News AMD adds mysterious RDNA 4 Navi 44 to ROCm hardware support list — with a new MI300X1 as well

[Admin]

AMD added a new 'Navi 44' GPU to its ROCm hardware list, the second RDNA 4 GPU so far. While similarities to the old Navi 24 name suggest this might be an entry-level GPU, it's more likely that there's a new codename scheme being employed for the upcoming GPUs.

AMD adds mysterious RDNA 4 Navi 44 to ROCm hardware support list — with a new MI300X1 as well : Read more

 

[Metal Messiah.]

the ROCm patch notes also added support for a new MI300 variant called MI300X1 ("gfx942"), which could be the rumored MI375 with a change in codename as well.

Don't think so.

Actually, the AMD Instinct MI300X/A already has the "gfx942" ID. So I doubt the the rumored MI375 chip will also fall under the exact same ID.

Last year the following commits were added for GFX941 and GFX942, to the LLVM compiler. The architecture codename for MI300 is GFX942, a member of the Vega (GFX9) family.

https://github.com/llvm/llvm-project/commit/1fc70210a6a585bad941f64bd3fca7909eeafdda

https://github.com/llvm/llvm-project/commit/9d0572797233857397f3fdc35fffcfb490354f56

 

[JarredWaltonGPU]

Don't think so.

Actually, the AMD Instinct MI300X/A already has the "gfx942" ID. So I doubt the the rumored MI375 chip will also fall under the exact same ID.

Last year the following commits were added for GFX941 and GFX942, to the LLVM compiler. The architecture codename for MI300 is GFX942, a member of the Vega (GFX9) family.

https://github.com/llvm/llvm-project/commit/1fc70210a6a585bad941f64bd3fca7909eeafdda

https://github.com/llvm/llvm-project/commit/9d0572797233857397f3fdc35fffcfb490354f56

MI300X1 is a new addition, though, right? I'm not off my rocker there?

 

[Metal Messiah.]

MI300X1 is a new addition, though, right?

No. It cannot be since the Instinct MI300X/A both already have the "gfx942/941" as the LLVM target name.

 

[JarredWaltonGPU]

No. It cannot be since the Instinct MI300X/A both already have the "gfx942/941" as the LLVM target name.

My question was whether the line of code with “MI300X1” is new, at least for the ROCm stuff, though “gfx942” obviously existed elsewhere prior to now. I have not seen “MI300X1” before at least.

 

[NeoMorpheus]

Given how the media trashed the 7900XTX regardless of its merits and by consequence, the sheep ignored them, i dont blame them for not going that route.

Lets see what the new ones will bring and of course, media will drill down on the still useless RT (lets be honest, there are literally a couple of games that justify the insane performance hit, the rest are just a waste of resources) so the reviewers will trash them also, regardless of what else they bring to the table.

 

[Metal Messiah.]

My question was whether the line of code with “MI300X1” is new, at least for the ROCm stuff, though “gfx942” obviously existed elsewhere prior to now. I have not seen “MI300X1” before at least.

Can't say with full surety whether it is a new line, but I think I have seen similar entries for few other products, in which there are two different versions/revisions being added with the 1 or an 'x' suffix in some cases. But I will show a proof with the patch link once I find one.

Assuming it is a new entry, not sure why the gfx192 ID was used though.

But in any case, we do know that AMD's MI375 is in the company's pipeline, so it might just be a matter of time more concrete info emerges.

https://www.barrons.com/articles/amd-stock-nvidia-ai-upgrade-5760ee3c

https://twitter.com/x/status/1785713952338346445

View: https://x.com/najfah/status/1785713952338346445?t=q1xfo1bgu6_wzFreygVjFw

 

[Metal Messiah.]

Lets see what the new ones will bring and of course, media will drill down on the still useless RT (lets be honest, there are literally a couple of games that justify the insane performance hit, the rest are just a waste of resources) so the reviewers will trash them also, regardless of what else they bring to the table.

Couldn't agree more ! I guess we should hold Nvidia accountable for over-hyping this whole RT stuff and making it vendor locked on its latest RTX hardware, instead of fully going open source. In pure rasterized performance AMD's cards still offer more bang for the buck, IMO.

But hey, everyone wanted to jump on the RT bandwagon, and get some extra "fake" frames via DLSS too as well (required for some RT games, if not all)! So it's also a sort of "psychological" impact this has had on gamers lately.

But, Ray traced scenes DO look good in some games, but that doesn't justify the performance hit it comes with (depends on the RT effect used, and game's engine though).

Also, I'm not sure how many gamers do actually pay THAT much close attention to RT/Path traced reflections, lighting, shadows and stuff like that while gaming.

I mean we need to first focus on the environment rather than on the gameplay itself to admire any graphical/visual difference RT provides. Very fast-paced games/shooters could make this even harder to notice.

The media and sheep will still trash these new entries, assuming AMD is not even trying to capture the high-end GPU market and these are just upper mainstream cards.

 

[NeoMorpheus]

I guess we should hold Nvidia accountable for over-hyping this whole RT stuff

I hate that company with a passion, but in this particular case, i dont blame them.

I blame the current crop of so called reviewers (in my book, their real title is influencers), whom blindly followed the marketing team and pushed the stupid RT narrative.

I mean, those free 4090s aren’t really free…

 

[Metal Messiah.]

Well, on some update, it appears that AMD has already started shipping early reference boards and early NAVI 48 GPU samples to its engineering labs.

This can be confirmed in the latest shipping log from nbd.ltd where several new GPU PCBs can be spotted.

A new Navi 48 XTX GPU can be seen here, mentioned as "REVB-PRE", which could be a reference evaluation platform board, and the card appears to be using the A0 silicon.

Samsung has also been also mentioned in this manifest. The PCB being used by the Navi 48 XTX GPU has the ID:102-G28201.

The production date says, "Q2 2024", so the board appears to be recent/new.

A total of Four PCBs most likely for the AMD RDNA 4 GPU family are being considered:

• 102-G28211
• 102-G28501
• 102-G28201 -----> Navi 48 XTX GPU.
• 102-C48701

 

[Metal Messiah.]

BTW, regarding Samsung being mentioned along with the production date, hints AMD might be using the process node technology from Samsung for these GPUs, most likely GDDR6/X memory as per rumors.

 

[TechyIT223]

Nice finding. So this one is flagship gaming XTX variant it seems. Navi 48 ?

Previous entries showed XT cards as well , both 44 and 48 I suppose.

 

[Metal Messiah.]

Yes, it's the Navi 48 high-end or flagship card at least for this generation of RDNA 4 lineup It's surprising to find two variants in the making, XTX and XT.

Maybe AMD will release two different configurations of the Navi 48 die. But only time will tell whether this is indeed the case.

AMD ROCm Github patch entries have already confirmed the existence of Navi 48 and 44 GPUs as well, with their respective graphics IDs, Navi 44 "GFX1200" & Navi 48 "GFX1201" SKUs.

https://github.com/ROCm/clr/blob/94...ltst/module/runtime/OCLDeviceQueries.cpp#L122

At this point I'm pretty sure we won't be getting any more SKUs in the RDNA 4 lineup, unlike NVIDIA's Blackwell series, since these 44 and 48 GPUs would be targeting the upper mainstream gaming segment.

And AMD has no plans to capture the enthusiast segment either.