AMD Reveals More Vega GPU Architecture Details

[eydee]

Vega when running DOOM on Windows is estimated to perform between a GTX 1070 and GTX 1080.

With the 1080Ti coming, it would be very unfortunate. Unless it's a lot cheaper of course. But what's the hype for then?

The linked article says air ventilation was a big issue with the test machines. It may have been downclocking automatically to avoid overheating, or may have been artificially nerfed. We'll see in a little while.

Also Vega seems to have new features over Polaris. Will there be a lower-end Vega?

[Ansla]

c) whole Linux distribution in it or in it or just EFI to boot from network storage on vc33 slide at

http://diit.cz/clanek/kompletni-prezentace-architektury-vega

An Early Port Of GCC To AMD's GCN Architecture

Don't get your hopes to high, that was just an overly sensational title, gcc will just gain a backend for GCN just like LLVM already has, there will be no Linux running on the GPU in the near future.

[Passso]

With the 1080Ti coming, it would be very unfortunate. Unless it's a lot cheaper of course. But what's the hype for then?

The linked article says air ventilation was a big issue with the test machines. It may have been downclocking automatically to avoid overheating, or may have been artificially nerfed. We'll see in a little while.

Also Vega seems to have new features over Polaris. Will there be a lower-end Vega?

As the 1080 is way too expensive and dedicated to end costumers, an AMD card like the 1070, sold for less than 300$ is a good deal imho.
My only doubt concerns the driver... will amdgpu be mature enough to have this level of performance under Linux? That would be great! If not I would stick to NVidia...

[eydee]

As the 1080 is way too expensive and dedicated to end costumers, an AMD card like the 1070, sold for less than 300$ is a good deal imho.
My only doubt concerns the driver... will amdgpu be mature enough to have this level of performance under Linux? That would be great! If not I would stick to NVidia...

Just don't forget the RX 480 fiasco. Marketing is about $200, sold for 300-350€ actually. People are already sticking with NV.

[starshipeleven]

Just don't forget the RX 480 fiasco. Marketing is about $200, sold for 300-350€ actually.

In a part of the world where NVIDIA prices also translate like that (Euros = Dollars + taxes + wtf + give-me-your-money-and-fuck-you), let's not forget this trivial detail.

For example, a NVIDIA 1080 goes from 650 $ to 750+ euros.

People are already sticking with NV.

Lololololololololol keep this up.

[srkelley5]

Here is where I saw it from someone who had first-hand play time with it: http://techreport.com/review/31224/t...architecture/4

That's pretty cool. I think that actual performance of the release cards will be significantly higher than the estimate given. With true ventilation and optimized drivers we should see better performance from real world machines. That's actually pretty amazing. That level of performance with the heat vents taped up and using drivers that most likely are still beta level at best? Nice.

[L_A_G]

Could someone tell me what the benefit is of being able to interface with NAND flash is please?

Speeds up pretty much everything where you have to read and write data onto disc during computation. You don't have to stop the computation and ask the processor to load the data that you need into device local memory and can instead fetch the data directly.

As for real world applications this will help I can come to think of games with really big and/or detailed worlds, a wide variety of high performance computing tasks, video editing and rendering, high detail 3D modelling and animation, high detail CAD/CAM and various machine learning systems.

I guess it's a good thing for me that this didn't come out years ago when I started my Master's Thesis project. The main purpose of that project was to expand the functionality of a GPGPU high performance computing application using OpenCL so that it could compute data sets so large they wouldn't be able to fit into device memory all at once. In this case it was a job one where you couldn't just run the data set in separate runs as data points at one edge of the data set could affect the results for data points at the other edge of the data set. I had to implement a solution that split the job into multiple segments and then re-ran the segments whenever data "spilled over" from one segment to another.

Joined the competition on ve.ga and if I win one of these cards, I might adapt that application (technically I don't own the source code because I did it as a paid project for a government organization, but I still have the source code and made some bug fixes to it this summer) to use this direct access to disc rather than the multi-GPU solution that I created for my Master's Thesis.

[koneu]

So nobody else is confused about 512 TB of virtual address space?
Firstly, why 512 TB?
Secondly, isn't VAS defined on a different abstraction layer?

Also thanks for that link, "Make. Some. Noise." really made my day.

[CrystalGamma]

Just don't forget the RX 480 fiasco. Marketing is about $200, sold for 300-350€ actually. People are already sticking with NV.

I'd have to check my invoice, but I think my RX480 cost me about 250€, which may not be as cheap as $200 (even subtracting taxes), but is definitely less than 300€.

[pal666]

Could someone tell me what the benefit is of being able to interface with NAND flash is please?

http://www.pcworld.com/article/3099964/hardware/amds-new-ssg-technology-adds-an-ssd-to-its-gpu.html