phase and there wouldn't have been any issues. Also, HBM2 only requires 12-24W of power most of the time (according to its single phase VRM), so that could probably be shoved somewhere small on the board of a Vega Nano card.

As we saw with the R9 Nano, cutting the number of phases doesn't prevent the card from running, but it does expose it's 'spikey' power draw to the PSU to handle. The R9 Nano was bad enough at 180W average spiking past 400W, I'd hate to think what instantaneous draw a 300W average Vega could hit!

 

[Therandomness]

As we saw with the R9 Nano, cutting the number of phases doesn't prevent the card from running, but it does expose it's 'spikey' power draw to the PSU to handle. The R9 Nano was bad enough at 180W average spiking past 400W, I'd hate to think what instantaneous draw a 300W average Vega could hit!

But that was the Nano with a 4 phase VRM, which was quite inadequate. If they had put the chokes on the back or something, they could've easily fit 6 phases onto the card, as well as 2 for the memory. That exact configuration was enough to support the 290/X quite easily, which had a very similar TDP/power draw, so I wouldn't be surprised if RX Vega (the gaming one) had a 6 phase.

 

[EdZ]

But that was the Nano with a 4 phase VRM, which was quite inadequate. If they had put the chokes on the back or something, they could've easily fit 6 phases onto the card, as well as 2 for the memory. That exact configuration was enough to support the 290/X quite easily, which had a very similar TDP/power draw, so I wouldn't be surprised if RX Vega (the gaming one) had a 6 phase.

Cramming double the number of phases (4 vs. 6 + 2) is going to increase the size of the card beyond the R9 Nano and closer to non-HBM ITX cards (even ignoring HSF size). Putting large components on the rear also runs into problems with cooling, and violating keep-out for the slot above.

 

[Therandomness]

Cramming double the number of phases (4 vs. 6 + 2) is going to increase the size of the card beyond the R9 Nano and closer to non-HBM ITX cards (even ignoring HSF size). Putting large components on the rear also runs into problems with cooling, and violating keep-out for the slot above.

I guess my expectations when it comes to component sizes and PCB design are a little too high.. Ah well. How about a 5+1? I mean, if they extended out the PCB just a little bit they could attempt 4+2 or 5+1 or 2, though they'd have to use inductors in the slim/tall package, and all the heat would be concentrated in a tiny area. They'd have to pull out an amazing heatsink design if they managed a VRM design like that.

 

[Boil]

...an amazing heatsink design if they managed a VRM design like that.

That is where water cooling (custom loop with CPU) comes in...!

 

[grumpyrobin]

My rule of thumb is to spend the cash needed for water cooling into a better gpu.
If there is nowhere higher to go, then I would go for water cooling.

 

[SeñorDonut]

PCPer have some benchmarks of the Vega FE, as well as photos of the board.

Looks like there's enough 'wasted' space to just about cram into a slightly shorter PCB, but the VRMs required to power Vega's 270W-300W hunger are going to take up more room than the R9 Nanos (and we all know the power spike penalty the Nano paid for reduced phases). That high power consumption nets performance as expected, usually around the 1070, sometimes between the 1070 and 1080, so dropping power to move to a smaller PCB is likely to net performance not far off of the R9 Nano's.

I died a little inside when I saw those pics

 

[grumpyrobin]

I died a little inside when I saw those pics

Don't radeon PCBs always have a huge blank left open for the fan?

 

[Phuncz]

Not really, it's only possible in the higher end models because of the HBM2 and the reference R9 Fury and Nano (the only other HBM cards) all have short PCBs. Other cards have the power delivery circuitry in that area.

 

[Maestria]

http://www.tweaktown.com/news/58387/amd-radeon-rx-vega-xtx-xt-xl-teased/index.html

"there will be no Nano-like RX Vega card"

:'(

 

[grumpyrobin]

 

[br3nd0]

http://www.tweaktown.com/news/58387/amd-radeon-rx-vega-xtx-xt-xl-teased/index.html

"there will be no Nano-like RX Vega card"

:'(

Poor form.

 

[iFreilicht]

SAD!