AMD Announces Advancing AI 2025

[btarunr]

Today, AMD (NASDAQ: AMD) announced "Advancing AI 2025," an in-person and livestreamed event on June 12, 2025. The industry event will showcase the company's bold vision for AI, announce the next generation of AMD Instinct GPUs, AMD ROCm open software ecosystem progress, and reveal details on AI solutions for hyperscalers, enterprises, developers, startups and more. AMD executives and AI ecosystem partners, customers and developers will join Chair and CEO Dr. Lisa Su to discuss how AMD products and software are re-shaping the AI and high-performance computing landscape. The live stream will start at 9:30 a.m. PT on Thursday, June 12.



View at TechPowerUp Main Site

 

[clopezi]

No CUDA no party

 

[ymdhis]

I don't give a fuck about AI, give me better APUs instead.

 

[Caring1]

Advancing towards Judgement Day.
Sarah Connor, where are you?

 

[Firedrops]

AMD keeps announcing AI this AI that, but don't even support fundamental matrix operations [1, 2].

 

[igormp]

AMD keeps announcing AI this AI that, but don't even support fundamental matrix operations [1, 2].

Your sources are about DirectML, which is a bad supported backend overall, it has nothing to do with any specific hardware, and would fail with any kind of accelerator.
In fact, in your first example there was a comment with issues regarding Apple chips.

 

[samlebon2306]

No CUDA no party

Are you CUDAing?

 

[Firedrops]

Your sources are about DirectML, which is a bad supported backend overall, it has nothing to do with any specific hardware, and would fail with any kind of accelerator.
In fact, in your first example there was a comment with issues regarding Apple chips.

If you actually used AMD, you'd know that their rocm doesn't support Windows, and hence many use DirectML implementations.

And yes, some issues were raised and resolved for the Apple variant of DirectML over the past year, but none ever made it to the AMD-compatible side of things. And why is DirectML considered a "bad supported backend" today? Because the main vendor that it's used with doesn't bother supporting it? Chicken or Egg?

Why are you even bothering to reply and white knight in domains you're not familiar with? You think AMD will reward you with some free stuff?

 

[igormp]

you'd know that their rocm doesn't support Windows

It does:

HIP SDK installation for Windows — HIP SDK installation (Windows)

HIP SDK installation for Windows

rocm.docs.amd.com

It is somewhat recent, and likely even worse than the linux version, but it is supported now nonetheless.

But yeah, I don't use AMD for compute nor recommend anyone to do so. It does have its fair share of issues and unsupported stuff, but I just wanted to point out that those specific example of yours were not really factual to the point you were trying to make.

and hence many use DirectML implementations.

No, DirectML has worse support overall and most people avoid it, no matter if you're using Nvidia, AMD or Intel with it.

And yes, some issues were raised and resolved for the Apple variant of DirectML over the past year, but none ever made it to the AMD-compatible side of things.

It seems like you're confusing some things. DirectML is a backend made by microsoft and that's hardware-vendor agnostic.
MPS is Apple's backend for their AS devices.
The AMD-compatible side of things would be ROCm, not DirectML.

And why is DirectML considered a "bad supported backend" today? Because the main vendor that it's used with doesn't bother supporting it? Chicken or Egg?

Because it's Windows-only, and it's something made by microsoft, with subpar performance when compared to any other backend such as MPS, CUDA or even ROCm.
Given how most stuff is meant to be run on linux, it doesn't make much sense to spend efforts supporting a Windows-only backend.

Why are you even bothering to reply and white knight in domains you're not familiar with?

I'm not white-knighting anything, I'm correcting misinformation. I'm well familiar with the domain since I do work daily with it and contribute a bit to many of that stuff as well.

You think AMD will reward you with some free stuff?

No, I'm not defending AMD, nor would I recommend anyone using it for any serious ML stuff, there's a myriad of arguments for that, but the ones you raised were not it.

 

[Firedrops]

It does:

HIP SDK installation for Windows — HIP SDK installation (Windows)

HIP SDK installation for Windows

rocm.docs.amd.com

Please just stop talking. You have gone so completely off-track it's not even funny. You are just cobbling together unrelated facts, pretending that they are part of a cohesive argument that addresses anything I was discussing. You have less than zero clue on how vendor-"backend" collaborations work. At this point I just hope that if anyone else stumbles upon this thread in the future and sees the nonsense you've posted, they also check my reply beneath yours before wasting their time on a wild goose chase.

 

[Visible Noise]

It does:

HIP SDK installation for Windows — HIP SDK installation (Windows)

HIP SDK installation for Windows

rocm.docs.amd.com

It is somewhat recent, and likely even worse than the linux version, but it is supported now nonetheless.

We‘ve been over this before. HIP isn’t ROCm. HIP is a SDK, not a platform.

Also, HIP is neutered on Windows.

Porting Tools has been “coming soon” for over a year.

I don't give a fuck about AI, give me better APUs instead.

Not surprisingly, AMD doesn’t give a fuck about you.

 

[igormp]

Please just stop talking. You have gone so completely off-track it's not even funny. You are just cobbling together unrelated facts, pretending that they are part of a cohesive argument that addresses anything I was discussing. You have less than zero clue on how vendor-"backend" collaborations work. At this point I just hope that if anyone else stumbles upon this thread in the future and sees the nonsense you've posted, they also check my reply beneath yours before wasting their time on a wild goose chase.

Welp, your lack of any argument whatsoever and plain name-calling really helps your point, buddy.
Feel free to continue spreading misinformation and showing your ignorance around.

We‘ve been over this before. HIP isn’t ROCm. HIP is a SDK, not a platform.

Also, HIP is neutered on Windows.

ROCm is the stack which is not fully available on windows, yeah. But all you want is the runtime to be able to run stuff, which is available and works on Windows, way better than DirectML in terms of performance.