XGMI.....I thought it was some new sort of sex thing.

Clawedge:

XGMI.....I thought it was some new sort of sex thing.

It can be... gigitty Looks like SLI/Crossfire is back in a new form

Not surprising, AMD slides shown infinity fabric between CPU, APU and GPU.

Will be nice if this connection can finally show multiple gpu's as if they where one at a hardware level eliminating the need for software implementation in multi GPU.

icedman:

Will be nice if this connection can finally show multiple gpu's as if they where one at a hardware level eliminating the need for software implementation in multi GPU.

Not enough to make 2 cards into one. Same issue as with nVLink. But likely good enough to do massive linking of 2 GPUs on one PCB... similar to connection between 2 CPU sockets. But I would rather seen MCM solution, there you do not need to pretend or overrides.

...and I can bet the connector won't cost 80 euro either.

Can it overcome the bad support of CF for games? or only Game Dev can?

Does it Really matter? I am a Huge SLI/CF supporter but the games are just nos supporting SLI anymore, is declining more and more, the only recent title I can think supports sli is FC5 nothing else

AlbertX:

Does it Really matter? I am a Huge SLI/CF supporter but the games are just nos supporting SLI anymore, is declining more and more, the only recent title I can think supports sli is FC5 nothing else

my last two rigs were sli but when I moved to the 1080ti I said no more due to poor game support

AlbertX:

Does it Really matter? I am a Huge SLI/CF supporter but the games are just nos supporting SLI anymore, is declining more and more, the only recent title I can think supports sli is FC5 nothing else

For gaming, probably not.. but Vega is also used in datacenter products and in those applications it matters big time.

if has 4096-bit HBM2 higher clocks on VRAM/GPU and lower power draw, is that really matters now, Nvidia is over 50% higher performance per watt greetings

Amx85:

if has 4096-bit HBM2 higher clocks on VRAM/GPU and lower power draw, is that really matters now, Nvidia is over 50% higher performance per watt greetings

Comparing 2 unreleased chips/architectures. Should we laugh or cry?

Amx85:

if has 4096-bit HBM2 higher clocks on VRAM/GPU and lower power draw, is that really matters now, Nvidia is over 50% higher performance per watt greetings

If that's what AMD has got lets hope it make a huge difference as the GPU market really needs AMD maybe there not interested in high end and hope to dominate the mid range

Spider4423:

It can be... gigitty Looks like SLI/Crossfire is back in a new form

HWgeek:

Can it overcome the bad support of CF for games? or only Game Dev can?

AlbertX:

Does it Really matter? I am a Huge SLI/CF supporter but the games are just nos supporting SLI anymore, is declining more and more, the only recent title I can think supports sli is FC5 nothing else

Barry J:

my last two rigs were sli but when I moved to the 1080ti I said no more due to poor game support

This feature is mainly for enterprises running deep learning software, and Vega20 itself is also an enterprise focused product which I don't think will be released for the gaming market.

Fox2232:

Comparing 2 unreleased chips/architectures. Should we laugh or cry?

im comparing actual GTX1080Ti to actual VEGA lol and expecting, VEGA20 🙄

I really think that we cannot expect anything from AMD at least for 2 years, i hope Navi architecture will be created from ground up... a very influential thing is that GPUs are not only to game anymore, and that is changing GPUs development everywhere.

Vega 20 is, to my understanding, first targeted toward servers. So yes, XGMI does have a practical use. The thing I don't really get is AMD has been using the PCIe bus for inter-card communication for years. I question how much of this is mostly just marketing fluff with minor efficiency improvements over their "old" method.

Clawedge:

XGMI.....I thought it was some new sort of sex thing.

Well, I suppose you could think of it as a GPU orgy.

AlbertX:

Does it Really matter? I am a Huge SLI/CF supporter but the games are just nos supporting SLI anymore, is declining more and more, the only recent title I can think supports sli is FC5 nothing else

crossfire/sli is still viable, its up to devs to support. the big issue imho is that the latest unreal engines can't do mGPU at all or very well and EPIC chooses not to support it in their engine citing technical reasons. Check this recent review from a redittt user benchmarking a bunch of recent games, surprising results in how many games still can do it: https://www.reddit.com/r/Amd/comments/8ylz2i/vega_64_crossfire_scaling_testing_in_24_games_5/ and recent strange brigade just added mGPU Vulkan support with near 2x scaling! https://wccftech.com/vulkan-radeon-mgpu-performance-in-strange-brigade-tested/

the use of "infinity fabric" is not news or even surprising (to me), the new xgmi connector is. AMD has roadmapped it for a while. indeed, NVLink was being researched in response as well as Intel's "Omni-path" as the previous interconnects are insufficient for modern bandwidth/latency req's.

Fox2232:

Comparing 2 unreleased chips/architectures. Should we laugh or cry?

I'm with the "laugh" crowd...;) Nothing like seeing pre-production marketing for non-shipping hardware, is there?...! Countdowns...secret this and secret that...yada-yada-yada...false leaks galore, etc. Such fun. Well... sorta--maybe...nahhhhhhh.... 😏 Not really, after you've seen it a dozen times already. Thing is, sadly, some people always believe the hype--but, fortunately, the cure for that is simply living a few years longer! nVidia is egregiously bad about it, & Brian Burke has an affair going with "50%" this or that, he obviously inherited from nV--going all the way back to the original set of "Detonator drivers," guaranteeing "up to 50% performance increases!" (With most people not seeing the "up to" qualifier, and before Burke's time at nV--while he was still at 3dfx, IIRC.) I think in those days after the fantabulous, incredilissimo !!! original Detonator drivers were released (I had either a TNT or TNT2 w/a couple of V2 6MB cards--so I could test 3dfx-GLIDE acceleration and compare it to nVidia's OpenGL 3D acceleration all day long--had a big GLIDE game library--and D3d/OpGL at the time was slim pickins!)--- we found one place where the frame-rate of the Dets sky-rocketed--HUGE- fps increase. It was in one location in a popular FPS shooter, and you could see it every time, as soon as you moved your character over in a corner and facing the wall--it was incredible!! Long as you sat still looking at the wall while nothing at all happened graphically--darned if the frame-rates didn't spike through the roof! Heh-Heh...I doubt I will ever forget things as over-blown as the "Detonator" driver scam, or the nV driver-on-rails 3d-Mark FutureMark scandal, or the Nv30-cum-leaf-blower scandal, or...well, I'd run out of room very soon so I should quit. Oh, almost forgot to mention, that in all other respects that could be seen--the Detonators were indeed duds--and in most cases there was no performance improvement whatsoever to be observed. Thus the name "Dudanators" was soon applied....Haven't seen a new "Detonator Driver Release" out of the ol' nV in many a year...such a shame as they are so much fun to to critique...;) Outlandish, hugely exaggerated performance improvement claims are among the easiest to debunk--naturally. DOesn't seem to stop nV's PR department, though--they are consistent if nothing else...:) Myself...if AMD or nV, for that matter, can ever come up with a way to put two GPUs together on a card and have the OS, the GPU drivers, *and* the 3d game APIs see it as not two, but just one single GPU...then we might actually see something worthy of even nVidia PR exaggerations and exclamations..!! We may be closer than any of us thinks...D3d12 contains the stuff to get it done--as it moves multi-GPU support into the API transparently--but the catch is a d3d12 game has to be written ground-up to do that--as D3d still supports the old non-support of multiGPUs in the API, too. We may have to wait awhile--but this does sound interesting. Still, may be awhile...hard to judge.