AMD brings Multi-GPU support to ROCm™ 6.1 Software for AMD Radeon™ desktop systems

[@BuddiiiBoiii]

I saw multiple gpu in the title and was hoping crossfire was making a comeback..

[@Dr_Mario2007]

It sort of did, this time it's general purpose supercomputing tasks including AI. It would be awesome if they also bring back parallel graphics processing feature with the latest version of Crossfire, as Crossfire no longer use bridges, rather through PCIe interfaces as they're obviously massively faster than the bridge interface.

[@commanderoof4578]

Crossfire is unlikely to make an official return... Would be pointless unless your in a 3x ultra wide 11520x2160 or something

[@DelticEngine]

In a sense, it never went away. The name 'crossfire' went obsolete along with the 'bridge'. Having multiple GPUs has always been supported so can still do a 'crossfire' equivalent if you want to, provided the application or game or whatever supports more than one GPU. The difference is that any GPU bridging is done through the PCIe bus, rather than a dedicated bridge, which has the advantage of combining several GPU cards rather than just two. The exception to this is where a game or other application is using the Vulkan API which will, at least in theory, use all the processing power you have in your system to improve performance including multiple GPUs and multiple CPUs. I have a multi-CPU workstation. The Vulkan API used every processor core in the system when I tested it and gave a corresponding boost in performance. Vulkan really is that good.

[@hasiktak1544]

Sure, pay more for less

[@MaddJakd]

Technically, it never left outside of gaming

[@KianFloppa]

Team AMD ❤

[@novantha1]

To be honest, I think AMD is catching up with ROCm in the GPU space, but I still contend that they're missing their biggest advantage; AMD has the cheapest GPU development platform with their iGPUs, making a very attractive platform with theoretically insanely high (V)RAM count, but it's not an officially supported configuration.

[@sailorbob74133]

Hopefully that'll change with Strix Halo later this year... 4070 class iGPU with unified memory...

[@jonasschulze8097]

There are so many gfx90c out there. I hope AMD starts to support every GPU in either ROCm or something else that Kokkos or SYCL support.

[@ProjectPhysX]

It took AMD only an entire year to figure out that making a workstation card 3-slot wide turns it into a useless brick, as 2 of them won't fit side by side... Btw, does the OpenCL extension function for P2P copy between AMD GPUs still segfault?

[@konserva_privet]

Слава АМД! Корпорации ДОБРА!!!

[@HilLiao]

Why does ROCm support so few GPUs on Linux? AMD should know better Most production deployments are on Linux. Developers want to develop in near Production Linux environment for smooth promotion to produciton.

[@mikoajneronowicz582]

I'm counting on AMD! I am rooting and a fan and at the same time I am very afraid. I hope AMD will not abandon support for RDNA 2 architecture Radeons RX 6000 series? I recently bought my GPU and wouldn't want to replace it in a while. Will the new ROCm releases support the previous generation of graphics cards? Can you count on some material on the use of RocDecode, HIP, Implementation of HIP in Davinci Resolve for example? AI and stable diffusion work great on Linux but there is no easy way to run Krita AI, for example. These are tools for creators not necessarily programmers. Are there plans to make these types of tools easier to use on Linux as well?

[@JonitoFischer]

I wonder if the hardware table support for ROCm in Linux is up to date. I read about windows having support for RX7600 but not linux, why?

[@Dr_Mario2007]

I am kind of surprised that Crossfire features made a comeback, this time for supercomputing acceleration workloads rather than gaming (even though it would be awesome to have both as some folks like me have either Epyc and / or Ryzen Threadripper server CPUs so we have multiple PCIe interfaces we can take advantage of in the new or secondhand builds).

[@DelticEngine]

Multi-GPU support with AMD never went away. Though the original crossfire and the electrical bridge connection was dropped, the use of multiple GPUs or 'multiGPU' never went away. The linking or bridging is done through the PCIe bus. This means that more than two GPUs can be supported. A game properly supporting the Vulkan API should use every GPU in you system as well as multiple CPUs if you have a multi-processor workstation. This is where gaming on Linux can potentially wipe the floor with Windows in performance as the DirectX goes through through Vulkan automatically. This means, potentially, that any game can be accelerated my several GPUs and CPUs even if it was never originally coded to do so.

[@moeinshakiba1102]

Please release rocm for low to mid-range GPUs on Linux like your competitor Nvidia , ai scientists may use linux , but thanks because rocm is open source.I love AMD.

[@franzpleurmann2585]

So no support for the Radeon Pro W6800 anymore? It is not listed in the hardware compatibilty chart.

[@rodneyengland4014]

I think AMD and Nvidia need to bring back crossfire and SLI, not for gaming but for computational tasks on the higher end enthusiasts gpu market. I thought of this a couple days ago and this video got me excited thinking that's what it was about.

[@patrickgronemeyer3375]

lol It's been here the whole time. It's called explicit multi adapter or mgpu, There there's a bunch of different names for it. The public just seemed stupid and hyperfixated on sli or crossfire.... I'm going to tell it to you directly it still exists but it goes by a different name... That's why a lot of the companies gave up on it. The public user was too stupid and stubborn to re educate the new verbiage.

[@Wi11i_St1g]

Dear AMD, AI is nothing without sea of data. To clean data, you need something like pandas or Spark. nVidia has RAPIDS which can accelerate spark pipeline. May you make ROCm to be able to accelerate Spark or provide something similar?

[@DeanClean42069]

I have a question AMD, with mGPU can you utilize any of the RX7000 family cards together ? Like for example, if someone were to pair a RX 7900 GRE with an RX 7600, would that work? Having the 7900 for main graphics card and the 7600 for ai processing for npcs? Is that possible with ROCm? If so that is pretty cool

[@patrickgronemeyer3375]

"Strange Brigade" is the only game I know that fully Supports mGPU combinations. Make sure you have MGP U turned on in your radeon drivers.

[@tor3203]

I wish I understood what was going on in the video 😅

[@PMCGroup-rb4jn]

Please release AMD adrenalin AI... 😅

[@TheTaurus104]

Does this mean that we now have full Rocm support for Windows like we do with Linux? Please for more customer GPUs. When AMD Software Pro Edition for RDNA3 Windows?

[@holthuizenoemoet591]

no unfortunately only via the linux subsystem for windows...

[@Accuaro]

Sorry bro just get an Nvidia GPU or use Linux

[@HazzaJenko]

​@holthuizenoemoet591 do we have full rocm support in wsl?

[@Accuaro]

@timer4times2 Hilarious that people on Windows resorting to ZLUDA because AMD won't bring ROCm to windows. They even killed the ZLUDA project, thankfully it went open source in the end but that's its end.

[@TheTaurus104]

@@Accuaro AMD has announced that they will bring full ROCm Support for Windows. But it will take time. This was since the day they bring HIP SDK for Windows. In dont give up to hope. And for the Waiting Times i work under Linux.

[@JohnEagle-l3e]

How is it possible to get the PC that you showing, it looks amazing as a Workstation intents to be

[@Accuaro]

How about you bring this to Windows?

[@bladestech]

0:18

[@Accuaro]

@@bladestech That is still Linux, that isn't Windows specific. Read carefully what that says and google it.

[@MathieuPrevot]

@@bladestech It is not on Windows, it is on linux and it works on WSL.

[@MathieuPrevot]

ROCm indeed would be great to have on Windows.

[@BKXZ]

So when are you going to start paying attention to mobile graphics?

[@CFI-2018]

TeraScale~GCN based? Or RDNA based?

[@BKXZ]

@@CFI-2018 Of course based on RDNA.I am using 7600M XT now.

[@CFI-2018]

@@BKXZ There are various Samsung Exynos GPUs, and one of them is called Xclipse GPU. Search for it.

[@preefix33_3]

need the availability of AMD laptop APU

[@77wolfblade]

That sounds great now give us good prices.

[@jasper123t]

AMD YES!

[@Shahzad12357]

What does this do exactly?

[@miko3085]

Have you heard some aplications use NVidia's CUDA as acelerators? Maybe a simple example is like in Blender for rendering, using CUDA will make the render faster, but CUDA is only usable with NVidia GPUs. ROCm is something like a translator for CUDA, so AMD GPUs can use them kind of. ROCm have been around for awhile for their professional products, I think starting now they're making consumer class AMD GPUs compatible with ROCm.

[@Anonihmus2567]

Multi GPU support for ai acceleration and other productivity stuff

[@harryniedecken5321]

My understanding is that it combines the computing power of the cpu and multiple GPUs together. That way you can do things like heavy duty video editing or difficult simulations much faster. For example start off with one W series pro graphics card and if you have a project that needs more processing power, you can just add more cards modularly. For instance I thinking about getting a lower core count threadripper system with a W7700 for some blender rendering and light duty video on Linux. Pretty sure that I can later add another W 7 series card with even more capabilities later. Not an expert so still learning about it.

[@IamSH1VA]

for AI training

[@sean8102]

It's for AI related work loads. I believe it's AMD's equivalent to Nvidia's CUDA API.

[@BlueHound]

I think I will wait until ROCm has full Windows support and CUDA compatibility built in.

[@sean8102]

"CUDA compatibility built in" Is that even possible? Pretty sure CUDA isn't open source, and no chance Nvidia would ever liscence it to anyone else. I'm pretty sure AMD would have done that as quick as they could if it was possible. That's one of their biggest hurdles fighting Nvidia and their domination in the space of AI related workloads.

[@kraio-sfu]

@@sean8102There is something called ZLUDA, which was, for a time, officially sanctioned by AMD. However, they recently pulled the plug on it

[@holthuizenoemoet591]

what we need is a Jupyter notebook using AMD to train a model on a dataset. We don't learn anything from these promo-videos. Give use actual code to work with!

[@nempk1817]

Code on a promo video😂

[@GabrielAlejandroDiazFierro]

Bad, why old versions of Rx are not supported? Why CUDA on NVIDIA old devices how Gt series .... Amd needs to improve on applying romC on slightly older graphics cards, otherwise they are forcing the user to purchase new GPUs. I was expecting a little more...

[@extremetruth]

AMD Awesome! 💯

[@brunognoato8016]

Its how 6900xt cant run pytorch when a 3050 can ....😢

[@TheBigJohny]

AMD. I bought RX 6750XT some time ago and I feel betrayed by this. No support for Pytorch etc. its absolutely lame

[@TheTaurus104]

No AI Cores. No Support for Full Rocm. But you can use HIP with the AMD Software Pro Edition for Stuff like Handbrake and Blender.

[@holthuizenoemoet591]

I understand, i also feel like AMD can do better. However pytoch does suport rocm 6.0 on linux, however im not sure how well training a model would work. Inference should work though.

[@БрухБрух-щ7и]

Bring rocm to rdna 1.0!!!

[@TheTaurus104]

Has no AI Cores and to less Vram

[@БрухБрух-щ7и]

@@TheTaurus104 you don't need ai cores to run neural networks. They help, but don't do all of the lifting

[@TheTaurus104]

@@БрухБрух-щ7и correct. But you need also Vram and RDNA1 has 6 or 8 GB and that is not enough in 2024 especially without AI Cores.

[@БрухБрух-щ7и]

@@TheTaurus104 ai cores do not magically decrease the amount of vram you need,they just handle the matrix calculations better than regular stream processos, and 8gigs are sufficient for amateurish ai workloads(plus the cards can pool vram)

[@VRMMOTH]

❤🌹😍 AMD PRO W7900 48GB Good I'm ask form Thailand really

[@kalidesu]

Over priced sadly, just get x2 7900XTX's for AI

[@Nattymane]

CrossFire is back 😮

[@patrickgronemeyer3375]

It's called MGPU... It never disappeared it just changed names.. Stop referring to it as crossfire stop googling crossfire... Stop getting on mGPU forms and squawking about crossfire.

[@z_MVP_Z]

time to bring back crossfire or similar for gaming GPUs

[@DelticEngine]

It kind of never went away. The original crossfire requiring a physical bridge between two cards was dropped, but having multiple GPUs carried on being supported. It's now called 'multiGPU' and works through the PCIe bus. To use gaming as an example, any game fully supporting the Vulkan API should be accelerated by as many GPUs and CPUs you have in your system.

[@Scott99259]

Where is Fsr 3.1?

[@Accuaro]

Q2, so most likely June 30 if not a bit later.

[@dgillies5420]

Its out now ...

[@OG_Gamers_Live]

To be clear most of us could care less about the Current AI hype.

[@Letsflipingooo98]

speak for yourself

[@AngelA-tj9ok]

And most of us as he said. You're speaking for yourself.​@@Letsflipingooo98

[@justene89]

Its no hype, AI is here to stay.

[@sean8102]

These products aren't for consumers.

[@lolonim]

Do something with the stock, god damn amd!!

[@survivor303]

just give me great video render performance in linux with kdenlive. thank you.. and please just add external power supply to your gpu:s, then we dont need to use internal psu to power these power hungry cards.

[@TheTaurus104]

HIP SDK/API

[@survivor303]

​@TheTaurus104 you mean that i can do it.. well i cannot, that's why I ask from the source 🤷🏼‍♀️

[@shera_1204]

Amd 8700g is over price please dicris price 18500₹ all tex

[@MauriBT]

Bring GPU-Multi support in gaming i want 7900XTX CROSSFIRE!!!!

[@DelticEngine]

Crossfire in its original for is long dead. Support for multiple GPUs never went away. You can still have multiple GPUs in your system. It's called 'multiGPU'. Any games using the Vulkan API should result in acceleration from all GPUs and CPUs in the system.

[@MauriBT]

@@DelticEngine i dont want this

[@patrickgronemeyer3375]

It's called MGPU. To been here the whole time. You're just Google searching the wrong things.... Stop Googling crossfire and start Googling MGPU. That's partially why it went away nobody could wrap their minds around the new name

[@patrickgronemeyer3375]

@@DelticEngine Yeah I completely agree with you Unfortunately most Vulcan games don't support MGPU

[@alfonso5177]

Yeah

[@Da-Bolt]

rocm for rx6700xt

[@MathieuPrevot]

Too bad you don't support ROCm on Windows natively.

[@Accuaro]

They honestly most likely never will, or it'll be 5+ years or a decade.

[@TheTaurus104]

There is restricted Support for Scenarios via newer Driver. And we have HIP SDK for the other not AI Stuff. But i agree rocm must come to Windows.

[@kalidesu]

@@TheTaurus104 Yep we need full pytorch and compile on Windows not just Linux, I don't want to dual boot. No wonder people just go pay extra and get an Nvidia card.

[@Rmm1722]

Nice ❤🎉

[@lejeffe1663]

oh great so you can abandon us again like you did with adrenaline mgpu?

[@JamesFox1]

SO EVERYONE STOPPED SLI / CROSSFIRE = JUST SO MAJOR CORPORATIONS ARE B THE ONLY WAY YOU CAN OBTAIN SUCH ACCESS TO THESE FEATURES ? soooo lame

[@DelticEngine]

The original AMD Crossfire with the bridge connector was dropped. The support for multiple GPUs or 'multiGPU' never went away and works through the PCIe bus with multiple cards. If your game properly supports the Vulkan API then you should get acceleration from every GPU and CPU core you have in your system.

[@JamesFox1]

@@DelticEngine , yea using what = a crypto account ???

[@TheTaurus104]

Multi GPU in for HPC and AI Stuff has nothing to do with the old Gaming Features für Gaming as Sli or Crossfire.

[@JamesFox1]

@@TheTaurus104 thats what i said above , if you wrent paying attention

[@patrickgronemeyer3375]

No its been here the whole time the name changed. its called MGPU now...

[@yorozuya61377next]

please windows native support...

[@gh0stgl1tch]

gfx1103 😢 pls😅