[News] AMD Adds new Radeon Pro Duo with two Polaris 10 GPUs

[StyM]

http://www.guru3d.com/news-story/amd...s-10-gpus.html

Today, AMD announced the world?s first dual-GPU graphics card for professionals, the Radeon Pro Duo, designed to deliver outstanding performance and flexibility in today's most complex multi-tasking workflows.

Today AMD(NASDAQ: AMD) announced the world?s first dual-GPU graphics card designed for professionals: the Polaris-architecture-based Radeon?Pro Duo. Built on the capabilities of the Radeon?Pro WX 7100, the Radeon ProDuoprofessional graphics cardis designed to excel at media and entertainment, broadcast, and design and manufacturing workflows, delivering outstanding performance and superior flexibility that today?s creative professionals demand.

The Radeon Pro Duo is equipped with 32GB of ultra-fast GDDR5 memory to handle larger data sets, more intricate 3D models, higher resolution videos, and complex assemblies with ease. Operatingat a max power of 250W, the Radeon Pro Duo harnesses a total of 72 compute units (4608 stream processors) for a combined performance of up to 11.45 TFLOPS of single-precision compute performance on one board, and twice the geometry throughput of the Radeon? Pro WX 7100.2The Radeon Pro Duo enables professionals to work up to four 4K monitors at 60Hz, drive the latest 8K single monitor display at30Hz using a single cable, or drive an 8K display at 60Hz using a dual cable solution.

?Today?s professional workflows continue to increase in complexity, often demanding that creators switch between a wide variety of applications to progress their work, pausing efforts in one application while computing resources are focused on another. We designed the Radeon Pro Duo to eliminate those constraints, empowering professionals to multi-task without compromise, dedicating GPU resources where and how they need them. It?s a continuation of our promise for Radeon Pro: to providegreater choice in howprofessionals practice their craft, enablingsuperior multi-tasking, accelerated applications, and powerful solutions for advanced workloadslike VR,? said Ogi Brkic, general manager, professional graphics, Radeon Technologies Group, AMD.

[BeepBeep2]

This makes absolutely no sense.

4096 SP Vega will match / destroy this(???)

[EternityZX9]

This makes absolutely no sense.

4096 SP Vega will match / destroy this(???)

It's mid tier graphics card for professional content creators...ya know for those that don't want to spend $1500 for a professional card but are ok with $1000.

It's a small niche market to be honest.

Vega 10's market is supposed to be mainly for high end gamers.

[niobium615]

This makes absolutely no sense.

4096 SP Vega will match / destroy this(???)

It has substantially more VRAM though at the price point. I don't believe that 32GB, single GPU, HBM2 memory configurations are possible yet, so the only competitor would be single GPU Vega @16GB. The pro version of Vega will most likely also launch at quite a bit more than $1000. Historically it seems as though the dual-GPU AMD cards are launched ~1year after the consumer cards are initially launched, putting the dual-GPU Vega card with equivalent memory size at about Q1 2018.

[vario]

It has substantially more VRAM though at the price point. I don't believe that 32GB, single GPU, HBM2 memory configurations are possible yet, so the only competitor would be single GPU Vega @16GB. The pro version of Vega will most likely also launch at quite a bit more than $1000. Historically it seems as though the dual-GPU AMD cards are launched ~1year after the consumer cards are initially launched, putting the dual-GPU Vega card with equivalent memory size at about Q1 2018.

As it is a CF in one package, it will have 16gb of available memory each gpu having own 16gb of the same stuff.
I dont see a much use to this card, unless vega is in some kind of problem.

[niobium615]

As it is a CF in one package, it will have 16gb of available memory each gpu having own 16gb of the same stuff.
I dont see a much use to this card, unless vega is in some kind of problem.

Of course, but two 16GB pro Vega cards would likely be 2-3x the price. The way I see it, this card is for those who needs lots of VRAM and not much raw computational power on a budget.

[BeepBeep2]

It's mid tier graphics card for professional content creators...ya know for those that don't want to spend $1500 for a professional card but are ok with $1000.

It's a small niche market to be honest.

Vega 10's market is supposed to be mainly for high end gamers.

But that's the thing - Vega 10's market is not going to be mainly for high end gamers, it is going to be AMD's catch-all competitor to GP100 and GP102. Vega will have ~12-13 TFlops of 32-bit FP, and over twice the 16-bit FP of Polaris within the same TDP and up to 16GB HBM2, so why spend the time/money/energy to release this card? Is 16GB HBM2 not going to be enough for the intended market?

Not to mention the current Pro Duo is still on sale and in stock for ~$800-1000 with 8096 SP's and similar perf/w, though 4GB per GPU looks much less attractive than 16GB per GPU.

It has substantially more VRAM though at the price point. I don't believe that 32GB, single GPU, HBM2 memory configurations are possible yet, so the only competitor would be single GPU Vega @16GB. The pro version of Vega will most likely also launch at quite a bit more than $1000. Historically it seems as though the dual-GPU AMD cards are launched ~1year after the consumer cards are initially launched, putting the dual-GPU Vega card with equivalent memory size at about Q1 2018.

This is true, but upon release of 16GB Single-GPU Vega, this card will be rendered useless for anything but workloads that A. Need to aggregate more than 16GB memory and B. Somehow manage that without requiring high memory bandwidth - Polaris' bus width is still too low to compete per-GPU with 2x HBM2 stack Vega which will have higher bandwidth.

As it is a CF in one package, it will have 16gb of available memory each gpu having own 16gb of the same stuff.
I dont see a much use to this card, unless vega is in some kind of problem.

Exactly. I also wonder about how much the architecture is improved in Vega, with all of the marketing done so far - ES seemed to be running lower clocks and having less gaming performance than one would hope for, and High-Bandwidth Cache seems to be a 12 year old idea only applied to high end GPUs. (In case the 8GB VRAM doesn't happen to be enough?)

Of course, the draw stream binning rasterizer is new, patent filed after Polaris was released so I suppose we will see.

For what Ryzen has accomplished vs. Core i7 w/ 14nm LPP, I wonder sometimes why Polaris is not better than it is. RX 580 now slots in between the GTX 970 and 980 in both performance and power consumption. It's almost beginning to feel like GPUllozer™

Of course, but two 16GB pro Vega cards would likely be 2-3x the price. The way I see it, this card is for those who needs lots of VRAM and not much raw computational power on a budget.

Fair enough, but one 16GB Vega card would do everything this one needed to do, faster, provided the workload did not need more than 16GB VRAM.

The only use I could think of this would be image ray-traced rendering like Blender @ 12K+ resolution and lots of huge textures.

[zanzabar]

It's mid tier graphics card for professional content creators...ya know for those that don't want to spend $1500 for a professional card but are ok with $1000.

It's a small niche market to be honest.

Vega 10's market is supposed to be mainly for high end gamers.

the old pro duo with 2xfury (2 high end parts) sells for like $700 ATM and you can still get them.

[kromosto]

Amd started to make strange things instead of making people focus on VEGA release. Suspicious!!!

[zanzabar]

Amd started to make strange things instead of making people focus on VEGA release. Suspicious!!!

hmb2 still has no part numbers and you normally have to make orders 60-90 days before it ships. they really screwed up with no having a gddr or hybrid fall back. at this rate i dont think we will see any vega 10 cards, and we may never see any vega cards.

[StyM]

AMD Announces New Radeon Pro Duo - Polaris x2

[niobium615]

What really makes me think this is solely for VRAM is the old pro duo. A Fury GPU is more powerful than one of those Polaris 10 GPU, so the old pro duo is objectively more powerful(and available for $700), yet it only comes with 8GB VRAM total. Otherwise, if you didn't need the VRAM, there's no reason not to go with the old pro duo.

[AbortRetryFail?]

A single Fiji pro card (think R9 Nano @ <170w, or the Radeon Instinct 'Mi18') has roughly 30%+ more FP giddy-up than a single Polaris pro card ...

My understanding is throwing a bunch of VRAM and bandwidth at deep learning has diminishing return (if any).

I'm guessing the Polaris Pro-Duo is more intended for "Instinct Clusters" and AMD's vision of 'cloud-power' for multiple users. It would also appear to hit a sweet spot in their professional card price points based upon power consumption and performance.