AMD Announces Radeon Pro Duo: Dual GPU Fiji Video Card For VR Content Creationby Ryan Smith on March 14, 2016 7:00 PM EST
At AMD’s GDC 2016 “Capsaicin” event, the company has announced their long-awaited dual Fiji video card. Being released under the name Radeon Pro Duo, the new card is a departure from the usual for AMD, with the company specifically targeting it towards VR content creation rather than end-user gaming.
AMD originally teased their then in-development dual-Fiji card back at the Fiji launch event in June of 2015. At the time the card was expected to launch towards the end of 2015 as the company’s flagship gaming card. However at AMD’s Polaris event in December, the company announced that they were realigning the card to focus on the VR market, and would be holding it back to 2016 to launch alongside the major VR headsets.
Officially AMD’s commentary was limited reiterating their desire to have the card tied to the VR industry. However I believe that AMD also delayed the card due to the poor state of AFR scaling in recent AAA games, which would make a dual-GPU card a hard sale in the typical PC gaming market. VR, by contrast, is a much better fit since through technologies such as AMD’s affinity multi-GPU, the two perspectives that need to be rendered for VR can be mapped directly to each GPU, avoiding AFR’s dependency and pacing issues.
In any case, with the launch of the major VR headsets finally upon us, AMD is formally unveiling their dual Fiji card, the Radeon Pro Duo. That AMD is still not going after the consumer market means they have once again defied expectations, but first let’s take a look at the specs as we know them so far.
|AMD GPU Specification Comparison|
|AMD Radeon Pro Duo||AMD Radeon R9 Fury X||AMD Radeon R9 Fury||AMD Radeon R9 295X2|
|Stream Processors||2 x 4096||4096||3584||2 x 2816|
|Texture Units||2 x 256||256||224||2 x 176|
|ROPs||2 x 64||64||64||2 x 64|
|Memory Clock||1Gbps HBM||1Gbps HBM||1Gbps HBM||5Gbps GDDR5|
|Memory Bus Width||2 x 4096-bit||4096-bit||4096-bit||2 x 512-bit|
|VRAM||2 x 4GB||4GB||4GB||2 x 4GB|
|Transistor Count||2 x 8.9B||8.9B||8.9B||2 x 6.2B|
|Typical Board Power||350W||275W||275W||500W|
|Manufacturing Process||TSMC 28nm||TSMC 28nm||TSMC 28nm||TSMC 28nm|
|Architecture||GCN 1.2||GCN 1.2||GCN 1.2||GCN 1.1|
|Launch Date||Q2 2016||06/24/2015||07/14/2015||04/21/2014|
Officially, AMD promotes the Radeon Pro Duo as having 16 TFLOPS of performance; unsurprisingly this translates to two fully enabled Fiji GPUs, clocked at around 1GHz. This puts the maximum performance of the card very close to a Fury X Crossfire setup, however there is still the matter of TDP to get to.
Otherwise as this is Fiji, the rest of the specifications should not come as a surprise. Doubling up on Fijis gives us 64 ROPs and 256 texture units per GPU. And 4GB of HBM per GPU, clocked at 1Gbps for an effective memory bandwidth of 512GB/sec/GPU.
Meanwhile on power consumption, though not in the initial press release, AMD has confirmed that the card is configured for a typical board power of 350W. In practice what this means is that the card is not too far removed from two R9 Nanos glued together, seeing as how R9 Nano had a TBP of 175W. However this also means that like the R9 Nano, the card is expected to power throttle under most heavy tasks, and the full 1GHz clockspeed will rarely see action. Again if it’s anything like the single Nano, this would put the average clockspeeds at around 875MHz, so it’s not quite a single card Fury X Crossfire analog.
Even with the 350W TBP, the shots of the card in AMD’s press materials all show 3 8-pin PCIe sockets, which would put the maximum power draw as officially allowed by the PCIe specification at 525W. Given the target professional market, all signs point to AMD opting to be conservative here and stick to the PCIe specification rather than risk pulling too much power over two sockets. These matters are rarely important to consumers, but it matters a lot to OEMs who may bundle the card, or at least offer it as an option. Meanwhile this also means that the Radeon Pro Duo is set to consume quite a bit less power than AMD’s previous generation dual-GPU card, the Radeon R9 295X2. That card was rated for 500W and could come very close to actually drawing that.
With the 350W TBP, AMD has once again resorted to a closed loop liquid cooler setup in order to keep the card at two slots wide. We don’t have detailed specifications for the radiator, but it is a single 120mm radiator, and it looks quite similar to the Fury X’s. The Fury X was essentially overbuilt in this regard, so it’s easy to see why AMD wouldn’t need a larger CLLC. This also means that the Radeon Duo Pro improves upon the R9 295X2 in a small but important way: everything is cooled by the radiator; there isn’t a small fan at the center on top of all of this.
As for display I/O, AMD’s official specs list 4x DisplayPort. However based on the admittedly limited pictures AMD has released, I believe this is an error on their part. The 4th port on the card looks a great deal like an HDMI port, which would make a lot of sense to have in place since HDMI is needed to drive the HTC Vive and Oculus Rift.
But perhaps the bigger news though isn’t the specifications, but the target market for the card. While I had initially expected AMD to target the card at the VR consumer market, AMD has gone in a different direction. Rather the Radeon Pro Duo is being pitched as a content creation card, making this an unusual halfway point between a Radeon and a FirePro.
As I’m writing this up in advance I haven’t heard AMD’s formal reasoning for why they aren’t heavily promoting it for the consumer market – though clearly the card will work there – but after giving it some thought I suspect it has to do with the system requirements for VR gaming. Both Oculus and Valve are pushing the idea that a Radeon R9 290/GeForce GTX 970 should be the performance level VR games are designed around. If developers actually follow through on this, then having a faster card is not especially useful since VR displays are locked to v-sync and can’t exceed their cap. If a 290 delivers 90fps, what would a Pro Duo when developers are targeting a fixed level of quality?
In which case content creation is the next best thing to do with the card. Games under development have yet to be tuned for performance, so it’s sound reasoning that developers would want something as fast as possible to do their initial development on. The catch for AMD is that this does limit the market for the card; besides the high price tag, the market for developers is much smaller than the market for consumers.
And since it’s a content creation card, it will also be receiving special driver attention from AMD. The Radeon Pro Duo will still use the Radeon driver set, but the drivers for it will be validated for a selection of major content creation applications (modeling, animation, etc). Validated drivers won’t come down the pipeline until a month or so after the card launches, but given AMD’s workstation aspirations here, they want to give users a FirePro-lite experience and not force users into picking between a powerful card and drivers that the major tool developers will support.
Finally, let’s talk pricing and availability. AMD has announced that the card will retail for $1499. This is the same price that the Radeon R9 295X2 launched at in 2014, however it’s more than double the price of a pair of Fury Xes, so pricing is arguably not aggressive there. On the other hand it’s more compact than a pair of Fury Xes (or even a pair of Nanos), so there is the space argument to be made, and as AMD’s positioning makes clear this is first and foremost a development card to begin with. Meanwhile the Pro Duo will be shipping in “early Q2 2016”, which means we should see it become available in the next one to two months.
Post Your CommentPlease log in or sign up to comment.
View All Comments
Valantar - Tuesday, March 15, 2016 - linkThe post originally talked about 525W, the maximum power draw available through PCIe+3 8-pin connectors. The post did state that this was unconfirmed. After AMD fleshed out the specs, the post was updated, but this was never noted anywhere.
Wreckage - Monday, March 14, 2016 - link24 pins for power! I bet the lights dim when you turn the system on. A shame it's crippled by the 4GB memory limit.
looncraz - Monday, March 14, 2016 - link4GB VRAM is not yet a limiting factor for gaming. The 970 can get away with 3.5, can't it?
russomd - Monday, March 14, 2016 - linkAt $1500 it's a future investment and 4 gig's isn't enough for 4K games maxed out. The 4 gigs of ram is a bad point about the card.
Kutark - Monday, March 14, 2016 - linkExactly. Anybody who drops $1500 on this right now IMO is out of their minds. Particularly with the fact that NVidia is getting within a few months of Pascal, which will likely have 8gb per card, and you could SLI 2 of the x80 variants for ~$1100 (based on usual pricing at launch of top end card), and likely get significantly better performance, double the VRAM, and my guess using probably 60-80% off the power of this card.
Kutark - Monday, March 14, 2016 - linkI meant so say 60-80% OF, not off.
Valantar - Tuesday, March 15, 2016 - linkThis is probably why they're aiming it at VR developers. 4GB of RAM per eye is more than enough - we've yet to see any VR headsets even reaching total resolutions of 4K, let alone per eye. Most seem to run around QHD or something similar, in which case 4GB is plenty. And given proper support for Affinity Multi-GPU, you can halve those requirements per eye. In other words, this should be very future proof for VR development.
Also, given promised "semi-FirePro" level driver validation and support, this should be a far better solution than CFX/SLI-ing a couple of cheaper GPUs together. After all, professional GPUs are EXPENSIVE: $2000 gets you a Quadro M5000, which is essentially a GTX 980 with better driver support and ECC memory, and while dual 980s in SLI would only cost you $1000, driver validation for necessary applications would be nonexistent.
mobutu - Tuesday, March 15, 2016 - link"future investment"?
with obviously superior tech as pascal and polaris coming in half a year or so.
close - Tuesday, March 15, 2016 - linkIs 8GB enough? Maybe I'm missing something with your count but this thing has 2 GPUs with 4GB each.
toxicfiend1957 - Tuesday, March 15, 2016 - link8 gig of vram not 4 dual gpu 2 x 4096 so 8 gig not 4.