AMD’s Radeon Pro SSG Features 1TB Of GPU Memory, Unheard-Of Rendering Capability

Status
Not open for further replies.
Interesting. I bet amazon and Netflix will be getting these dev kits. Re scaling video on the fly is a real boon for both the server and client. It saves the host company on bandwidth nd improved the rendering quality on the client...win win for large libraries.
 

Beholder88

Honorable
Oct 14, 2013
38
0
10,530
I'm curious to see how this thing is configured on the board level. Will there be any VRAM at all, or is the SSD acting AS the VRAM? Correct me if I'm wrong here, but HBM memory has a bandwidth of 256Mbps and SSDs can see more than twice that. Depending on the interface and controller being used this seems like a really good idea. Very obviously not for gaming, but the use cases could expand with advancement of the design. I like it, hopefully this give team red a bit of a boost.
 

Kelgrin

Commendable
May 9, 2016
3
0
1,510
Maybe its just the fact that I do not deal with video editing or rendering, but the fact that the article talks about one machine rendering in FPS and the other in Hz confuses me. For me, it makes it difficult to compare the speed of the two machines since they appear to be using different measurements. I would assume the two (FPS and Hz) are not equal to each other. If they were, then I don't understand why there was a need to use two different terms to describe the same thing.
 

kenjitamura

Distinguished
Jan 3, 2012
195
3
18,695
They should have demonstrated a different use case because editing 8k video real-time is a non-issue with proxy clips. Should have shown it doing something like rendering a massive 3d model scene which this should also benefit really well.
 

bit_user

Polypheme
Ambassador
Even most SSDs have DRAM buffers. Cheap SSDs are starting to use system memory as a buffer, but 'till now it's been embedded in the drives.

Okay, you're wrong. AMD's Fury is spec'd at about 512 GB/sec. SATA 3 tops out at about 500 MB/sec. You're only off by about 1000x. Okay, so M.2 SSDs can do over 2 GB/sec. Still, not even close.

Only 3D XPoint gets close to the speed of DRAM, but DDR4 is still several times faster. I don't know about Memristor-based memory, but I think it might be in the ballpark.

So, I think we can safely say this thing has HBM, GDDR5, or GDDR5X. I wish they posted some other specs on it, but you can't really blame them. It's not actually a product launch.

And I'm not too impressed with 8k video "rendering". I mean it's good, but I'm sure some people read that as 3D rendering at 8k, which it's not.
 

bit_user

Polypheme
Ambassador
If you give a professional the option of using proxies or native, all else being equal, they'll go native every time. It eliminates the issue of things you don't notice with the proxy, or artifacts that are only in the proxy.

Especially when you start getting into effects, you want to see the final render in realtime, so you don't have to go back and fix it later.
 

SwissBeatz

Commendable
Jul 26, 2016
1
0
1,510
Hoooah! 1TB of Gpu memory for buffer!???.. Well, for a multimedia company or for any digital content creator like me whos into hi res digital film editing, photography and 3d renderings this video card is a HEAVEN SENT! With that amount of rendering power, my high end 64 gb ram, dual Intel xeon e7-8891 2x10 core (40 thread total) processor workstation / gaming rig with gigabyte custom air cooled geforce gtx 980 wtih 6gb of ram would perform like a cheap ass $800 dollar workstation when compared to this. I hope this would be priced as Amd's high end workstation cards so that buying one of these is enough to match a very pricey performance of four high end Intel xeon 20 core seperate workstations for rendering realistic 3d designs / animations and editing long 4k videos with lots of complex effects, this will SURELY CUT THE AMOUNT of rendering times (that usually took days to complete) specially on complex 3d animation projects and will definitely increase anyones productivity as far as digital content creation is concerened.
 

alextheblue

Distinguished


Correct, and the article needs to be corrected (and retitled). However, I will say that it isn't a slow SSD cache. They're using a pair of 950 Pro SSDs (PCIe 3.0 x4 M.2) in RAID 0. In addition this is directly tied to the GPU, so it's actually quite darn speedy for a cache. The SSD array can feed the GPU very rapidly from very large fixed sets of data. RAW (uncompressed) 8K video just happened to be a good demo of what you can do when not restricted by a conventional chipset-linked SSD setup.
 
Re M.2 SSD->
I suspect these "10x" latency or whatever claims turn out to be much, much less. Put another way, when it's common to see the GPU at near 100% load how could adding local SSD as an additional buffer help much?

It probably does, but even so as programs get better at swapping data between system memory and GPU memory we should see minimal benefit to this approach.

Caching is always like this and happens at many levels. It helps until it doesn't, and then more is pointless.
 

bit_user

Polypheme
Ambassador
Well, if the GPU could go directly to the SSD, then I'd agree that there's little benefit in putting it onboard. But, since the SSD might not be on the same bus and due to various configuration headaches, the reality is that the GPU would have to go through the host OS to access data on it. And that would be many times slower. I'd say 10x would even be optimistic.

I think your error is trying to use applications you know as a basis for comparison, here. This wasn't done to speed up things that GPUs are already optimized for. This was done to speed up things where the GPU isn't traditionally the bottleneck. It's a good idea.
 

JakeWearingKhakis

Commendable
Apr 5, 2016
153
0
1,690
I think the SSDs inside of this thing should give this product a longer warranty. Since SSDs have a lifetime measured in terabytes written, and using it as a cache for graphics I'd assume it will reach that limit much quicker than usual? Correct me if I'm wrong.
 
"Sure, these demos always favor the product being shown, but assuming there are no games being played here, let that 90Hz sink in for a second.
The Radeon Pro SSG appears to be capable of faster-than-real-time rendering of 8K raw footage. The card was rendering 4.5GB/s at that pace."

Can someone show me the math there? 7680pxX4320pxX90hz gives me 2.9Gpx/sec assuming non-hdr 24bit color depth that gives me 71Gb/s or 8.9GB/s which is only the transfer rate too support that uncompressed display resolution, but wouldn't the card need to render each of those pixels? I can't imagine it compressing to nearly 50% and rendering like that, but I don't know all the things that go on inside the GPU.
 

bit_user

Polypheme
Ambassador
If they're using 4:2:0 @ 8-bits then, I get 4,478,976,000 bytes/sec. That rounds to 4.5 GB/sec.
 
Status
Not open for further replies.