HBM and HBF Speculation

[sm625]

In theory AMD isnt even limited to stacking DRAM. Instead of having 4 (or 8) 1GB stacks, they could go with 3 (or 7), and use the last spot for a 64GB-128GB high bandwidth flash die (HBF). Games could load their entire texture library onto that flash and memory access would still be several times faster than loading that data over the PCIe bus.

Once you remove the need to load data over the PCIe bus, you no longer need a lot of actual VRAM. You just load the entire game onto the video card's internal flash and be done with it. I doubt they will do this for this generation, but I would expect it eventually. It would also make one monster of a SoC once they make an APU version. Give me a SoC with 4 CPU cores, 2000 shaders, 7GB of unified HBM, and 128GB of HBF for my OS and whatever game I'm playing at the moment, and you're looking at an incredibly powerful and ultraportable gaming solution. At the very least, I absolutely expect the next console to have this type of design. But there is no reason AMD couldnt package it up and sell it as their own branded laptop.

It also opens the door to a real external GPU solution, since bandwith over a USB 2.0 bus would be sufficient to drive the thing once all the actual graphics data resides on the video card's onboard flash. AMD has a real chance here to revolutionize gaming. I'm talking about a real external GPU solution, not some pie in the sky lightning thunder nonsense that never had a chance to succeed because its too proprietary and expensive. Imagine docking your ultrabook into a cheap and simple Radeon external GPU, downloading your game onto it, and then getting a massive fps boost!

Of course, AMD also has the chance of completely bulldozing on HBF.

 

[tential]

32gb would be enough? 64-128 sounds like overkill.

 

[xthetenth]

64-128 would be nice if you wanted to futureproof or do a RAID 0 equivalent across chips. For actually practical things it's likely excessive.

That would potentially be extra cool though. Heck, for consoles imagine having literally the entire memory and storage space physically on the die.

 

[sm625]

32gb would be enough? 64-128 sounds like overkill.

There are games that are already using more than 32GB of HDD space. I just figured it would be super easy and dummy proof to load the whole thing once. It would make loading times hella fast (after the initial one-time load). Granted, certain things wouldnt need to be stored on there, such as unrendered scenes. Those could just be streamed off the HDD. But I imagine all the audio would have to fit onto the HBF too. But maybe not. That could save a few hundred MB.

 

[xthetenth]

I'd be worried if they made a game that used literally all of the texture and geometry data at once.

 

[geoxile]

That's an interesting idea, but would it be fast enough to make it reasonable on a GPU?

I do recall both Samsung and AMD suggesting various multi-tier memory hierarchies for HPC clusters, one of which sounds similar to this: high bandwidth memory + high capacity memory (flash being one of the proposed memories).

 

[ShintaiDK]

Makes zero sense.

And you need quite a flash array plus controller to get over 8GB/sec.

Not to mention nobody would buy your "monster soc" to begin with.

 

[swilli89]

Makes zero sense.

And you need quite a flash array plus controller to get over 8GB/sec.

Not to mention nobody would buy your "monster soc" to begin with.

Plenty would. Ideal candidate for a small form factor steam machine.

 

[ShintaiDK]

Plenty would. Ideal candidate for a small form factor steam machine.

Define plenty while setting up the business plans pros, cons and price. Current APUs is a sales disaster. And desktop APUs is an even bigger disaster as the numbers show each time with 20-30% sales decrease every quarter.

 

[sm625]

Current APUs are a disaster because they bring no benefits to integration, and add some drawbacks. The reason you integrate is to get the same performance onto a smaller form factor, to reduce the BOM and enable more form factors. But AMD APUs still require basically the same size motherboard. They just recently got rid of the southbridge. That should have been gone long ago. And the DRAM performance went way down, dragging down performance, making them completely pointless. An APU simply has to have HBM to be worth anything. Once it does, then the loss of fps per compute core (vs discrete) disappears and actually reverses. The HBF would be on a 128 or 256 bit bus, so 8 GB/s would be very easy target to hit, and the latency would be cut way down, similar to an ULLtraDIMMs. It is going to happen one way or another. If you think its a bad idea then you think smartphone SoCs are a bad idea and we should just go back to using Gordon Gekko bricks. The only question is will AMD profit from this transition or will they continue to drown.

 

[ShintaiDK]

128 or 256bit doesnt matter if the NAND and/or controller isnt fast enough. Even the fastest SSD you can buy for money today cant saturate a PCIe x16.

And the business plan needs to work. You can buy discrete+CPU for the same or lower price than an APU today. And its not the graphics performance thats the reason APUs dont sell.

 

[therealnickdanger]

A lot of the bloat of modern games is the uncompressed audio. It's good bloat IMO, but still bloat. No reason to load audio files into GPU flash, even with TrueAudio. That would cut down on a lot of storage space required.

 

[sm625]

On a game that takes up 40GB of space, like Witcher 3, how much of that space is audio?

 

[ShintaiDK]

On a game that takes up 40GB of space, like Witcher 3, how much of that space is audio?

Witcher doesnt take up 40GB. Its 26GB. And about 5GB of that is textures. 6GB or so goes to audio.

 

[duskkk]

Realistically few people would ever need more than 16gb unified memory.
An APU with 8 or 16GB HBM unified in a notebook, there would not be any need for additional memory. The whole thing would fit on a tiny logicboard which leaves plenty of space for heatsinks, fans and battery.
Seems ideal for mobile SoCs. Since RAM comes soldered already quite often it would not be much of a change and 16GB of HBM shouldn't be too expensive. That really needs to happen soon not just with AMD, Intel could easily make it happen and it would be way better than their 128MB cache.

 

[ShintaiDK]

Its quite clear we get CPUs in the future with fixed memory sizes we cant upgrade for the mainstream and down. But prices needs to come down on HBM/HMC first and densities up. A good guess would be around 2020.

 

[sm625]

It's happening! Well, sort of.

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

It's not HBF, but it is nice to see that they are putting flash on a graphics card, finally.

 

[IEC]

Yep, good call on this one. It has some good practical uses and I can see it being easily worth the cost of entry.