AMD Announces New Radeon Embedded GPUs

[btarunr]

AMD today announced the AMD Embedded Radeon E9170 Series graphics processing unit (GPU). The new processor is the first "Polaris" architecture-based AMD Embedded discrete GPU available in multi-chip module (MCM) format with integrated memory for smaller, power-efficient custom designs, as well as PCI Express and MXM formats for standard form factor systems. The E9170 Series GPU is ideal for devices that require premium graphics and expanded display capabilities while meeting exacting power and thermal efficiency demands. AMD is extending its core graphics technology, delivering crystal clear resolution and a stunning and seamless 4K experience across multiple displays to a growing number of markets, including digital casino games, thin clients, medical displays, retail and digital signage, and industrial systems.

The AMD Embedded Radeon E9173 GPU, based on the "Polaris" architecture, leverages an optimized 14nm FinFET manufacturing process to provide up to 3X the performance-per-watt over previous generations of AMD Embedded GPUsi. By offering sub-40W TDP in a small package, AMD enables a broader range of products, adding a new level of scalability to the AMD Radeon Power-Efficient Embedded GPU portfolio. With support for up to five simultaneous 4K displaysii, the E9170 Series GPU virtually eliminates the need for additional processors and duplicate hardware to create an immersive multimedia environment. Additionally, the option to select from MCM, MXM and PCI Express modules increases design flexibility while minimizing complexity.



"Developers are continuing to push the boundaries on what's possible for embedded systems, demanding more performance, more features and more design options, all while significantly reducing power consumption. At the GPU level, it is critical to provide versatile solutions that do not compromise on graphics performance or 4K multi-display capabilities," said Colin Cureton, director of product management, AMD Enterprise Solutions. "The new AMD Embedded Radeon E9170 Series GPU not only achieves up to 3X the performance-per-watt over previous solutions, but is the first 'Polaris'-based Embedded GPU available in the MCM form factor to enable sleek designs in a variety of sizes, striking the optimal balance of performance and power efficiency."

Key Application Benefits of the E9170 Series

• Digital casino games: Helping reduce energy consumption and operating costs, the E9170 Series GPU cuts down on operating costs for entry level and midrange casino gaming systems, as well as lowers thermal dissipation that impacts facility air conditioning costs. Support for five independent and simultaneous displays allows companies to create an eye-catching experience with more pixels and more action. DirectX 12 support enables higher frames per second and reduced latency, and cross-platform Vulkan API support helps ensure better image quality while decreasing CPU processing bottlenecks.
• Thin clients: Thin clients powered by the E9170 Series GPU can have a smaller hardware footprint than traditional PC systems, and can have lower heat output and provide cooler workspaces in densely-staffed and space-constrained environments such as stock exchange trading floors. Additionally, 4K graphics resolution provides crisp visual clarity when thin clients are in split-screen display mode, improving the end-user experience.
• Medical displays: The high-performance 4K graphics resolution in the E9170 Series GPU helps improve the accuracy of clinical diagnoses in X-Ray machines and ultrasounds as well as with bone tissue and analysis, helping medical professionals to enhance patient care.
• Retail and digital signage: Cost-conscious signage companies integrating the E9170 Series GPU can drive up to five simultaneous 4K displays for a stunning visual experience that holds viewer attention without incurring a premium processor price.
• Industrial: The Radeon E9170 Series GPU in the reliable MCM format enables the stability required for the most stringent shock and vibration tests in environments such as airplane cockpits, in addition to driving multiple displays.

Key Features of the E9170 Series GPU

• Up to 1.25 TFLOPS at sub-40W TDP board power, setting a new benchmark within the AMD Radeon Power-Efficient Embedded GPU portfolio by delivering up to 3X the performance-per-watt over previous generations.
• 4K optimization and flexible multi-display configurability to enable rich multimedia and design versatility.
• Features include 4K HEVC/H.265iii and AVC/H.264 decode and encode, and up to five display outputs using HDMI 2.0 and/or DisplayPort 1.4.
• Available in compact MCM, standard MXM and PCI Express card form factors.
• Planned availability through 2024 to ensure support continuity for long-lifecycle products and next-generation designs.
• 4K and 3D support ensures unparalleled visual immersion.
• 4GB memory (128bit GDDR5) for high-performance needs.

Partner Support
Quixant
"Quixant is a long-term AMD partner and we have successfully implemented several generations of high-performance GPUs for various casino gaming systems," said Nick Jarmany, Chief Executive Officer, Quixant. "The new AMD Radeon GPU in the MCM form factor provides us with the competitive performance advantage we expect from AMD, while also driving down power consumption so we can deliver advanced, eye-catching content on a 4K display in a compact, cost-effective package, keeping our customers at the top of their game."

Eizo
"Over the last 10 years, Eizo Rugged Solutions has partnered with AMD Embedded to successfully bring highly reliable and rugged graphics, video capture and encoding products to several large defense programs," said Selwyn L Henriques, President and Chief Executive Officer, Eizo Rugged Solutions. "AMD's unwavering commitment to the embedded market, including long product lifecycles and GPUs with high performance-per-watt, has enabled us to offer our customers exceptional video solutions that meet or exceed their needs and minimize TCO. The new Eizo XMC and 3U VPX products, based on the AMD Embedded Radeon E9170 Series GPU, will help us deliver even higher performance and continue to expand our customer base."

Availability
The AMD Embedded Radeon E9170 Series GPU in PCI Express and MXM module configurations are planned to be available in October 2017, and the E9170 Series GPU MCM module configuration is planned for November 2017.

View at TechPowerUp Main Site

 

[damric]

That's a huge amound of GPU power for such low TDP, better than my old HD 6850s!

 

[GoldenX]

I see they are still overclocking the HD7750.

 

[Vya Domus]

I see they are still overclocking the HD7750.

Polaris*

 

[GoldenX]

Polaris*

And still same TDP than Southern Islands, with the same amount of CUs.

 

[Vya Domus]

TDP and core count comparisons are irrelevant across different architectures on different nodes. Oh well , whatever call it an overclocked 7750.

 

[GoldenX]

Same arch, GCN, just slightly refined, just like with Intel in recent years.

GCN has survived thanks to price cuts. Vega is a good example of a tired design, it's tuned way past it's efficiency to try and stay relevant.

 

[S@LEM!]

boring name for boring architecture

 

[sergionography]

Same arch, GCN, just slightly refined, just like with Intel in recent years.

GCN has survived thanks to price cuts. Vega is a good example of a tired design, it's tuned way past it's efficiency to try and stay relevant.

Vega is not GCN*
If forza 6 is any indication then vega has alot of potential. And if we follow historical trends then vega 64 will be trading blows in benchmarks with gtx1080ti once fully optimized probably in a year or so

 

[Kanan]

Same arch, GCN, just slightly refined, just like with Intel in recent years.

I wouldn't call a move from 28nm to 14nm and 3 generations difference "slightly refined".

 

[Prima.Vera]

1.2Tflops only? Jeez. Cannot even game properly on 720p....

 

[GoldenX]

I wouldn't call a move from 28nm to 14nm and 3 generations difference "slightly refined".

The IPC increase from Sandy Bridge to Kaby Lake is only 20%, and that's a jump from 32nm to 14nm, over 6 generations.
Look at Vega vs Fury IPC comparisons, GCN hasn't evolved very well. Check a review of an A12 9700 and compare it to previous APUs (that's Polaris vs GCN1.1 and 1.2), even DDR4 doesn't make a big jump to those 8CUs.

 

[Kanan]

The IPC increase from Sandy Bridge to Kaby Lake is only 20%, and that's a jump from 32nm to 14nm, over 6 generations.
Look at Vega vs Fury IPC comparisons, GCN hasn't evolved very well. Check a review of an A12 9700 and compare it to previous APUs (that's Polaris vs GCN1.1 and 1.2), even DDR4 doesn't make a big jump to those 8CUs.

a) I wouldn't compare CPUs to GPUs, it's simply not the same. The CPU space is barely moving, while the GPU space moves a lot in performance (Ryzen aside).
b) We should wait a few months or even years before making assumptions on the true performance of Vega / its IPC, it's simply too early to tell - examples such as Forza 7 are a good example why I say this.

 

[GoldenX]

Vega is not GCN*
If forza 6 is any indication then vega has alot of potential. And if we follow historical trends then vega 64 will be trading blows in benchmarks with gtx1080ti once fully optimized probably in a year or so

Vega is GCN 1.4, or 5th GCN, with more changes applied than previous versions:
https://llvm.org/docs/AMDGPUUsage.html

gfx900           amdgcn     dGPU         

    Radeon Vega Frontier Edition
    Radeon RX Vega 56
    Radeon RX Vega 64
    Radeon RX Vega 64 Liquid
    Radeon Instinct MI25

AMD won't just simply add compatibility to Vega in LLVM if it was a different arch. Same story with the drivers, both radeonsi and AMDGPU are GCN exclusive drivers, you have something new, you need new drivers.

 

[Kanan]

Vega is GCN 1.4, or 5th GCN, with more changes applied than previous versions:

Vega is GCN 2.0 by that logic, they also changed name of the CUs. GCN 1.4 (or 4th gen) is still Polaris.

It's a different architecture but not completely. Same as the difference between Sandy Bridge and Haswell or Haswell and Skylake. It's a bigger difference, but nothing entirely new.

 

[FordGT90Concept]

Embedded PCIe...wait, what?

 

[GoldenX]

Vega is GCN 2.0 by that logic, they also changed name of the CUs. GCN 1.4 (or 4th gen) is still Polaris.

It's a different architecture but not completely. Same as the difference between Sandy Bridge and Haswell or Haswell and Skylake. It's a bigger difference, but nothing entirely new.

Subtract one to use the point nomenclature, Polaris is 1.3, or GCN4.0, Volcanic was 1.2 or GCN3.0, Sea was 1.1 or GCN2.0 and Southern was 1.0.
Weir thing is, on a driver level Fiji seems the same as Polaris.

 

[Kanan]

Subtract one to use the point nomenclature, Polaris is 1.3, or GCN4.0, Volcanic was 1.2 or GCN3.0, Sea was 1.1 or GCN2.0 and Southern was 1.0.

Tahiti is GCN 1.0 (HD 7970)
R9 290 or Hawaii is GCN 1.1 or 2nd Gen
390 same.
380 or 285 is GCN 1.2 or 3rd Gen
480/580 or Polaris is GCN 1.3, true, or 4th Gen

There is no "1.4" because Vega is GCN 2.0 by that logic. Either that or 500 series is 1.4 because they slightly refined it, making higher clocks possible. Anyway, I call them by their true names given by AMD, that is GCN Gen 1/2/3/4/5.

 

[DeathtoGnomes]

I see they are still overclocking the HD7750.

Still butthurt the HD7750 got left in the dust? how many generations old is a 7750 now? 3? please troll with modern day cards.

 

[sergionography]

Vega is GCN 2.0 by that logic, they also changed name of the CUs. GCN 1.4 (or 4th gen) is still Polaris.

It's a different architecture but not completely. Same as the difference between Sandy Bridge and Haswell or Haswell and Skylake. It's a bigger difference, but nothing entirely new.

Thats true buts its somewhat more complicated than that. Mostly because the way amd build gpus is very different than nvidia for example. Amd is very modular and their gpu chips are like lego with a mix of the different parts that make up the whole chip, and with each new chip they work to improve certain parts, be it the tmus, the memory controller, the shaders etc. Nvidia on the other hand tweak aswell but also reshuffle the configuration to do the most work using the least efforts(power). And once they get to a good configuration they build every chip of that particular series using that configuration or ratio of resources, and end up with predictable and consistent set of gpus from top to bottom. This leads me to vega where amd reworked most of the parts of the gpu that its a much bigger jump than all previous GCN iterations that its practically a new architecture for the most part. Needless to say amd retains the same methodology of being able to build smaller, or bigger chips, but with different configurations/ratios of memory width/tmus/shaders etc.

 

[GoldenX]

I say GCN is getting old and all I get is free fanboy hate, thanks I guess?

GCN1.0 still is a GCN product, like Vega. A 8CU (512 shaders) Polaris is pretty much the same as an old HD7750, if not weaker if they don't use a 512/32/16 layout.
As I said, check if the new AM4 APUs are any better than the old FM2+ ones, IPC hasn't changed in recent years in GCN, only the clock frequency/power consumption balance thanks to refinements and process reduction.

You want a real change in design and performance that is not only visible in compute work? Do what GCN did to Terascale, or Maxwell to Kepler.

Did someone test if the performance of Vega in Forza is not also visible on a Fury X?

 

[Vya Domus]

GCN1.0 still is a GCN product, like Vega. A 8CU (512 shaders) Polaris is pretty much the same as an old HD7750, if not weaker if they don't use a 512/32/16 layout.

You are simply wrong. One cannot help but classify your comments as trolling.

Let's just stop spamming this thread with the same thing over and over.

 

[jigar2speed]

1.2Tflops only? Jeez. Cannot even game properly on 720p....

The usecase of this chip is not gaming but to drive displays on Medical imaging machines, Casino digital games, ATM displays, menu displays etc.

 

[TheinsanegamerN]

Vega is not GCN*
If forza 6 is any indication then vega has alot of potential. And if we follow historical trends then vega 64 will be trading blows in benchmarks with gtx1080ti once fully optimized probably in 3 years or so

FTFY. By the time vega "trades blows" with the 1080ti, volta will have already been retired in favor of whatever replaces it. Also, just because one game favors an arch does not mean that arch has tons of untapped power. If that were true, then the 1060 should be way faster then the 480 based on a few nvidia centric games.

Aint nobody but fanbois got time to wait that long for AMD to perform.

 

[Kanan]

As I see it, Forza 7 is a exception because it utilizes pretty much the whole feature set of Vega. Other games don't or mostly don't and it's a mystery how many games will be like that. Companies always need to optimize first and foremost for Nvidia that's the problem and dilemma for AMD/Vega and that means, games like Forza 7 will be rare.

GCN1.0 still is a GCN product, like Vega. A 8CU (512 shaders) Polaris is pretty much the same as an old HD7750, if not weaker if they don't use a 512/32/16 layout.
As I said, check if the new AM4 APUs are any better than the old FM2+ ones, IPC hasn't changed in recent years in GCN, only the clock frequency/power consumption balance thanks to refinements and process reduction.

How could the "same" GCN which is not the same, it was changed a lot, you simply don't happen to know the details, be weaker or on same speed with much higher clock speeds?

IPC has in fact increased, there are reviews on the net, if you don't believe me. But please research yourself I'm not here to do your homework and as I see it you have a lot of work to do and I won't do it for you.

You want a real change in design and performance that is not only visible in compute work? Do what GCN did to Terascale, or Maxwell to Kepler.

I concur to that though. It is true that AMD lacks a proper gaming engine without compute nonsense gamers 99% don't need anyway. Vega / Fury X etc. are not gaming engines, they are all multi purpose engines comparable to Kepler/Fermi/Tesla but not to Maxwell or Pascal (consumer variant). On top of that it's obvious Fury X and also Vega lack ROPs, 64 aren't enough, it's also the reason why after 1440p the card is falling in performance whereas the 1080 Ti is still rising. Reminder: Hawaii also had 64 ROPs, a GPU with a lot less shaders 2816 compared to 4096 each. The design of Fiji / Vega is simply not balanced.

Did someone test if the performance of Vega in Forza is not also visible on a Fury X?

Fury X was included in the same test and also had very high performance rivaling that of GTX 1070 which represents a overclocked GTX 980 Ti - suffice to say, Fury X normally had no chance against a overclocked 980 Ti.