[DX12] Fable Legends Beta Benchmarks

[.vodka]

Representative or not, it's another data point where we see GCN shoot up to Maxwell's DX11 performance levels at worst, beating it at best. Fiji as noted isn't benefiting as much, but else the gains are amazing. Hawaii steals the thunder, Tahiti puts an impressive showing for what is nearly a 4 year old chip. What a run. Fiji is, sadly, unbalanced, and too workload dependent to put out performance in accordance to its paper specs.

AMD's DX11 performance is still a sore spot in some games, but well, we're fortunately leaving that behind.

You know Hawaii cards in their 2xx/3xx flavors are the 8800GT of the 28nm generation when you see the performance they put out while costing so little (at least in the US, but many other countries would be around the same vs 970/980 prices). Just perfect for 1080p and maybe higher resolutions.

If this ends up being the DX12 landscape until next gen cards arrive, it's win win for everyone, AMD could start regaining market share back (to be frank, there isn't a single nV card that's attractive or priced adequately relative to its performance apart from the 980Ti if the trend continues for the upcoming DX12 games, and no, a 50w power difference isn't a matter of life or death) while nV gets stronger competition and could start to get its act together for Pascal (hoping for no bizarre vram setup like the 970, etc. usual nV shenaningans, you know what I mean).


I like it, this is a clean slate for both sides of the fence and much needed for the market as a whole. Still, I haven't heard anything about nV having an optimized driver for this benchmark/game, so that could change things. AMD's could bring more performance to the table.

 

[Azix]

It's representative in a sense. For pure graphics sure. I said that because it misses a major feature of dx12. Neither ashes nor fable legends have shown us what asynchronous compute can really do. All the wonders are left for consoles.

These performance numbers are going to die when people actually start playing in the world but works for graphics as far as relative performance goes.

 

[RussianSensation]

Fiji is front-end bottleneck (refer to original Fury X review at Computerbase.de where they talked to AMD directly about design challenges & tradeoffs, AMD is fully aware of the potential bottleneck), same setup for many more shaders. AMD do need to optimize it better to extract peak performance.

Thanks, I'll check it out via Google translate.

Glad I kept the 2 R9 290s in CF for rig 2 below!

Whoever got your Sapphire Tri-X 290 for sub-$200 got an absolute steal of a deal.

2013 Full Hawaii chip is the elephant in the room.

Even R9 290/390 are doing amazing for their price. I would have liked to see this benchmark allow for 1440P scores because 4K performance is too slow for everything.

Seems that the tables turned so we may see Nvidia eeking out AMD at the higher resolutions now, while AMD winning at the lower resolutions. Its mostly about the architecture at this point.

Might only be true for GTX980Ti vs. Fury X. In all other pricing segments, NV lost so I don't see any of this generation Maxwell cards having any serious advantage at high resolutions in the sub-$550 segments. Maybe for Pascal vs. Arctic Islands?

Fiji at this point is ROP limited which proves it because the memory bandwidth and shaders are not helping. Hawaii is the winner in performance gains.

R9 290X/390X are showing great scaling against the HD7970 but the scaling falls off a cliff once we get to Fury/Fury X and the scaling is especially poor at 4K where theoretically Fury/Fury X should have the greatest lead over 290X/390X.

AMD should needs to go to 96-128 ROPs for next generation or figure out some way to improve efficiency of those ROPs.

It's representative in a sense. For pure graphics sure. I said that because it misses a major feature of dx12. Neither ashes nor fable legends have shown us what asynchronous compute can really do. All the wonders are left for consoles.

In theory heavier use of AC shaders should only improve performance of GCN cards. If Fable Legends and Ashes are not the poster child examples for ACE utilization, then this is still an amazing result for AMD's DX12 as a whole.

 

[antihelten]

I believe the results from Extremetech have been posted before in this thread, but it's probably worth mentioning that their tests was apparently done with the new beta drivers from AMD, and they basically have the Fury X and the 980 Ti running neck and neck at 4K and 1080P (980Ti wins at 720P if anyone feels like running at that resolution on such a card):

 

[Osjur]

Looking at how original 7970 just destroys 960 in this Nvidia sponsored Unreal engine DX12 benchmark, we can soon give this card the same legendary status as 9700Pro and 8800GTX have. And those original 7970 cards OC'd like beasts: from 925mhz to almost 1.3ghz on air with golden samples. Mine was game stable at 1320mhz on water and I did benchmark runs with almost 1.4ghz clocks.

Card which was released almost 5 years ago still manages to get acceptable framerates at 1080p and seems to just get better with DX12.

And looking at the 290 results, it seems they can also find themselves in that legendary GPU list at some point if the trend with DX12 goes the way it has been going now.

 

[Zodiark1593]

Damn, impressive results from AMD, especially on the 7970. While Nvidia goes for efficiency (utilization of available resources) and small die sizes (especially in Maxwell), AMD favors throwing as many compute resources in as possible, and improve utilization later through drivers. DX12 in particular seems to focus on maximizing utilization of the card's potential. So AMD's approach, while iffy due to die size and power use, may finally be paying off with DX12.

If Nvidia is utilizing their Maxwell chips to the fullest (which they probably are as that was the focus of Maxwell to begin with), anything below the 980 TI it's gg. We'll see what happens here.

What would be interesting is if AMD carries out it's shader threading on the cpu (then handing it all off to the gpu cores) instead of threading the shader cores in hardware as Nvidia has been doing. This would mean GCN is very highly dependent on the underlying software support to utilize it's potential while Kepler nor Maxwell can alter theit threading characteristics quite as freely through software. This hypothetical would explain how the 7970 kept for so long (new drivers allow higher % utilization), and how the 680 fell so far behind without including conspiracy theories.

 

[hawtdawg]

good. dual 980ti's should probably last me until big Pascal.

 

[zlatan]

So you are saying the bench is using async compute, and its unoptimized for GCN. But what about Maxwell? Did NV finally added async compute support in their drivers? Last time we heard any on this, apparently the feature wasn't ready and not enabled yet.

Edit: Got some more info, Joel Hruska at Extremetech said that Lionhead Studios did inform him that the benchmark has a low amount of async compute usage, within each frame is 5% worth of compute that could be performed asynchronously. Given that Zlatan says the AC is 30% slower than normal on the PC build vs Xbone (compiler issue needs driver optimization), that 5% is therefore about 3% performance gain that GCN can flex over anything that doesn't support AC. Which is so small, it's pointless in the overall performance.

The async compute & DMA is ready and enabled in the AMD/NV drivers - if the hardware support it.

I have tested the benchmark on a Fury X:
1080p async on: 105,8 fps
1080p async off: 87,8 fps

I don't have Maxwell so I can't provide data on it.

 

[happy medium]

anyone have a release date for this game?

 

[Gundark]

Hmm. I wonder how would Pitcairn do in 1080p. Mid to high in 30 fps possible?

Edit: Oh, it seems it would be possible even on high settings.

 

[Spjut]

PCgameshardware.de has tested more cards since yesterday. The 780 Ti and 970 are neck and neck.

 

[AtenRa]

PCgameshardware.de has tested more cards since yesterday. The 780 Ti and 970 are neck and neck.

nice thanks,

R9 290 is 13% faster than GTX 970

GTX 980 is only 3% faster than R9 290.

R9 380 is 20% faster than GTX 960

R7 260X 13% faster than GTX 750Ti

 

[antihelten]

nice thanks,

R9 290 is 13% faster than GTX 970

GTX 980 is only 3% faster than R9 290.

R9 380 is 20% faster than GTX 960

R7 260X 13% faster than GTX 750Ti

And with this being at 1080P, it would appear that AMD GPUs (with the slight exception of the bottlenecked Fiji based ones), have gained some 20-25% on Nvidia GPUs, relative to how you would normally expect them to line up (based on experience from DX11 games).

Pretty impressive, although I would hope AMD can figure out a way around the Fiji bottleneck (the new driver does seem like it might help a bit with the scaling, but it's still quite poor).

 

[guskline]

Good showing by AMD. Congrats. The big question is whether or not AMD's PR team can translate it into improved sales. I'll buy the game to see how it performs on both of my rigs below.

 

[parvadomus]

At this point idk if Fiji has a bottleneck or its just the effect of severe throttling to stay within its configured TDP, the difference with Hawaii is too small.

 

[Techhog]

At this point idk if Fiji has a bottleneck or its just the effect of severe throttling to stay within its configured TDP, the difference with Hawaii is too small.

If it were throttling, I'm sure that it would have been reported by someone by now. That can easily be ruled out.

 

[arandomguy]

If you look at Techreport and Pcper data sets scaling seems to be relatively proportionate among Maxwell, CN 1/1.1 chips, and the GCN 1.2 chips as groups. This breaks down when you compare GCN 1/1.1 and GCN 1.2 chips. Tonga and Fiji scaling seems to be there.

Looking at the individual timings in the Pcper article compare the times of GCN 1/1.1 and GCN 1.2. The faster GCN 1.2 cards are slower in some of the specific categories compared to a slower GCN 1/1.1 card.

What is interesting about this as well is we know Mantle behaves differently with GCN 1.2 and GCN 1.0/1.1.

Is there an article that features both Tahiti and Tonga?

 

[parvadomus]

If it were throttling, I'm sure that it would have been reported by someone by now. That can easily be ruled out.

Showing a constant 1050mhz clockspeed does not mean it does not throttle/has aggresive power gating at a lower level, that can add latencies inside the chip itself.

 

[railven]

Is it conclusive that the Fiji need a new driver to really open up? Seeing Hawaii/Grenada that close to Fiji is a little...well...

That isn't how you sell a flagship.

EDIT: Just saw post #88. Woof, all them other sites need to rebench with those new drivers, like immediately.

 

[Headfoot]

Hawaii still showing legs. Anyone who snagged the original 290 @$400 in 2013 has seriously gotten a smoking deal. In 2013 you could still mine litecoins and make good money on that 290 and pay it off (what I did), and now its getting another boost from DX12. Truly today's 8800 GT. I wonder which side will have the next one on 14/16nm...

 

[Azix]

The async compute & DMA is ready and enabled in the AMD/NV drivers - if the hardware support it.

I have tested the benchmark on a Fury X:
1080p async on: 105,8 fps
1080p async off: 87,8 fps

I don't have Maxwell so I can't provide data on it.

which benchmark is this? Sites have said there are no configurable options

was doing some checking and saw that the gigabyte 980 that techreport used cost $629 back in the day. boy oh boy.

 

[zlatan]

which benchmark is this? Sites have said there are no configurable options

Fable Legends bench.
The press may not get the same version.

 

[tential]

I think you shouldn't test oc cards vs stock cards in some situation certainly.

The 980ti cards are 670 usually and the fury x is 650 or higher as well. So I really see no issue using an aib 980ti card that's 20 over the msrp. It's what the vast majority of people care about.

Using a 980ti reference card that very very few people are interested in, in a test means nothing to me. Show me the $670 980ti that are vastly superior that's what matters to me. The only reason people are whining is because it makes the 980ti look great.

Is it unfair if people don't use amd cards that are aib models when it makes sense? Sure... But we should tell them to make their tests better. Not to make other people's tests worse because we're mad at some "slight" from years prior....

 

[railven]

I think you shouldn't test oc cards vs stock cards in some situation certainly.

The 980ti cards are 670 usually and the fury x is 650 or higher as well. So I really see no issue using an aib 980ti card that's 20 over the msrp. It's what the vast majority of people care about.

Using a 980ti reference card that very very few people are interested in, in a test means nothing to me. Show me the $670 980ti that are vastly superior that's what matters to me. The only reason people are whining is because it makes the 980ti look great.

Is it unfair if people don't use amd cards that are aib models when it makes sense? Sure... But we should tell them to make their tests better. Not to make other people's tests worse because we're mad at some "slight" from years prior....

Bingo. But these same people won't say anything when an OC'ed Fury is used in a review. For me, data is data. As long as all the information is made available there shouldn't be an issue. I remember when they through the 460 FTW into the HD 6850/6870 review. It was a dirty move, but it was representative of what people could buy.

If anything with today's GPU Boosts, we don't know what these cards are boosting to at all times. We just get a rough idea. For example, the card used in one of the reviews (think it was Techspot) they openly said the card boosts almost 50mhz higher then what was rated on the box. That's good information to disclose. Other wise we'd all be thinking it's doing that performance at the box boost clocks.

 

[thesmokingman]

I think you shouldn't test oc cards vs stock cards in some situation certainly.

The 980ti cards are 670 usually and the fury x is 650 or higher as well. So I really see no issue using an aib 980ti card that's 20 over the msrp. It's what the vast majority of people care about.

Using a 980ti reference card that very very few people are interested in, in a test means nothing to me. Show me the $670 980ti that are vastly superior that's what matters to me. The only reason people are whining is because it makes the 980ti look great.

Is it unfair if people don't use amd cards that are aib models when it makes sense? Sure... But we should tell them to make their tests better. Not to make other people's tests worse because we're mad at some "slight" from years prior....

It really depends on what type of article/review one wishes to portray, .ie motive. Comparing a nearly 1400mhz oc card to a stock one is hardly fair. If you have no problem with that, then there's not much to debate really.