New AMD Memory Tech ?

Dicehunter · 31 Mar 2019 at 09:03

So was talking to someone in a little gaming group I'm in, He's in America and shall remain anonymous but he's claimed several times on different occasions to have worked with AMD on being able to use system memory in lieu of GPU memory to "augment" a memory starved GPU.

It sounds a bit like HBCC that was introduced with Vega but apparently works on older cards like his dual 6990 setup, I'm taking it as a pinch of salt right now but would be interesting to see if it is true.

@LtMatt Would be interesting to get your take on this considering he's claimed his little dev team worked directly with AMD which is a pretty big claim.

james.miller · 31 Mar 2019 at 09:38

Why would they bother? Where would this be an advantage?

Dicehunter · 31 Mar 2019 at 09:42

james.miller said:
Why would they bother? Where would this be an advantage?

Could be a selling pint for the AMD brand as a whole "Your GPU's life span will be longer with our memory tech" or something to that extent.

d_brennen · 31 Mar 2019 at 09:48

Remember sideport RAM? Probably something like this. Companies always R&D even when it's unlikely to come to fruition, important discoveries are made that might be transferable and experience gained by techs etc

MasterOC · 31 Mar 2019 at 09:51

i dont trust anyone with a pc that looks like that

Dicehunter · 31 Mar 2019 at 10:03

MasterOC said:
i dont trust anyone with a pc that looks like that

Well he told me his company is Janssens Dynamic which makes a real time 2d/3d conversion tool called ViewVerge but when I asked him about specifics about the hardware and GPU chipset changes he claims to have made he got very aggressive.

LeMson · 31 Mar 2019 at 10:13

A ram disk is fast compared with HDD, SSD, not tried one vs m.2.

But versus any recent gfx ram on a somewhat decent card, it'll be slow

MasterOC · 31 Mar 2019 at 10:16

Dicehunter said:
Well he told me his company is Janssens Dynamic which makes a real time 2d/3d conversion tool called ViewVerge but when I asked him about specifics about the hardware and GPU chipset changes he claims to have made he got very aggressive.

Not surprising , Did a quick search on google , ''patents pending'' with no application's available or proof or concepts since the conception with a failed kick start in 2016... obviously a dreamer

gpuerrilla · 31 Mar 2019 at 10:24

They pushed RAMDisk years ago but something to compliment gpu mem would be handy.

bru · 31 Mar 2019 at 10:27

Isn't this what the HSBC thing does with the HBM memory on the FUJI cards onwards?

Dicehunter · 31 Mar 2019 at 10:33

bru said:
Isn't this what the HSBC thing does with the HBM memory on the FUJI cards onwards?

HBCC *high bandwidth cache controller* it was introduced with Vega and can help give higher minmums and average FPS as well as smooth out memory intense operations but he's claiming his companies implementation can "add" VRAM by using system memory.

bru · 31 Mar 2019 at 10:35

Ah yes that's the one, was it vega I thought it was earlier than that, your probably right as my memory isn't as good as it was.

Dicehunter · 31 Mar 2019 at 10:36

bru said:
Ah yes that's the one, was it vega I thought it was earlier than that, your probably right as my memory isn't as good as it was.

It may have been earlier that they talked about it but my only knowledge on it is from Raja Koduri hyping it up prior to releasing Vega 56 & 64.

mmj_uk · 31 Mar 2019 at 11:06

This has been possible for decades, with AGP you used to be able set an "AGP aperture" which was a fixed amount of system RAM for GPU usage should you run out of local video memory. There were some old low end Geforce cards that used to ship with hardly any local memory at all, with the idea being that they just had a small amount of memory (32MB?) as a buffer and primarily used 256-512MB of system memory but I've forgot what they called it. Onboard GPU's have used system memory since the dawn of time. It's nothing new but the problem has always been that using non-local memory tanks performance because the motherboard data buses and system memory are far slower than what's on your average high end graphics card.

The only difference here seems to be they're talking about a modified chipset (presumably to minimise/remove bottlenecks) which they say is expensive, if it wasn't expensive we'd have probably seen this technology 20 years ago but big buses and fast memory has always been expensive.

Davedree · 31 Mar 2019 at 16:51

bru said:
Ah yes that's the one, was it vega I thought it was earlier than that, your probably right as my memory isn't as good as it was.

You are correct Bru, it was Fiji and it was a very primitive Shortlived Demotech.
Radeon pro SSG, it had x2 m.2 slots for additional dataset storage/access.

EastCoastHandle · 31 Mar 2019 at 22:12

This sort of stuff was talked about when we had ATI. Just before AMD bought them. I've not heard anything more about this since then.
This is the stuff of ole to be honest. As it stands now AMD is working on system memory via GDDR5 and HBM. Which I thought would incorporate this.
That would be an expensive MB. However, why wouldn't AMD create an entire PC ecosystem made for...AMD?

4K8KW10 · 31 Mar 2019 at 22:18

AMD isn't interesting now.
Look at intel's 32 GB 4D Xpoint with 8 TB/s bandwidth :eek:

Skynet5 · 31 Mar 2019 at 22:24

Sounds like that old solution that allowed you to share RAM with graphics card. Had it on an awful HP laptop in 2003. All I could afford but worst £900 I ever spent.

bru · 1 Apr 2019 at 02:56

4K8KW10 said:
AMD isn't interesting now.
Look at intel's 32 GB 4D Xpoint with 8 TB/s bandwidth

While your looking, check out the date.

Rroff · 1 Apr 2019 at 03:04

I've no idea if he is talking BS or not but stuff like that has been done - it generally isn't particularly useful outside of very specific situations - nVidia did something similar for a bespoke Mosaic setup IIRC near 20 years ago and there were simulators built that used 8x ATI cards (off the top of my head x1800) that did something similar.

Even with a modified driver and some crazy fast RAM you can't just replace or add onto VRAM seamlessly and retain the same performance - usually far too much latency involved even when bandwidth is reasonable and in most cases system DRAM is both far too slow and different in terms of how it is accessed to supplement VRAM directly without major performance impact.

GPUs generally page out to system RAM and even storage (pagefile) if they run out of VRAM anyhow but it ain't pretty especially as it is handled indiscriminately with no knowledge of what the GPU data is being used for so as to best optimise it for minimal performance penalty.