RDNA 5 speculation

[Marees]

If the new xbox has the 9070 xt successor in it, what will the system cost at launch 999? I feel like they need another generation like the Xbox One and PS 4 where they pull back on the specs a bit, things are way too expensive right now and I'm not so sure it's wise to sell at console at a huge loss.

Yeah, the way Microsoft is going they could make the next xbox as $1000 plus

But hopefully TSMC 3nm prices & GDDR7 prices have cooled down by then

If AMD launches the successor of 9070xt for a street price of $600 then that could keep the prices of xbox in check too

 

[Marees]

My expectation of RDNA 5

RDNA 2 was a full stack top to bottom launch

This is more like Radeon VII & Polaris

• AT0 — big bad card 512 bit bus max 192 CU (equivalent to Radeon VII)
• AT1 — missing card 256 bit bus 96 CU (scrapped because makes no business sense to compete with 80/80s/80ti/80tis cards)
• AT2 — 9070xt/7800xt replacement 192 bit bus 72 CU (the polaris of RDNA 5. Also shared with xbox)
• AT3 — 9060xt replacement 128 bit bus 48 CU ? (mainstream / entry level)

 

[NKD]

Yeah, the way Microsoft is going they could make the next xbox as $1000 plus

But hopefully TSMC 3nm prices & GDDR7 prices have cooled down by then

If AMD launches the successor of 9070xt for a street price of $600 then that could keep the prices of xbox in check too

xbox is going to use mid-range die. So i don't suspect it being 1000. It will probably land between 600-700. I also wouldn't be surprised if they sell it at 499.99 and eat some losses to grow some sales. 599.99 is probably where I suspect it will land.

 

[SeymourGore]

I wonder if there will be 'XBox' editions of AMD's Radeon GPUs. These guys could have Xbox backwards compatibility built in (like the next gen Xbox will apparently have) and promise base 'Xbox'-like performance and features.

 

[Decko]

I wonder if there will be 'XBox' editions of AMD's Radeon GPUs. These guys could have Xbox backwards compatibility built in (like the next gen Xbox will apparently have) and promise base 'Xbox'-like performance and features.

I hope not lol, also seems like they work more closely with Sony these days than MS. Also that sounds mostly like a branding exercise depending on the implementation. I'm hoping Xbox doesn't bifurcate their products again, the Series S sucks. We'll see what chips cost when the time comes, I'm expecting 699 for next gen consoles. I would think MS and Sony already have their projections in for chip costs of next gen machines, that and general inflation trends.

 

[Marees]

I hope not lol, also seems like they work more closely with Sony these days than MS. Also that sounds mostly like a branding exercise depending on the implementation. I'm hoping Xbox doesn't bifurcate their products again, the Series S sucks. We'll see what chips cost when the time comes, I'm expecting 699 for next gen consoles. I would think MS and Sony already have their projections in for chip costs of next gen machines, that and general inflation trends.

Microsoft & AMD very recently publicly announced some kind of tie-up. Something is cooking for sure...

 

[Marees]

MLID hallucinations / leaks

AMD’s RDNA 5 Based 10900 XT May Feature 36 GB VRAM & ~10K Cores [RUMOR]​

https://hardwaretimes.com/forget-th...00-xt-may-feature-36-gb-vram-10k-cores-rumor/

 

[zehoo]

As with all things AMD, it's best to wait for the reviews. But I'll probably pick up whatever is their top tier, assuming it isn't priced stupidly and performs reasonably. And since ray tracing will be getting more important next console cycle I'll probably have to consider the green team's products again. This is all assuming that there are even any interesting games to play on it, games that interest me are getting few and far between as I'm not part of the mythical modern audience.

 

[Marees]

We now have the full RDNA 5 lineup

1. AT0 (>5090) — 144 CU to 192? CU (384 bit gddr7 to 512 bit gddr7)
2. AT1 — scrapped ?
3. AT2 (~5080) — 72? CU (192 bit gddr7 = shared with xbox) — there could be cut down versions
4. AT3 (~5070?) — 48 CU (384bit lpddr6) — shared with medusa halo mega
5. AT4 (~3060 12gb +) — 24 CU (128bit lpddr6) — shared with medusa halo mini

View: https://youtube.com/watch?v=K0B08iCFgkk&t=610s

 

[Gideon]

https://wccftech.com/roundup/amd-rdna-5-gpus/

Latest info I could find.

 

[Marees]

https://wccftech.com/roundup/amd-rdna-5-gpus/

Latest info I could find.

The memory info is not as per the leaks

1. AT0 (90xt) — 512 bit gddr7 that is cut down to 384bit gddr7
2. AT2 (70xt) — 192 bit gddr7
3. AT3 (60xt) — 384 bit lpddr6 or 256 bit lpddr5x
4. AT4 (50xt) — 128 bit lpddr5x

Given current memory situation, I expect below SKUs

1. 50xt 12gb — $300
2. 60 16gb — $400
3. 60xt 16gb/24gb — $500/$550
4. 70 18gb — $600
5. 70xt 18gb/24gb — $700/$800
6. 80xt 24gb — $1200+
7. 90xt 36gb — $1500+

 

[Mackintire]

There used to be a 60 or 72 CU variation and I still think it's a bad idea for Team red not to use that.

 

[Marees]

There used to be a 60 or 72 CU variation and I still think it's a bad idea for Team red not to use that.

70xt should have close to 72 CUs

Also the next xbox (series X2?) should use that same GPU chiplet

 

[Gideon]

The memory info is not as per the leaks

1. AT0 (90xt) — 512 bit gddr7 that is cut down to 384bit gddr7
2. AT2 (70xt) — 192 bit gddr7
3. AT3 (60xt) — 384 bit lpddr6 or 256 bit lpddr5x
4. AT4 (50xt) — 128 bit lpddr5x

Given current memory situation, I expect below SKUs

1. 50xt 12gb — $300
2. 60 16gb — $400
3. 60xt 16gb/24gb — $500/$550
4. 70 18gb — $600
5. 70xt 18gb/24gb — $700/$800
6. 80xt 24gb — $1200+
7. 90xt 36gb — $1500+

I tend to be highly skeptical of hardware leaks this early. I have a feeling the memory that gets attached will be highly evolving based on the market conditions.

 

[LukeTbk]

this early.

ideally it should not be that early right now, if they want something lock in Q1 2026 to have a launch at volume for an october 2026 q4, if they skip 2026, they would go a full 5 years before beating a RTX 4080 of 2022.... and quite a long time before a new full top too bottom stack (with mobile dgpus) to replace the 7000 series.

 

[Agent_N]

Yeah prior to the 6900XT I can't remember the last time AMD was going toe to toe with Nvidia's flagship. Was hoping they would be able to do the same thing with RDNA3 but here they are back to competing with Nvidia's 2nd best instead and next gen looks to be either the same thing where they can only compete up to the 2nd best from Nvidia, or even worst where they can only compete with maybe the 3rd/4th best in Nvidia's lineup. Maybe RDNA5 can get things back to the 6900XT days but meh I'm even doubting that.

About the time you joined Harforums, AMDs R9 290X. High end cards are not where the money is at, it is at the higher volume mid and low-end cards that more people can afford to buy. I don't care if AMD cannot compete with the 5090, I care that I can get solid performance for a good price. However, AMD is making a lot more on AI hardware because companies with deep pockets are ready to throw billion at them for AI hardware.

 

[Gideon]

ideally it should not be that early right now, if they want something lock in Q1 2026 to have a launch at volume for an october 2026 q4, if they skip 2026, they would go a full 5 years before beating a RTX 4080 of 2022.... and quite a long time before a new full top too bottom stack (with mobile dgpus) to replace the 7000 series.

Nobody is rushing to serve the consumer market right now. I expect a mid to late 2027 release. AI everything first than the loser chips that don't make that cut will be tossed to us, that will give them the time to build enough inventory to throw to us. Nvidia will likely do the same. Were just going to get scraps until the AI market stalls out.

 

[Mackintire]

Ideally the skus they need to build are need to complete with:


B580/5060 level of performance at 1080, selling at $250-$300
5060 TI level of performance at 1440, selling at $400
5070 TI level of performance at 4k, selling around $600
5080 Super performing very well at 4k, selling at $1000
Build a 5080 SUPER GRE version for $800 with reduced RAM

 

[LukeTbk]

Nobody is rushing to serve the consumer market right now. I expect a mid to late 2027 release. AI everything first than the loser chips that don't make that cut will be tossed to us, that will give them the time to build enough inventory to throw to us. Nvidia will likely do the same. Were just going to get scraps until the AI market stalls out.

A lot of those desktop dgpu chips are not really enterprise bin down one (5090 is a bit on an exception in that regard), failed B200 cannot be turned into 5080/5090, they need to be good enough to be binned down into B100 or they are lost.

Nvidia will likely have hbm memory/packaging/networking bottleneck with too much silicon capacity for the GB200 and its successor it can make and will continue to make desktop gpus, like they did all along since 2022, the gpu die silicon are on quite mature node and usually not the main bottleneck

 

[Marees]

A lot of those edsktop dgpu chips are not really enterprise bin down one (5090 is a bit on an exception in that regard), failed B200 cannot be turned into 5080/5090, they need to be good enough to be B100 or they are lost.

Nvidia will likely have hbm memory/packaging/networking bottleneck with too much silicon capacity for the GB200 and its successor it can make and will continue to make desktop gpus, like they did all along since 2022, the gpu die silicon are on quite mature node and usually not the main bottleneck

A lot depends on when the vram prices will cool down

It appears that will happen only H2 2027

 

[LukeTbk]

that an issue for budget high vram card, for $1000 sub 450mm gpu like the 5080 and probable 6080, probably lot of room to make the added $80 work, specially at the real retail price at launch those cards tend to be, they could have been over 60% margin for Nvidia product. Nvidia as an history of releasing bigger card first and low vram on lower end that can make high price memory window "natural" on a new generation.

Will be harder on AMD that tended to rely on better vram capacity/price to attract buyer.

 

[Marees]

AMD GFX13 gfx1310 target appears in LLVM, hinting RDNA 5 development​

Published 2 hours ago by Hilbert Hagedoorn

An AMD “GFX13” identifier has resurfaced in open-source tooling after a long period of little to no public movement, and this time it shows up where GPU enablement often begins: the LLVM compiler stack. Late on Friday, January 23, a public LLVM Project update was highlighted by Kepler_L2, revealing that initial support has been defined for an AMD “gfx1310” target. While the change is small on the surface, new “gfx” IDs are typically the first visible signs that a future architecture is being prepared across the software pipeline. A developer note attached to the update suggests gfx1310 is currently handled as a placeholder, treated as equivalent to RDNA 4-era identifiers, specifically GFX12 and GFX1250. That kind of temporary mapping is common when a new target is introduced: compiler infrastructure needs a recognized target name and basic plumbing before it can meaningfully express architectural differences. In practical terms, it means the target exists in the tree, but it is not yet differentiated in a way that reveals new instruction behavior, scheduling changes, or feature flags.

Even so, the return of “GFX13” is notable because it echoes older references from prior development notes. Before 2026, various patch notes and ecosystem discussions mentioned early “GFX13” work in the same orbit as next-generation AMD graphics architecture naming, including “RDNA 5” and “UDNA.”

The new LLVM entry does not confirm what AMD will ultimately call the architecture, and it certainly doesn’t confirm retail product branding. However, it does reinforce the idea that a post-RDNA 4 GPU target family is being staged on the compiler side. Some observers have pointed at LLVM 23-related activity and suggested the work is already tied to RDNA 5 specifically. At this stage, that’s still interpretation rather than confirmation: the “identical to GFX12/GFX1250” note implies gfx1310 is not yet exposing meaningful differences. The more grounded takeaway is that the target has been introduced early, and follow-up commits will be the ones that matter for understanding what gfx1310 actually represents.

As always, speculation quickly drifts toward product naming and timelines, including talk of a theoretical “Radeon RX 10000” desktop GPU family and a vague mid-2027 timeframe. Those ideas may align with typical generational pacing and even potential next console cycles, but none of that is established by a single compiler-target entry.


What the LLVM change does indicate is that the software groundwork has started—or restarted—in public view.


The next checkpoint will be further LLVM development. Phoronix founder Michael Larabel has already flagged interest in the LLVM 23.1 stable release expected around August/September, since that window could bring more commits that move gfx1310 from placeholder status toward real target definition. Until then, gfx1310 is best treated as an early breadcrumb: relevant, but not yet a roadmap in plain sight.


Source: LLVM Project, Kepler_L2, Phoronix, Wccftech, LLVM GitHub

“"GFX13" Target Spotted in AMD LLVM Update - Paving the Way for Radeon "RDNA 5" GPUs​

by T0@st Today, 15:58 Discuss (0 Comments)
The intriguing "GFX13" identifier has turned up again, after many months of silence. Late last Friday (January 23), Kepler_L2 highlighted a public-facing LLVM Project update that defines initial support of Team Red's "gfx1310 target." One team member comment points to this new property being identical—as a temporary arrangement—to the RDNA 4 generation's "GFX12" and "GFX1250" IDs. Prior to 2026, a bunch of AMD patch notes mentioned early "GFX13" commitments, alongside next-gen "RDNA 5" or "UDNA" graphics architectures.

It is believed that last week's LLVM 23 compiler release only refers to "RDNA 5," leading to speculation about a forthcoming theoretical "Radeon RX 10000" discrete graphics card family. This next wave is not expected to launch any time soon; industry observers posit a vague mid-2027 launch window—maybe coinciding with unleashings of next-gen home consoles, although present-day market conditions could spoil Microsoft and Sony's best laid plans. Phoronix's founder and Editor-in-Chief, Michael Larabel, is looking forward to the August/September arrival of AMD's LLVM 23.1 stable release—by that time "GFX13" support could have advanced to another stage of development, thus providing further insights.“