AMD Vega Hardware Reviews

It still reminds me of R600, the R600 won in selected games even against the GTX8800 and people believed that a driver would unlock the full potential of the card. And while the miserable FSAA performance was improved by the driver, the overall performance never came close to the competition.

 

I'd say it was exactly opposite situation. Performance without MSAA was comparable to GTS, later to GTX, but with MSAA it was always poor. It was never fixed in drivers, because it was caused by HW bug (HW MSAA resolve had to be disabled to provide stable operation).

 

Sounds like a good test for hbcc. I skimmed tweaktown but didn't see any mention of HBCC, did I miss something? Is it still ineffective and/or in development?

 

As far as I'm aware, no existing game uses 8K-specific assets, so I don't think there would be increased memory capacity requirements at that definition (for now).

 

HBCC is not activated by default in the drives, other than that it's good to go. Of course, you will have more and better defaults for different access types, maybe even application profiles when drivers mature.

 

It works when enabled, but AMD is still messing with page sizes and levels to make it more efficient.

 

Won't the ridiculous increase in resolution lead to a ridiculous increase in memory consumption for shadowmaps, effectively putting into question if the 8GB are indeed sufficient for this case scenario?

I remember reading somewhere that using a 8K monitor with a current game would be a great test subject for HBCC.
It's a bit of a let down that tweaktown didn't touch that subject.

 

Hmm, I don't know. @sebbbi ?

 

Has AMD explained why Vega needs such a high idle clock speed?

The four prior generations of GCN were all happy enough at 300mhz but Vega is clocked at 852mhz, could this be because of Infinity Fabric somehow?

 

Per AMD, the fabric has its own dedicated clock domain. The GPU should theoretically have more flexibility in varying from it.

 

Curious... Still, the card only draws a few watts when idle, so it doesn't seem to matter much.

 

There's a "by 32" clock divider, so the potential idle clocks are - on the contrary - extremely low. It seems not to be enabled in all available drivers though.



That's with 17.8.2 Optional, I set the fan rpm manually to a higher level.

 

But the available memory pool would have schrunken using larger rendertargets, adding higher pressure on the resource coherency managed by the runtime (swap in/out).

 

As someone who owns the card(56) and has had a few hours to play with it i dont see what all the negativity is about.

1. it over clocks really quite well, i just wish i could go higher then 50% power
2. if you want power efficiency you loose very little performance just by moving slider to power save
3. if your happy to play with undervolting it both decreases power and increases performance and is really really easy.

its biggest issue is that its 1 year later then pascal, it will be interesting to see what 1 year of driver development brings......

 

Have you tried to flash it?

 

I'm getting 28MHz idle core clock on my Vega 64:

 

Using latest drivers for RX Vega and GTX 1080.

 

Not yet, i like to be methodical and i am no where near ready for that.

 

28Mhz looks like a another iteration of the clockbug. How many CPU and GPUs are able to clock that low?

 

With Vega there is a new deep-sleep-state:
https://www.techpowerup.com/reviews...cture_Technical_Overview/images/slides-46.jpg

Computerbase (German) writes that doing nothing yields in ~27 Mhz, moving a window ~50 Mhz and surfing at complex websites ~100 Mhz.
The HBM2 will be downclocked to 167 Mhz, if there is a certain amount of load it will jump to 500 Mhz.
https://www.computerbase.de/2017-08...st/#abschnitt_10_vorteile_der_vegaarchitektur