We have a SSD attached to a GPU now.

[Maverick177]

http://www.anandtech.com/show/10518/amd-announces-radeon-pro-ssg-polaris-with-m2-ssds-onboard

 

[Stuka87]

Its a cool idea, for the types of workloads that require tons of memory, the memory does not have to be super fast.

 

[master_shake_]

Its a cool idea, for the types of workloads that require tons of memory, the memory does not have to be super fast.

shouldn't it be as fast as what is already on the card?

 

[Elixer]

Only $10,000, I wonder how much the founder edition of the card will cost? (yeah, I know this is a dev kit)
All kidding aside, this one is also colored opal blue, which looks pretty nice.
And it seem like we are turning back the time machine, since I know that video cards of the past, you can add to it memory, it was just how it was done.
Interesting that they are now using SSDs (which have ECC and all that good stuff) for some kind of a cache usage.

 

[IEC]

I can think of a few immediate practical applications for this at my former company. I'm sure there will be at least one ordered there as a dev kit.

 

[ZippZ]

Semiaccurate has some great scenarios where it would be a massive help...

"...How about 8K movie streaming and cleanup in realtime. At 96FPS. Sure you can do this with traditional methods but the best of them will run the same task at 17FPS."

"Hugely complex and highly detailed CAD models that took the better part of an hour to load up and decompress will still take the better part of an hour to load up and decompress on an SSG GPU based system. Why bother? Because it takes the better part of an hour to load and decompress the first time, then it can stay resident on the GPU’s flash storage. The second time it should take seconds. "

http://semiaccurate.com/2016/07/25/amd-puts-massive-ssds-gpus-calls-ssg/

 

[Flapdrol1337]

How is this supposed to be faster than the cpu ram?

 

[sirmo]

I can definitely see applications for this.. GPUs with thousands of cores can crunch equal amounts of of simultaneous streams.. not being bound to a CPU or the OS stack, allows for a much lowered overhead..

It's definitely a neat technology and yet another AMD first.

 

[Dribble]

"Hugely complex and highly detailed CAD models that took the better part of an hour to load up and decompress will still take the better part of an hour to load up and decompress on an SSG GPU based system. Why bother? Because it takes the better part of an hour to load and decompress the first time, then it can stay resident on the GPU’s flash storage. The second time it should take seconds. "

As someone who works in that area I can categorically say that's rubbish. It might take ages to load, but it's loading to a program that does all sorts of stuff, not just display pictures. Rendering is just a small part of the application. Yes if you have some big design review you will load up your data first, but it's the whole program you need to load up. Then it stays resident in main memory so available in seconds.

 

[selni]

How is this supposed to be faster than the cpu ram?

It probably won't be and Ryan comments on this in the article, but presumably you don't have as much storage there.

 

[stuff_me_good]

Oops... did AMD just caught nvidia with their pants down? The whole market for themselves... free money, woop woop.

 

[antihelten]

Oops... did AMD just caught nvidia with their pants down? The whole market for themselves... free money, woop woop.

These cards aren't launched officially until 2017 (only developer versions available currently). So it's probably a bit premature to start saying company X or Y has the market for themselves.

 

[Yakk]

1TB+ of storage accessible at over 4,500GB/s AND not saturating the system CPU, RAM, and system bus. All of these are essentially untouched by the massive amounts of data being processed.

All kinds of awesome here!

 

[NTMBK]

Finally, a card that can run Shadow of Mordor smoothly with max texture quality.

 

[selni]

1TB+ of storage accessible at over 4,500GB/s AND not saturating the system CPU, RAM, and system bus. All of these are essentially untouched by the massive amounts of data being processed.

All kinds of awesome here!

You probably mean 4.5 GB/s, 4500GB/s would indeed be all kinds of awesome.

 

[therealnickdanger]

1TB+ of storage accessible at over 4,500GB/s AND not saturating the system CPU, RAM, and system bus. All of these are essentially untouched by the massive amounts of data being processed.

All kinds of awesome here!

Yeah, definitely a game changer for video professionals. Better than real-time 8K video rendering in such a small package is monumental. I've worked with some multi-TB lidar point cloud data and CAD files that could certainly benefit from this as well.

 

[jhu]

Now it just needs USB ports for keyboard/mouse and a Linux port and it can be your desktop computer.

 

[boozzer]

Now it just needs USB ports for keyboard/mouse and a Linux port and it can be your desktop computer.

haha, why not

 

[Stuka87]

shouldn't it be as fast as what is already on the card?

No, definitely not. GDDR5 is *WAY* faster than what an M.2 slot can provide. But for this work this card will do, it doesnt need to be.

 

[guachi]

I know nothing about the uses of such a card, but I hope it's a success and leads to something more elegant in the future.

 

[Jaskalas]

Storage on my GPU?

Someone will have to convince me why I'd want this, but maybe the product isn't for a typical home user or gamer.

 

[IEC]

Storage on my GPU?

Someone will have to convince me why I'd want this, but maybe the product isn't for a typical home user or gamer.

It isn't for the typical home user or gamer. Certainly not at a $9999 introductory price for the dev kit

 

[therealnickdanger]

Storage on my GPU? Someone will have to convince me why I'd want this,

If you don't already know, it's not for you.

but maybe the product isn't for a typical home user or gamer.

Bingo.

This GPU is specifically built for application requiring direct communication of many, many GBs of fast storage completely addressable by the GPU internally. RAM would be ideal from a speed perspective, but the pricing for 1TB (or more) RAM on a GPU would be astronomical. On the other hand, SSDs with reduced obstructions are very quick (compare NVMe v. SATA).

You can call it a stopgap for something greater if you want - but for a number of markets, it will be a very, very popular stopgap. Video, CAD, and Lidar are the three that come to mind.

 

[Headfoot]

I dont get all the confusion. It's local SSD storage. It's just basic computer science -- closer storage = faster storage all else being equal, and caches come in heirarchies.

L1 > L2 > L3 > (eDRAM / eSRAM / SRAM / L4) > Local high speed memory > Non-local high speed memory over fast bus > Local fast bus SSD > Nonlocal fast bus SSD > Non local HDD.

Of course having a terabyte of L1 cache would be incredibly fast. Also incredibly impractical. In fact, a terabyte of anything doesn't even approach reality until you hit multi-socket massive DRAM arrays, and that's the absolute high end. This is why 3dXpoint / HMC / HBM are such game changers. It brings significantly larger amounts of capacity closer on a wider/faster bus.

In short, this is for data sets that are worked on by the GPU that are so large that can't get close to fitting in VRAM and RAM combined.

 

[Shivansps]

I dont understand how this will work, its an SSD where only the GPU has access to? Otherwise it will just go over PCI-E like any other M2, latency will be lower trought.