Report: Wolfenstein 2 Optimized for AMD Vega with FP16 Shader Support

Subject: Graphics Cards | August 1, 2017 - 12:05 PM |
Tagged: Wolfenstein 2, vulkan, Vega, id Tech 6, id software, half-precision, game engine, FP16, amd

According to a report from Golem.de (German language), with the upcoming Wolfenstein 2: The New Colossus game AMD Vega owners will have the advantage of FP16 shader support from a new version of the id Tech 6 engine. The game supports both DX12 and the Vulkan API, but the use of half-precision calculations - the scope of which has not been specified - will potentially offer higher frame-rates for AMD Vega users.

View Full Size

AMD provided some technical details about Wolfenstein 2 during their Threadripper/Vega tech day, and this new game includes “special optimizations” in the id Tech 6 game engine for AMD Vega hardware:

“For what exactly id Software (is using) FP16 instead of FP32, AMD did not say. These could post-processing effects, such as bloom. The performance should increase in the double-digit percentage range, (though) id Software did not want to comment on it.” (Translated from German.)

Source: Golem.de

August 1, 2017 | 02:06 PM - Posted by Anonymouse (not verified)

Does this seem shady to you?

https://www.techpowerup.com/235701/nvidia-unlocks-certain-professional-f...

August 1, 2017 | 03:00 PM - Posted by rock1m1

They are opening up Quadro features, what is so shady about it? Fact that it was never intended for the Geforce line but you are getting some of it for free is great.

August 1, 2017 | 03:35 PM - Posted by Anonymouse (not verified)

I guess crippling a card via software to push people into Quattro cards is perfectly reasonable. If there was no Vega FE you can bet your bottom dollar they would have never enabled it.

What's more disturbing is you don't find a problem with it.

August 1, 2017 | 04:26 PM - Posted by Clmentoz (not verified)

No disabiling the Pro drivers in the consumer cards because the Pro Drivers cost millions of dollars more to develop is why Nvidia/AMD do it. And just paying the consumer card prices in not enough to cover those millions, so even the Titan XP/Vega FE SKUs are a little more costly than the consumer variants. Engineering and software engineering hours for the pro drivers runs in the millions of EXTRA hours more for all that EXTRA engineering(Software/Hardware) and certification hours and those folks do not appear to want to work for free, what about you!

August 1, 2017 | 04:37 PM - Posted by serpico (not verified)

"I guess crippling a card via software to push people into Quattro cards is perfectly reasonable."

Yes, it is. And if I ever owned a company and produced a non-essential product at whatever price I wanted I would definitly ignore you when you tell me that it isn't fair that I get to maximize my profit.

September 19, 2017 | 03:26 PM - Posted by tatakai

People will say this is ok. But if and when AMD does it, there will be hell to pay on reddit and forums. Ok for nvidia tho.

August 1, 2017 | 04:18 PM - Posted by Clmentoz (not verified)

Nvidia is not doing anything shady in enabling more of the Quardo types of driver features for its Titan XP SKUs than AND has done with its Radeon Pro Duo and Radeon Vega FE SKUs. So Nvidia is just offering developers some of the same types of Pro driver feature parity that AMD offers with its Radon Vega FE SKUs.

The purpose here is to give developers a lower cost hardware option with the same ability to target the same Pro drivers that are available with the more costly Quadro/Radeon Pro WX GPU hardware systems. It's so software development houses can save money by getting the Titan XP/Radeon Vega FE SKUs and use these lower cost SKUs to develop graphics/other software for the higher cost Quadro/Radeon Pro WX SKUs.

Now for graphics/engineering production purposes you still need the professional Quadro/Radeon Pro WX SKUs because they are the only SKUs with the ECC enabled/certified hardware, while for development purposes ONLY the software development houses can make use of the pro drivers made available on the Titan XP/Radeon FE lower cost SKUs to create software that targets the Quadro/Radeon Pro WX SKUs.

No civil engineering, aeronautical, or structural engineering, medical/other Production workloads can be run on anything but the Fully certified for ECC professional Quadro/Radeon Pro WX SKUs and associated Epyc/Xeon professional certified for ECC SKUs. And even the software developed on the Titan XP/Radeon FE SKUs will have to have their final testing/certification runs done on the Quadro/Radeon Pro WX SKUs.

So most development houses can get by with mostly using larger numbers of the Titan XP and Radeon Vega FE SKUs for software development then do the final certification workups on a few Quadro/Radeon Pro WX final certification units that are used for the final software certification/testing purposes only. That way there is no need to give every software developer on the software development project a more costly Quadro/Radeon Pro WX GPU and associated professional CPU/Motherboard based system.

You keep fewer Quadro/Radeon Pro/Epyc/Xeon based systems for final software certification/testing on the fully certified hardware, and use the more of low cost Titan XP/Radon Vega FE based, and non pro CPU based systems for most of the software development process, except the final testing/certification which has to be done on the Real Professional Workstation CPU/GPU hardware.

Most of the Folks commenting on that TechPowerUP article know NOTHING about the real world production workload usage of any Professional CPU/GPU SKUs for the professional market and the TOTAL need for the best possible accuracy on any PRODUCTION workloads where those workloads' results can put lives at stake. Workloads that require accuarcy at all costs like structural engineering production workloads, aircraft engineering production workloads, etc. Pro drivers cost millions more to develop and certify, that's why the Quadro/Radeon Pro WX SKUs cost more, to pay for all the hardware/software testing/certification expenses and millions of extra programming hours in takes to develop/certify software, and engineer the hardware used for professional production workloads.

Do you want the airplane you are flying in have its design done on a consumer/gaming computer using a consumer CPU/motherboard/RAM(non ECC) with non professional drivers/software or do you want things done of a professional Workstation/Computing platform with professional certified hardware dirvers and software. It's a long way down from 35,000 feet if the wing breaks off.

August 1, 2017 | 04:49 PM - Posted by Anonymously Anonymous (not verified)

not shady in the least bit. various companies in various industries do this all the time and it is perfectly legit.

I think the term you should have used is "does this seem shitty to you?" and I would wholeheartedly agree

August 1, 2017 | 06:18 PM - Posted by CNote

Just because bit involves shaders doesn't make it shady ;)

August 1, 2017 | 04:52 PM - Posted by Clmentoz (not verified)

Current games makers should be able to make use of the same types of tricks to use 16 bit FP computations for some color/other game computations as the eariler games industry did when 8 and 16 bit FP units where the only things available. There are ways to work with 16 bit values via extrapolation and interpolation and still have acceptable results. It only depends on how much the extra extrapolation and interpolation computations affect latency and how that latency can be hidden by software/hardware methods so that FPS metric is relatively unaffected.

August 1, 2017 | 07:13 PM - Posted by Godrilla

This could be the edge that amd needs to be competitive. If amd can't come on top in their own promoted title they will be in trouble. If they come on top then there is still hope.

August 2, 2017 | 05:40 AM - Posted by PeterG (not verified)

PS4 Pro has 8.4 TFLOPS of FP16 (Has Vega features)

Xbox One X does not support FP16 (Polaris arch, does not support any Vega features)

August 2, 2017 | 09:15 PM - Posted by TweakiLeaks (not verified)

According to tweaktown the RX Vega 56 is a GTX 1070 killer!

"Radeon RX Vega 56 benchmark results

. Battlefield 1: 95.4FPS (GTX 1070: 72.2FPS)
. Civilization 6: 85.1FPS (GTX 1070: 72.2FPS)
. DOOM: 101.2FPS (GTX 1070: 84.6FPS)
. COD:IW: 99.9FPS (GTX 1070: 92.1FPS)"(1)

"AMD Radeon RX Vega 56 leaked benchmarks: GTX 1070 killer
AMD Radeon RX Vega 56: the new $400 mainstream king"

http://www.tweaktown.com/news/58635/amd-radeon-rx-vega-56-leaked-benchma...

August 2, 2017 | 09:45 PM - Posted by TweakiLeaks (not verified)

P.S. Looking at the RX Vega 56's power figures on GamersNexus: 165W, 210W + Board, for the RX Vega 56. That makes the GTX 1070's power usage metric, at 150W, not so much lower than the Vega 56's 165W. So the RX Vaga 56 is more in line with Nvidia's stripped of compute gaming only focused SKUs and that's at RX Vega's 1156MHz/1471MHz base and boost clocks compared to the GTX 1060's 1657MHz/1860MHz base and boost clocks.

I wonder how close the RX Vega 56 overclocked can get to the 1080. And I'd love to see some dual RX Vega 56 benchmarks with the dual RX Vega 56 SKUs non-overclocked and maybe even underclocked/undervolted to save on power.

The RX Vega 56, like the Ryzen 7 1700 CPU counterpart, is shaping up to be the very same price/performance leader. I'll also be waiting for the RX Vega NANO's specifications to and that should be interesting also.

August 3, 2017 | 03:23 AM - Posted by Martin Trautvetter

Don't know why you keep perpetuating that AMD marketing obfuscation, but the power consumption comparison stands at 150W (1070) vs. 210W (56).

(and the 1070's clocks are 1506 MHz Base, 1683 MHz Boost)

"And I'd love to see some dual RX Vega 56 benchmarks with the dual RX Vega 56 SKUs"

Didn't you hear? Crossfire is supposedly being de-emphasized, if not deprecated.

August 3, 2017 | 10:00 AM - Posted by DeFUDingTrautSomething (not verified)

Take those power figures up with GamersNexus and they should be listing the GPU ASIC's power first then adding in any PCIe card power and Look at the leaked benchmarks that's on the RX Vega 56 at its lower clock rates relative to the GTX 1070. RX Vega 56 is beating the GTX 1070.

Sure they(AMD and others) are/should be de-emphasizing their COMPLEX/INEFFICTIVE Propitary GPU Driver based CF/SLI because the Vulkan/DX12 respective GPU driver models are more simple and to the metal. So that Multi-GPU load balancing responcibility has now been shifted to the Games/Gamimng engine makers. So it's up to the games/gaming engine developers to utilize their middlware Milti-GPU load balancing algorithms that call on DX12's/Vulkan's explicit multi-GPU functionality built into the DX12/Vulkan Graphics APIs.

Currently all the non-gaming graphics software can make use of more than one GPU for their rendering workloads so why not the gaming/game engine SDKs that can now call on the the DX12/Vulkan explicit multi-GPU adaptor functionality built into the new graphics APIs. Who cares about CF/SLI that technology has been obsoleted by the DX12/Vulkan APIs!

When you start to throwing those gaming card power usage figures around it's better to start with comparing the ASIC to ASIC power usage figures as the total card power usage will differ depending on power conditioning/cooling designs that are different for every refrence/AIB makers products.

Even the RX Vega 56 SKU with its reduced number of SPs still has more compute relative to the GTX 1070, but the RX Vega 56's ROP counts remain the same as the RX vega 64's ROP counts. So now here is the RX Vega 56's specs for fro TechPowerUp's GPU database.

For RX Vega 56

Shading Units: 3584
TMUs: 224
ROPs: 64
Compute Units: 56
Pixel Rate: 94.14 GPixel/s
Texture Rate: 329.5 GTexel/s
Floating-point performance: 10,544 GFLOPS
GPU Clock: 1156 MHz
Boost Clock: 1471 MHz
Memory Clock: 800 MHz 1600 MHz effective

Memory Size:8192 MB
Memory Type: HBM2
Memory Bus: 2048 bit
Bandwidth: 409.6 GB/s (look at that effective bandwidth and only at the HBM2's 800 MHZ DDR speeds(1600 MHz effective)

For the GTX 1070:

Shading Units: 1920
TMUs: 120
ROPs: 64
SM Count: 15
Pixel Rate: 107.7 GPixel/s (higher pixel fill rate the V 56)
Texture Rate: 202.0 GTexel/s (lower than the V 56)
Floating-point performance: 6,463 GFLOPS (lower than V 56)
GPU Clock: 1506 MHz
Boost Clock: 1683 MHz
Memory Clock: 2002 MHz 8008 MHz effective

Memory Size: 8192 MB
Memory Type: GDDR5
Memory Bus: 256 bit
Bandwidth: 256.3 GB/s (less effective bandwidth than the RX Vega 56's HBM2)

Founders Edition base/boost clocks: 1506 MHz 1683 MHz are listed in the database(See Note below).

[Note: the GTX 1070 is a released product with many AIB variants so in the TechPowerUp data base there are many varints listed with different base/boost memory clocks/etc.]

P.S. Trautvetter take the leaked RX Vega 56 benchmarks up with TeakTown! But RX Vega 56 design just needs a little more tweaking from AMD to remove enough compute to match Nvidia's compute gimping for power savings designs. Maybe AMD can afforde to whip up some mobile Vega variants with less compute with some of the extra Epyc/Ryzen/Threadripper revenues coming over the next few quarters, and the banks will surely be happy to lend to AMD short term and long term now that AMD is back in the game, and back big time with Zen, with Vega looking OK with some improvement on the way as always with AMD's driver/games optimizations over time.

Let it also be Known that DX12's and Vega's Explicit Multi-GPU adaptor usage is going to take over so, CF/SLI, do not let the door hit you on your way out!

August 3, 2017 | 05:16 PM - Posted by Martin Trautvetter

"Take those power figures up with GamersNexus and they should be listing the GPU ASIC's power first then adding in any PCIe card power"

I'm sorry inform you that the very site you're commenting on has (and has had) better equipment than Steve for testing and logging PCIe add-in board power consumption. What Steve is measuring as of a couple of weeks ago is the power used via the PEG connector or connectors, while PCPer is also monitoring power drawn through the PCIe slot itself.

But your sentence above explains to me why you seem so obsessed with GPU or ASIC power. You're misreading that as board power and believe that it equals the power measured through the PEG power connector, which it doesn't.

Short version:

Because layout and design of the power circuitry of add-in boards can vary, you need to log every interface capable of supplying power to the board simultaneously and add all separate source together to get the total board power consumption at any given time. (Which, incidentally, is exactly what PCPer does!)

In the case of the GTX 1070, that sum total is limited to 150W. According to AMD, Vega 56 is limited to 210W in the same measurement.

August 3, 2017 | 08:23 PM - Posted by NotPowerOnlyTrautKunItsRatiosToo (not verified)

And that RX Vega 56 SKU has more ShaderProcessors/compute than the GTX 1070 so the RX Vega 56 ASIC will use more power than the GTX 1070 ASIC, unless AMD has some alien technology in use. But still the RX Vega 56 is going to be fine as it has a pixel fill rate just about 94.14 GPixel/s or about 13.56 GPixel/s less than the GTX 1070's 107.7 GPixel/s fill rate. And the RX Vega 56's Texture Rate: 329.5 GTexel/s is about 127.5 GTexel/s more(A sizeable difference) compared to the GTX 1070's Texture Rate: 202.0 GTexel/s.

Look at RX Vega 56's Shading Units: 3584 with Vega's much deeper FP/INT pipelines needed for the higher clocks compared to the GTX 1070's Shading Units: 1920 figure and Nvidia has to use deeper pipelines the same for higher clocks. But look at these difference in shader units. The RX Vega 56 has 1664 more shader units so that extra power difference is not so bad on that extra power used for AMD's RX Vega 56. One thing to remember is to maybe look at a shader core processor per Watt metric along with some other unit ratios and then there is any future DX12/Vega games that may just put Vega's extra shaders to good use. Vega's primitive shaders also may be programmable for interesting things in future games along with using FP 16 more in games instead of all FP32.

I'm not obsessed with power usage but Nvidia made great headway into the Mobile/Laptop GPU market beginning with the Maxwell family of architectures, when Nvidia embarked on a “mobile first” design strategy for GPUs. So AMD only needs a little more tweaking with Vega's micro-arch to begin to strip out any unnecessary compute to save power on any mobile GPU SKUs based on the Vega GPU micro-arch. Maybe even AMD can leave in the more of the Vega NCU Shader cores and clock them lower to get the power savings and make use of Radeon Chill and get some low power parity with Nvidia's mobile/laptop offerings that way.

Navi will represent an all Raja/Raja's first fully designed/managed project under the RTG banner. So Vega is more of a Tweaking out of the previous GCN issues and Vega appears to be not so bad and will probably suprise after a time with the software/driver/gaming ecosystem still in the process of pivoting away from DX11 Towards DX12/Vulkan gaming with those new graphics APIs that are engineered to handle Vega's/GCN's extra compute hardware. Vega's HBCC/HBC IP and Infinity Fabric IP on Vega, as well as Zen, well that's somthing that could come into its own also gaming and compute wise for Vega's new features over and above Polaris and previous GCN GPU micro-archs.

And for sure Vega is brand new so yes the power testing is still underway and maybe GN/Steve is getting updates from his sources all the time, so things are still in a state of Flux as far as nailing down the proper specifications, and then there are the AIB designs to consider with all their custom engineering/tweaks. I'm using GN's and TechPowerUPs(GPU database) figures mostly. The NDA has still not expired so everybody(that signed the NDA) has to be careful until.

August 4, 2017 | 10:59 AM - Posted by Martin Trautvetter

Rather than comparing theoretical numbers across vastly different micro-architectures, I think it's more interesting, and closer to the available data, to compare how the different SKUs on each side stack up against one another:

Even if we didn't know about AMD's track record in scaling down its full-fat graphics cards (PRO vs. XT), the recently released data show Vega 56 much closer to its bigger brothers than the 1070 is to the 1080.

If Vega 56 lands at ~85% of Vega 64 with only 72% of the 64's power draw, while the 1070 lands at ~70% of the 1080 with 84% of its power draw, then based on roughly on-par performance of Vega 64 and GTX 1080, Vega 56 should rightfully beat the 1070, and quite handily depending on the specific tests.

Now efficiency might still paint a different picture, Vega 56 will likely have a hard time performing 40% higher then the 1070, which it would need to pull even based on their rated power use, but that's really not much of a victory in desktop environments.

Nvidia will likely need a better 1070 replacement, which I suppose shouldn't be an insurmountable problem given the large swaths of GP104 rendering resources disabled for its current guise. But introducing a 1070 v2 or 1075 at this point of the product life cycle seems highly undesirable, so they might want to avoid that even though Vega 56 will put pressure on a meaty part of their product stack. (all subject to availability, of course)

Personally, I find the business part of this game much more interesting, as I've moved beyond 1070-level performance a year ago.

August 3, 2017 | 10:39 AM - Posted by DeFUDingTrautSomething (not verified)

Also with respect to the above posts, I do not look at my GPU purchasing decisions with gaming only in mind as any Blender 3D rendering benchmarks are the more important, including acceleration of Ray Tracing/Ray Interaction calculations on the GPU's FP compute shaders/cores.

My reaons for wanting some dual-GPU(RX Vega 56) undervolting/underclocking benchmarking done is to suss out the RX Vega 56's optimal power uasge metrics that can be obtained from any Underclocking/Undervolting regimens as they relate to ainmation rendering workloads that see GPUs running for hours long animation rendering runs.

Dual RX Vega 56 usage for Blender 3D animation rendering and RX Vega 56's price/performance metrics may just make the RX Vega 56 as affordable for low cost rendering workloads and multi-GPU rendering as the respective RX 400/400 series Polaris GPUs are for coin mining.

WHO CARES ABOUT GAMING ONLY, FFS! IS'T NOT ALL ABOUT ONLY GAMING!

August 3, 2017 | 10:50 AM - Posted by DeFUDingTrautSomething (not verified)

Edit: RX 400/400 series Polaris GPUs
to: RX 400/500 series Polaris GPUs

Post new comment

The content of this field is kept private and will not be shown publicly.
  • Lines and paragraphs break automatically.
  • Allowed HTML tags: <a> <em> <strong> <cite> <code> <ul> <ol> <li> <dl> <dt> <dd> <blockquote><p><br>
  • Web page addresses and e-mail addresses turn into links automatically.

More information about formatting options

By submitting this form, you accept the Mollom privacy policy.