AMD Vega Hardware Reviews

I believe this has been discussed before but the high polygon throughput is likely a potential based on culling primitives, not magically being able to generate more polygons than its 4 setup engines is capable of. That test you linked isn't designed to test any kind of discard from what I know.
tdDOS5t.png


The data I posted is specifically for 100% culled primitives. Here is a larger version that includes Vega 64 at Fiji clocks.
 
The data I posted is specifically for 100% culled primitives. Here is a larger version that includes Vega 64 at Fiji clocks.
I see, I missed that part of the screenshot sorry. I thought it was just a polygon throughput test.

Why is Strip so low on Fury? I assume List is the indication that it isn't working.
 
Undervolted (1200->1025mV) and set to +50% power, Gamer's Nexus Vega 56 is drawing 30-70 Watts more than stock.
Oops, I only just realized that the Hardwareluxx article had them leaving the power target at +50% and still getting reduced power consumption compared to stock. I had assumed they had reduced the +% power offset to the minimum necessary to sustain the combination of clocks and undervolting they were using, as the newest driver supposedly made it possible to undervolt and increase the power target only enough to stop throttling behavior, rather than having to set the power target to the full +50%. I was inferring the same would be true of what GN achieved on launch drivers (i.e. if you could reduce the power offset to the minimum necessary to sustain the boost clock at the lower voltage the resulting power usage would be below stock settings).
 
I see, I missed that part of the screenshot sorry. I thought it was just a polygon throughput test.

Why is Strip so low on Fury? I assume List is the indication that it isn't working.
It's a different format, which prior to Polaris AMD's setup pipeline had lower throughput for.

Speaking of Polaris, the same set of tests shows that Vega has about the same throughput at 100% culled lists and strips, and also for 50% culled lists. That last case is one where Polaris does not have the same throughput as the fully culled cases, which hints that at least something is different for Vega.
 
I believe the explanation is that RX64 is suffering much more severely from being front end bottlenecked by its polygon throughput still being on the order of 4 triangles per clock since primitive shaders have not been enabled in drivers yet.
That bottleneck, if the case, would be no different than prior GCN products and readily bypassed with async compute. Primitive shaders likely needing dev intervention to cull more than before. The only automatic gains should be the deferred attributes which will be difficult to attach a number too. Less cache usage on a cache that is likely significantly larger than prior generations based on Scorpio's parameter cache increase. Along with less wasted bandwidth.

As a result the higher power target and CUs of RX64 are simply being wasted. RX56 is proportionally less effected by the front end bottleneck so its showing much better perf/watt even without primitive shaders enabled.
Not wasted, as the device should be dynamically scaling towards the demands. Not doing work will conserve energy and the cards still hit that limit easily enough. The cards would need to be extremely power limited if hitting that 4 triangle limit and power limit while most ALUs are idling from lack of work. As mentioned above, that limitation would be easily worked around with async.

So, primitive shaders have not been enabled in the drivers yet. Also, DSBR is not working for each and every application out there. Also, HBCC is disabled by default, even though it can potentially increase performance by double-digit percentages. Add all of that together (if those are accumulates) and you have your 1080 Ti competitor.
You forgot packed math and intriniscs/SM6 which should stack as well. Now AMD just needs to change Omega to Magic for their drivers and run with the free advertising.
 
So, primitive shaders have not been enabled in the drivers yet. Also, DSBR is not working for each and every application out there. Also, HBCC is disabled by default, even though it can potentially increase performance by double-digit percentages. Add all of that together (if those are accumulates) and you have your 1080 Ti competitor.

If that is acuatlly the case, the whole management of the Radeon Group needs urgent replacement. If you have a powerful piece of hardware and your driver hold it back that much, your organisation, planing and execution are just bad.
 
I was surprised that they released a huge driver update for old products before Vega release. Why they don't put all man power on Vega?
 
If we follow the path of apparently disabled features, what would be the reason to not state this? The only answer I can think of is that you can't release a product based on possible performance increases and they were required to release RX to meet financial expectations, much like the FE.
 
If that is acuatlly the case, the whole management of the Radeon Group needs urgent replacement. If you have a powerful piece of hardware and your driver hold it back that much, your organisation, planing and execution are just bad.
I doubt that is actually the case. I think its more people trying to figure out why vega suck that much(actual performance in games vs raw power vs last gen) when it was so hyped.
 
If that is acuatlly the case
This is not a matter of "if". The B3D Suite results directly show that Vega currently has a culled triangle throughput of 3.75 triangles per clock, which is not compatible with the claims made in the whitepaper about primitive shaders and their effect on culled triangle throughput. Unless someone is going to assert that RTGs claims in the whitepaper are fabrications, one is forced to infer from the B3D Suite results that this feature is not activated in drivers yet.

As for why RTG would throw RX Vega out the front door with the drivers in this state? Who knows. Personally, I would have delayed until the drivers were in better shape than this.
 
Was watching GamersNexus livestream with undervolting Vega 56 again with new drivers. Results are much better with Wattman now, AMD definitely did well to get those issues resolved fast. Dropping almost 100w with negligible performance difference, should have full article and video in a couple of days.
 
Nice grouping by time/age and API. Tempting to project the results to 2018 and DX12.1 or Vk1.1.
All the really bad outliers appear to have MSAA on.....

Like i can see why going forward AMD might not be caring about MSAA performance, but at the same time it kind of makes you want to slap your forehead................
 
This is not a matter of "if". The B3D Suite results directly show that Vega currently has a culled triangle throughput of 3.75 triangles per clock, which is not compatible with the claims made in the whitepaper about primitive shaders and their effect on culled triangle throughput.
Questions in no particular order:
  • Is the B3D suite working properly?
  • Does the driver detect the B3D suite as unsuitable for primitive shading?
  • Has the driver been coded to preserve image quality, at the cost of performance, when in doubt?
  • Is there a bug in the driver when running this test?
  • Is the driver just crap on the entire subject of primitive shaders?
  • Is the performance of Vega in the B3D culling test exactly what it should be with primitive shading active?
  • What else?
Unless someone is going to assert that RTGs claims in the whitepaper are fabrications, one is forced to infer from the B3D Suite results that this feature is not activated in drivers yet.
It's tempting to presume something, but there's so many choices!

As for why RTG would throw RX Vega out the front door with the drivers in this state? Who knows. Personally, I would have delayed until the drivers were in better shape than this.
Drivers, by definition, are always sub-optimal. There's always new code they've not encountered before, and tweaks are required. Compilers can always be improved. etc.

So, while it seems reasonable to assume the driver is indeed crap at extracting the best performance from Vega, the question is really, how close to "working as intended" is the hardware/software combination of Vega. 1 month? 6 months? A year?
 
Questions in no particular order:
  • Is the B3D suite working properly?
  • Does the driver detect the B3D suite as unsuitable for primitive shading?
  • Has the driver been coded to preserve image quality, at the cost of performance, when in doubt?
  • Is there a bug in the driver when running this test?
  • Is the driver just crap on the entire subject of primitive shaders?
  • Is the performance of Vega in the B3D culling test exactly what it should be with primitive shading active?
  • What else?
Fair enough. I still think it is fair to say that the results of this test, when combined with the apparent 100% identical performance between Vega 56 and Vega 64 at the same clocks in both this test and games strongly suggest that primitive shaders are not currently enabled in the RX drivers for any application, not merely the test suite.

Drivers, by definition, are always sub-optimal. There's always new code they've not encountered before, and tweaks are required. Compilers can always be improved. etc.

So, while it seems reasonable to assume the driver is indeed crap at extracting the best performance from Vega, the question is really, how close to "working as intended" is the hardware/software combination of Vega. 1 month? 6 months? A year?
I have no idea, but I certainly hope the answer is sooner rather than later as my curiosity is killing me.
 
Back
Top