AMD, Analysis, Breaking, Game Developer Conference (GDC), Graphics, Hardware, VR World

Radeon Pro Duo: How AMD’s Dual GPU went Professional

When AMD revealed the Radeon Pro Duo during the Capsaicin event, the actual product looked significantly different to the one AMD showed at last year’s E3 conference. In fact, it was a completely different product. Besides an occasional leak, destiny of AMD’s ‘Gemini’ board was unknown for the better part of 2015. In this article, we will reveal the full story of how Radeon Pro Duo came to be.

The original plan to launch AMD’s “Radeon R9 Fury X2” for gamers changed with the announcement of Radeon Technologies Group (RTG), a new organisation inside AMD led by Raja Koduri, specifically designed to focus on AMD’s FirePro, Radeon and now also Radeon Pro lines. RTG will drive the adoption of in-house designed technologies throughout AMD product line-up, as well as current and future licensees. AMD is known for licensing its IP to 3rd parties – such as processors inside Microsoft’s Xbox One and Sony’s PlayStation 4 console. Same story with Nintendo’s (Wii and Wii U consoles), meaning AMD has a 100% share in contemporary gaming consoles.

Fiji GPU brought revolutionary packaging, meaning RTG and AMD wanted to create halo products based on halo architecture. Given that Polaris and Zen, top-to-bottom shift to 14nm FinFET were a generation away, the decision came to design three families of future products based around Fiji, continuing onto Polaris / Zen. The RTG team decided to scrap a high-end gaming product and merge the ‘best of Radeon and best of FirePro’ into a single product. End result is the birth of prosumer Radeon Pro family, with Radeon Pro Duo being the first product.

Building a Graphics card for Project Quantum

Project Quantum was invented to demonstrate record performance in the smallest form factor. Fiji X2 GPU produces 17 TFLOPS in less than 25cm, approx. 4x more efficient than Intel.

Project Quantum was invented to demonstrate record performance in the smallest form factor.

Codenamed Gemini, R9 Fury X2 was originally designed as a discrete part that can fit inside AMD’s Project Quantum, a ‘computer for VR’ concept which made its debut alongside the Fiji family of products. This was also the first product from AMD to make it to the mainstream media in quite some time and carried a global appeal. Regardless of what you may think about this product, we think that this is one sexy piece of hardware, and received open approval from my wife and her female friends as ‘computer they would buy’.

Unfortunately, AMD hit the same limitations with the supply chain as Razer encountered when they wanted green USB ports. The state of current supply chain simply isn’t geared to support an innovative form factor such as Project Quantum. In our conversations with RTG’s leadership, we learned that they did not give up on ‘Project Quantum’, and the second window just might be the launch of ZEN architecture.

With Project Quantum being pushed onto the back-burner, green light was given to create an uncompromising design merging consumer and commercial aspects, the Radeon Pro Duo.

Learning from R9 Fury X and Fury X2 (Gemini)

Built for Project Quantum, Gemini board sports two Fiji GPUs.

Built for Project Quantum, AMD Gemini prototype sports two Fiji GPUs.

Picture above shows the original R9 Fury X2 (engineering prototype). Engineers behind the board demonstrated to AMD’s management that Fiji GPU can drive new, previous unimaginable form factors. Gemini design removed any internal doubt that R9 Nano can be a viable product. Personally, out of all Fiji-based graphics cards, I prefer the compact design and incredible compute performance from ITX-focused R9 Nano.

Gemini was designed to feed the GPus with 375 Watts; 300 Watts came from two 8-pin PEG connectors (seen in the right corner) and 75 Watts came through the motherboard. Given design of PCI Express bus, there was a valid concern that continuous load on data paths on with continuous draw of 75 Watts could lead to instability in some of motherboards on the market. The Gemini board was literally two Nanos next to each other – albeit at a lower clock. The prototype delivered 12 TFLOPS of single and 0.75 TFLOPS of double precision compute performance. Samples could achieve R9 Nano clock and performance (16 TFLOPS SP, 1 TFLOPS DP), but the margin for power delivery was minimal; ~355W out of 375W available.

One upping the original Gemini design proved to be an easy task for the hardware engineers, and a daunting task for its liquid cooling partner, Cooler Master. Spec for Radeon Pro Duo called for professional or industrial grade design capable of qualifying in workstation systems. In a way, this graphics card is more a FirePro than a Radeon. Engineers at AMD and Cooler Master used the same specifications demanding durability and 60-month, 5 year ’50/25/20/5′ cycle (50% full load, 25% half-load, 20% idle, 5% off). This is the first time AMD used FirePro workstation / server design guidelines for a consumer/prosumer product.

Initial R9 Fury X boards suffered from coil whine (marked red)

Limited number of R9 Fury X boards suffered from coil whine (chokes are marked red). Green markings show the base component of the new power design.

AMD R9 Fury X cooling design by Cooler Master did not cover the chokes.

AMD R9 Fury X cooling design by Cooler Master did not cover the chokes.

On R9 Fury X, Cooler Master designed an ASIC focused design, cooling down just the GPU and HBM memory. The liquid from the GPU/HBM block then turned into a heatpipe which passed through the hottest parts of power delivery system. Retail versions of the card came with an updated cooling pump which resolved the issues on review boards (the sticker went from this multi-color design to a single chrome sticker). One of lessons learned was that if you’re going for a fully passive design, without fan cooling the board (PCB) – you have to cover as much components as possible, especially ones that heat up to 100 degrees Celsius.

Meet the Radeon Pro Duo

AMD Radeon Pro Duo

For the Radeon Pro Duo, AMD went with no-holds-barred approach. The board is long – as long as the previous dual-GPU board from AMD, Radeon R9 295X2. First thing you will notice is the re-positioning of the liquid cooling tubes from the front of the card to the rear. These cables are longer than they were on R9 Fury X, giving you a multitude of options to integrate the Radeon Pro Duo even in big tower cases.

AMD Radeon Pro Duo PCB as shown on Capsaicin webstream.

AMD Radeon Pro Duo PCB as shown on Capsaicin webstream.

R9 Fury X featured two different types of power regulators (choke, inductor), while Radeon Pro Duo moved to a single-type design (inductor-only). We had eight R9 Fury X boards from the first production run and three of those displayed the ‘coil whine’ issue. By the time boards made it to production, coil issues would appear on approximately 1 in 30 boards, and if that happened to you – RMA is a non-issue – just like on whiny GeForce GTX 970 / 980 boards. After all, these design are Formula 1 cards of computing world – never forget Fiji silicon has a record 8.9 billion transistors, which put a lot of strain on the power delivery system. Going forward, AMD’s new power design philosophy should eradicate this potential issue for good.

In Radeon Pro Duo, engineers packed no less than 18 inductors (3 more than on Gemini). The board brings a total of 15 dedicated full-length power phases and additional filters, capable of delivering 525 Watts. While the stock clocks on Radeon Pro Duo are equal to ones on a R9 Nano, there is a significant headroom for overclocking. 175 Watts, to be precise. This should enable your Radeon Pro Duo to achieve equal clocks as overclocked R9 Fury X cards, albeit on a single board.

Next step, the cooling:

Cooling design on Radeon Pro Duo reveals workstation-class approach.

Cooling design on Radeon Pro Duo reveals workstation-class approach.

The new fully-enclosed design no longer allows for those nine activity LED’s that AMD featured on the Fury boards. Howwever, it’s good to see that AMD kept the customizable front panel. By removing five screws, you can insert a personalized plate. You can download a 3D model for the R9 Fury X and we are pretty sure AMD will offer a 3D model for the Pro Duo as well. Once removed, we can see all the changes AMD and Cooler Master did. Practically every heat emitting element is completely covered by the red colored liquid cooling block.

The pump is hidden beneath one GPU, while the second GPU is features a block and regulates the cabling. Right side of the board features a screw through which the liquid is being introduced into the system. Unless you want to experiment with different types of liquids (and void your warranty), don’t touch the screw. Not pictured is the radiator and the fan. They’re identical to the system used on R9 Fury X, which should give you 500 Watts of thermal dissipation capacity.

Consumer Meets Professional

After going through all the hardware, we come to the reason why this graphics card became Radeon FirePro Duo. Through the course of this article, you saw what kind of hardware changes AMD made – and they were not forced to do those changes, as the original Gemini board worked perfectly in games. AMD beefed up the design in order to be able to deliver high performance in professional applications. Given the price difference between Radeon (Consumer) and FirePro (Commercial), Radeon Pro Duo just might be the first real ‘prosumer’ graphics card – rising above consumer to meet the commercial demands.

The video above shows where are we heading with VR experiences (#VRex). Mixing not one, but two VR setups to drive VR experience puts strain on a system which no single-GPU can handle. Luckily, with the arrival of Microsoft Windows 10 and the DirectX 12 API, we have a situation where the API sees multiple frame buffers as one. This still works even if you mix cross-vendor multi-GPU combinations, which we all saw with an updated benchmark for Ashes of The Singularity. In a professional world, this approach is not unusual but as always it needed the strength of Microsoft to standardize frame buffer refresh and optimize Radeon Pro Duo brings unified 8GB buffer for the first time, and preliminary results in Adobe’s Creative Suite show much promise.

Most Virtual Reality Experiences (VRex) today are 360 degree videos recorded on cameras such as Google’s Jump (GoPro Odyssey), Ozo from Nokia, Jaunt VR as well as DIY sets from 360Heroes and Freedom360. In order to utilize video stitching, applications such as Jump Assembler need a lot of GPU horsepower. We’ll give you a brief example of what you need to calculate, taking Google Jump into account:

  • Single 1080p, Full HD video frame is 2.07 million pixels, which turns into 124.42 million pixels per second (1080p60 is needed for VR).
  • Google Jump: 16 1080p60 video streams generate 1.99 billion pixels each second. That’s 1.99 GPixel/sec.
  • 2.7K i.e. 2704×1520 is 4.11 million pixels. 2.7K60 generates 246.60 million pixels each second.
  • Google Jump: 16 2.7K60 video streams generates 3.95 billion pixels each second (Gpixel/sec).
  • Single 2160p, UltraHD video frame is 8.29 million pixels, which turns into 250.13 million pixels per second.
  • Google Jump can record 4K at 30fps, i.e. 4.00 billion pixels per second.

That’s frames. But once we convert the frames into megabytes, this is the picture you get (all numbers are for uncompressed data):

  • 8.1 MB – 1920×1080 frame
  • 16.06 MB – 2704×1520 frame
  • 32.57 MB 3860×2160 frame

Thus, taking the streams into the account, this is the demand GPUs face today – if they are going to be used to create VR experiences:

  • 7.77 GB/s – 1080p60 times 16
  • 15.42 GB/s – 2.7K60 times 16
  • 15.63 GB/s – 4K30 times 16

Do note that we skipped on 1080P at 120 fps, which in a Google Jump / GoPro Odyssey configuration demands 15.55 GB/s. As we all know, PCI Express 3.0 x16 can deliver 15.75 GB/s. PCI Express 3.1 brings improvements in functionality, performance and power management – but was not supported by any motherboard. Luckily, PCI Express switch on Radeon Pro Duo enabled the dual Fiji GPUs to continuously operate at 99.9%, satisfying the demands of all the popular 360 video standards. Do note than an idea of seamless data transfer through the CPU goes out the door due to introduced latency. By keeping everything on one PCB, latency is reduced to near zero.

The raw stream demand requires a lot of RAM, as-fast-as-you-can-find storage subsystem and a graphics processor capable of taking as much video information as possible. Mr. Roy Taylor (Corporate VP Alliances and Content, AMD) lives in Hollywood, some five miles from where Walter Jeremiah ‘Jerry’ Sanders, III, founder of AMD and legendary CEO lives to this date. Working with partners in Hollywood, AMD managed to gather a lot of feedback on what is missing in the production pipeline, and Radeon Pro Duo is a first generation reply to the demands from content creators.

Over a decade ago, Mr. Taylor was the driving force behind Nvidia’s ‘The Way its Meant To Be Played’ game developer program, which yielded numerous AAA Game franchises. By leaving Nvidia and moving to Hollywood the focus shifted from end product (games, movies, music) to content production, which is visible in VR experiences such as MATTERvr’s ‘First’ for the Smithsonian Museum. Or more recently, OTOY creating a wrapper which enables applications written in CUDA to utilize all the hardware, such as Radeon Pro Duo. Thus, Radeon family of products is shifting from focusing just on games – with Crimson driver, for example – into content production as well.

The only regret with the Fiji GPU we have is that AMD kept the double-precision performance artificially locked at 1/16th, meaning Radeon Pro Duo delivers 16 TFLOPS of Single Precision and 1 TFLOPS of Double Performance. Given that in compute, AMD GPUs architecture operates just like the CPU (delivering unlocked double precision floating point at 50%), Radeon Pro Duo might have delivered 8 TFLOPS of Double precision, making this an ultimate production-grade processor. Perhaps a FirePro version of this board might see the light of day with unlocked performance. In that case, 4GB capacity per GPU would not be detrimental as 512 GB/s of bandwidth takes care of quick memory refresh (compare vs. ‘best case scenario’ 66.65 GB/s on the Xeon E5 processors).

Devil is in Details… Drivers

AMD recently signed a collaboration with legendary news agency AP.

AMD recently signed a collaboration with legendary news agency AP.

Now that you know the demands of VR content, it is clear that the content creation community needed a serious amount of horsepower. This is why you can buy a system from BOXX Technologies featuring no less than six graphics cards (up to 11 GPUs as one graphics card needs to be single-slot). Radeon Pro Duo is a stepping stone to AMD’s new strategy: covering gamers (Radeon), prosumers (Radeon Pro) and professionals (Fire Pro). AMD recently signed a collaboration with Associated Press to launch a Virtual Reality Journalism platform and the news agency needs as much compute power as anyone.

Radeon Pro Duo comes with Crimson Driver that Radeon owners know as the gaming driver, and expand the performance by adding a host of profiles for commercial applications such as Adobe Creative Studio, Autodesk 3Ds Max or Maya, and many more. These profiles are typically present only on products such as AMD FirePro and Nvidia Quadro.

“Be more productive by completing tasks lightning fast. With dual GPUs enabled with high bandwidth memory (HBM) technology, Radeon™ Pro Duo graphics has the parallel computing power to speed through OpenCL™ supported applications such as Adobe® Premiere, Adobe® After Effects, Blackmagic DaVinci Resolve Studio and Blackmagic Fusion Studio.​”

The exact details on profiles are not known at the time of writing, but during the Game Developers Conference, Autodesk demonstrated performance gains on Radeon Pro Duo over the consumer cars in Maya, a popular 3D modelling suite. Performance increase measured in double digits over R9 Fury cards in Crossfire mode. We did not manage to confirm will these profiles be ‘certified’ as with FirePro cards, but we would not be surprised if that ends up being the case.


Radeon Pro Duo will come to market sometimes in April 2016, aligned with the launch of HTC Vive, coming just a week or two after Oculus starts shipping their VR part. For the $1499 the company asks, you will be able to purchase probably the best product that AMD ever produced, showing the maturity of just launched Radeon Technologies Group. If Radeon Pro Duo succeeds on the market, we would not be surprised if we see a Radeon Pro branded dual-GPU becoming an integral part of being launched alongside the single GPU line-up.