Radeon Pro SSG - 1TB GPU

[Macker]

Check this out

The RAM limitations of GPU's are the sole reason we haven't switched our workflow over to them, but this looks like it could potentially be a game changer does it not?

Vlado & the team, what are your thoughts? Will you get a developer kit?

[benb32]

Sadly not The SSDS just appear as additional drives available to the OS. The only advantage is the SSDs connected to the card are directly connected to the PCIe lanes to the CPU and not an intermediate chipset. Will require the graphics application to use custom APIs to stage data in the SSDs to access it and other than reduced latency is still slower than accessing system RAM. The card might be available in 2017 if you want it sooner a development version will cost $9999!

[Macker]

I am aware that it is going to be slower than accessing RAM, but given that most if not all of my scenes do not fit into VRAM then it is still preferable. Surely the SSD's show up as part of the graphics cards VRAM?

from AMD's website:

"With this new GPU technology, a one terabyte extended frame buffer is dedicated to support the GPU. This enables much larger datasets to be loaded locally, connected over a dedicated PCIe® bus. When the GPU requests content, it looks first in the local frame buffer and only needs to involve the CPU if the data is not already in the extended frame buffer."

[benb32]

No that's the point - they are advertised as SSDs with NVMe. Is no different to adding an PCIe SSD but writing custom code to cache things on that SSD. Data requests must still be marshalled by the NVMe controller and CPU - not directly between the GPU and SSD Realistically the PCIE bus is still much slower as well as increased latenence. I do appreciate that as a solution it would be quicker than an SSD drive attached via SATA or PCIe lanes outside those provided directly by the CPU.

[savage309]

Will you get a developer kit?

We will try to, yes.
Hard to tell how well will it work for our needs before testing.

Best,
Blago.

[Macker]

The article seems to suggest that it does bypass the current issues with vram as the bridge chip is on the GPU...

"The interesting card, however, is the Radeon Pro SSG, a new Radeon Pro GPU with integrated M.2 slots for adding PCI Express-based NAND storage. According to Raja Koduri, the GPU can add up to 1TB of SSD storage connected via a PEX8747 bridge chip. According to AMD, adding a 1TB switch to the GPU is a simple way to bypass a common problem: the GPU memory limit."

[super gnu]

yes but the question i have is how is this going to be any faster than going over pcie 16x to the main memory? its perfectly possible now to bypass the gpu memory limit, but going over pcie to system ram is so much slower than the local gddr5 storage that it kills performance badly. since a pair of nvme drives will give you max around 5GB /sec, thats much *slower* than the pcie bus/system ram...

basically id have much preferred a gpu with some ddr4 slots on and a fancy 8 channel memory controller.

[benb32]

The article seems to suggest that it does bypass the current issues with vram as the bridge chip is on the GPU...

"The interesting card, however, is the Radeon Pro SSG, a new Radeon Pro GPU with integrated M.2 slots for adding PCI Express-based NAND storage. According to Raja Koduri, the GPU can add up to 1TB of SSD storage connected via a PEX8747 bridge chip. According to AMD, adding a 1TB switch to the GPU is a simple way to bypass a common problem: the GPU memory limit."

You're right it does bypass any bridge chips on the motherboard, but there would still be latency introduced by the bridge on the graphics card.

This article has an update describing a bit more:
http://www.anandtech.com/show/10518/...2-ssds-onboard

[Macker]

So currently when VRay runs out of memory on the GPU what happens? My understanding was that it was cached to the HDD/SSD, not memory? If it is cached to memory, then this card isn't as exciting as I thought it was.

[Svetlozar Draganov]

Currently when V-Ray RT runs out of memory it prints "out or memory error" and stops rendering.
We are working on improving V-Ray RT GPU memory management and future version will be able to render more data.

[Macker]

We will try to, yes

Did you ever get hold of a dev kit? Has anybody heard any more news on this?

[savage309]

Did you ever get hold of a dev kit? Has anybody heard any more news on this?

We don't have nothing to share yet.

Best,
Blago.

[mitviz]

what about an out of core option like octane?

[savage309]

what about an out of core option like octane?

Real out of core rendering (like real unbiased rendering) slows down too much. What is usually done is optimize the memory needed by the render, so it can fit more memory than it used to. In the nightlies we have optimized the texture memory needed for some cases, also we are working on a thing that would allow to reduce the memory even more. I hope we manage to get it working fine for the next SP, but we will see.

Best,
Blago.

[mitviz]

Real out of core rendering (like real unbiased rendering) slows down too much. What is usually done is optimize the memory needed by the render, so it can fit more memory than it used to. In the nightlies we have optimized the texture memory needed for some cases, also we are working on a thing that would allow to reduce the memory even more. I hope we manage to get it working fine for the next SP, but we will see.

Best,
Blago.

uff, sounds like you guys are ontop of things