Rise of the Tomb Raider VRAM Testing

[Bacon1]

http://www.hardocp.com/article/2016/02/29/rise_tomb_raider_graphics_features_performance

Socker, they were completely wrong about the VRAM being the bottleneck during their SLI/CFX Testing. Fury holds strong with its 4GB of HBM even at the crazy settings.

From their previous review:

One of the main reasons the Fury X didn't do as well at high settings is VRAM capacity. We found that the AMD Radeon R9 Fury X would stutter some at higher settings. The limitations of its 4GB of VRAM became a problem at higher settings. When VRAM wasn't a bottleneck, as you can see, the Fury X can perform faster than a GTX 980 Ti at 4K, but only when there isn't a VRAM capacity bottleneck.

From their actual VRAM testing review:

Even though the AMD Radeon R9 Fury X has the least amount of dedicated VRAM here it too does not differ much in performance using "Very High" textures. It seems that dynamic VRAM usage is not harming performance here.

While this game may use a lot of VRAM with "Very High" textures, it doesn't suffer in performance because of it.

In regards to the AMD Radeon R9 Fury X we know that there is only 4GB of HBM onboard the video card itself, that is its dedicated VRAM. However, it appears that with the AMD Radeon R9 Fury X the dynamic VRAM is working in a big way to offset the difference. We see a large 1.8GB of RAM is being used through dynamic VRAM at "Very High" textures. When we drop down to "High" textures this drops off down to only 198MB, about the same as the R9 390X. This means those "Very High" textures are adding greatly to the Fury X's dynamic VRAM usage, but the dedicated VRAM is of course pegged at its maximum capacity at all times except for "High" textures.

In regards to the NVIDIA GeForce GTX TITAN X, which has 12GB of VRAM onboard it consumes a whopping 7GB of VRAM at "Very High" textures. This means the GeForce GTX 980 Ti may not only be reaching its maximum VRAM usage but may be bottlenecked with 6GB of VRAM onboard when running "Very High" textures in this game.

Ironically its the 6GB 980 TI that is bottlenecked!

 

[3DVagabond]

In regards to the AMD Radeon R9 Fury X we know that there is only 4GB of HBM onboard the video card itself, that is its dedicated VRAM. However, it appears that with the AMD Radeon R9 Fury X the dynamic VRAM is working in a big way to offset the difference. We see a large 1.8GB of RAM is being used through dynamic VRAM at "Very High" textures. When we drop down to "High" textures this drops off down to only 198MB, about the same as the R9 390X. This means those "Very High" textures are adding greatly to the Fury X's dynamic VRAM usage, but the dedicated VRAM is of course pegged at its maximum capacity at all times except for "High" textures.

In regards to the NVIDIA GeForce GTX TITAN X, which has 12GB of VRAM onboard it consumes a whopping 7GB of VRAM at "Very High" textures. This means the GeForce GTX 980 Ti may not only be reaching its maximum VRAM usage but may be bottlenecked with 6GB of VRAM onboard when running "Very High" textures in this game.

Ironically its the 6GB 980 TI that is bottlenecked!

More of Kyle's guesswork.

 

[littleg]

[H] lost any credibility they may have had a long time ago.

 

[Leadbox]

[H] lost any credibility they may have had a long time ago.

Only thing H has going for them is their benchmark runs are long, it gives the fps #s more cred. The analysis, conclusions and recommendations have always been limited at best.

 

[Mahigan]

Yep, I said as much in the Ashes thread. He had ZERO evidence that the VRAM was holding Fury back. Now a bunch of folks began parading this news across the web as though it were a matter of fact.

 

[Head1985]

Tomb raider runs great with 2Gb vRam
https://www.youtube.com/watch?v=6fxr8qnSg0o

 

[Mahigan]

The Fury series uses memory differently. So even if a game uses more than 4GB, as Rise of the Tomb Raider does, it won't impact performance. The 4GB of HBM is used as a cache for very fast local data access and the system ram is used for things which don't require a large amount of bandwidth.

The end result is that 4GB is enough for 4K.

 

[thilanliyan]

The Fury series uses memory differently. So even if a game uses more than 4GB, as Rise of the Tomb Raider does, it won't impact performance. The 4GB of HBM is used as a cache for very fast local data access and the system ram is used for things which don't require a large amount of bandwidth.

The end result is that 4GB is enough for 4K.

Is this fact, or your educated guess?

 

[96Firebird]

Is this fact, or your educated guess?

It is completely wrong.

 

[Mahigan]

Fact, it's in the HardOCP review under VRAM usage:
http://m.hardocp.com/article/2016/0...graphics_features_performance/14#.VtWfhOTS9oM

 

[96Firebird]

Per game basis by AMD.

Which I thought was frowned upon?

 

[Mahigan]

Per game basis by AMD.

Which I thought was frowned upon?

Look, goal post shifting...

Anyway, AMD had stated this prior to the Fury Launch.

“If you actually look at frame buffers and how efficient they are and how efficient the drivers are at managing capacities across the resolutions, you’ll find that there’s a lot that can be done.*We do not see 4GB as a limitation that would cause performance bottlenecks. We just need to do a better job managing the capacities. We were getting free capacity, because with [GDDR5]*in order to get more bandwidth we needed to make the memory system wider, so the capacities were increasing. As engineers, we always focus on where the bottleneck is. If you’re getting capacity, you don’t put as much effort into better utilising that capacity.*4GB is more than sufficient. We’ve had to go do a little bit of investment in order to better utilise the frame buffer, but we’re not really seeing a frame buffer capacity [problem]. You’ll be blown away by how much [capacity] is wasted.

http://wccftech.com/amd-addresses-capacity-limitation-concern-hbm/#ixzz41ey6H9fZ

People ignored what they said and kept the myth going. Now that we see it is true, people are shifting the goal post further.

 

[Dribble]

I take it "dynamic ram" is short for "doesn't fit so we've moved it to main memory"? In which case in the example it's using almost 2GB of dynamic ram Fury would be faster if it had 6GB or ram.
It's interesting that the 390's only use 4GB, looks like they've optimised the driver for 4GB to suit fury. I bet there's a few more % performance available to the 390 if they optimised for it's larger 8GB buffer.

 

[thilanliyan]

Fact, it's in the HardOCP review under VRAM usage:
http://m.hardocp.com/article/2016/0...graphics_features_performance/14#.VtWfhOTS9oM

Thank you for the link.

 

[Mahigan]

I take it "dynamic ram" is short for "doesn't fit so we've moved it to main memory"? In which case in the example it's using almost 2GB of dynamic ram Fury would be faster if it had 6GB or ram.
It's interesting that the 390's only use 4GB, looks like they've optimised the driver for 4GB to suit fury. I bet there's a few more % performance available to the 390 if they optimised for it's larger 8GB buffer.

There's 0 performance advantage going from 4GB HBM with these optimizations and 8GB HBM without. If there was we'd see a large performance hit going from high to very high textures.

Have a look:

Spoiler

 

[bystander36]

Ironically its the 6GB 980 TI that is bottlenecked!

How did you come to that conclusion? This is from your own quote:

One of the main reasons the Fury X didn't do as well at high settings is VRAM capacity. We found that the AMD Radeon R9 Fury X would stutter some at higher settings. The limitations of its 4GB of VRAM became a problem at higher settings. When VRAM wasn't a bottleneck, as you can see, the Fury X can perform faster than a GTX 980 Ti at 4K, but only when there isn't a VRAM capacity bottleneck.

I think you are misunderstanding the results of the Titan X that showed it used up to 7Gb, but that does not mean that it needs 7Gb for smooth results.

 

[xthetenth]

^I'm pretty sure he's mocking Kyle.

I wonder how they're doing the arrangement of data. I'm figuring that since caching isn't exactly a new or difficult technology it's heuristics based rather than per game. It'd be interesting to test with older versions of RoTR and drivers to be sure of it though.

 

[bystander36]

I take it "dynamic ram" is short for "doesn't fit so we've moved it to main memory"? In which case in the example it's using almost 2GB of dynamic ram Fury would be faster if it had 6GB or ram.
It's interesting that the 390's only use 4GB, looks like they've optimised the driver for 4GB to suit fury. I bet there's a few more % performance available to the 390 if they optimised for it's larger 8GB buffer.

Based on the first quote of the OP, you can see that it may not affect performance, but it does affect smoothness.

 

[Osjur]

The thing is, It did stutter on Fury X and Very High textures when the game first came out, at least on 1600p. I had to lower the texture quality to High to get rid of that stutter. I need to retest this at some point as I finished the game before the first official patch came out.

 

[crisium]

Only thing H has going for them is their benchmark runs are long, it gives the fps #s more cred. The analysis, conclusions and recommendations have always been limited at best.

Agreed. I mostly like the way they benchmark, but they too often spew garbage when they type their thoughts.

 

[thesmokingman]

[H] lost any credibility they may have had a long time ago.

Like when they ran the 580 vs 7970 triple gpu test? Back then they ran three 580s vs a 7990 + 7970. The AMD combo was able to fit neatly into their platform's two pciex limit whereas with the triple 580, one 580 would have to use a dreaded slot off the chipset. Obviously this tanks performance and the 580 tri-sli suffered, but they were too clueless to realize the audacity of their mistake.

Oh [H]...

 

[thilanliyan]

How did you come to that conclusion? This is from your own quote:

Isn't that from the 1st review? Which is then contradicted by the results from the 2nd review?

 

[thesmokingman]

lol, anyone remember the Physx mod for AMD gpus? The mod is a hack which by definition will get flagged by antivirus, so to get it to work you have to whitelist the mod. [H] did not know this... like how the heck do you not??? The mod had been around for years before they got around to writing about it. Then some of their green members got AV flags, and they knee jerked. They ended up banning/kick/limiting members who pointed their flaw out, then they banned ngohq.com because they thought since the dev came from there, then they must all be haxors! All to protect Kyle's readership right? Oh [H]...

 

[Despoiler]

Is this fact, or your educated guess?

It is completely wrong.

It is a fact. I correctly speculated how AMD's memory management worked before the cards released. Not that it was hard to guess given how memory management works in DX by default for all cards.

http://forums.anandtech.com/showthread.php?p=37530110&highlight=#post37530110


Why is anyone still posting anything from Kyle / [H]? Kyle has proven so many times over that he has no clue what he is talking about. Beyond that he is nothing more than a troll as seen by his nonsensical forum posts.

 

[3DVagabond]

I take it "dynamic ram" is short for "doesn't fit so we've moved it to main memory"? In which case in the example it's using almost 2GB of dynamic ram Fury would be faster if it had 6GB or ram.
It's interesting that the 390's only use 4GB, looks like they've optimised the driver for 4GB to suit fury. I bet there's a few more % performance available to the 390 if they optimised for it's larger 8GB buffer.

Where did you get all that from? And if you just made it up, just say so.