2027/2028 Nvidia/AMD/Intel GPU speculation

[Marees]

RDNA 5 Is Coming Earlier Than Anyone Expected. What That Means for Your Next Build​

Plan for Q3 (2027), be pleasantly surprised if it’s Q2, and don’t lose sleep if it slips to Q4.

The DRAM and NAND Problem Nobody Is Talking About Loudly Enough​

RDNA 5 arriving in mid-2027 is not automatically good news for build costs. Tom’s Hardware reported in July 2026 that AI-driven demand is keeping DRAM and NAND prices elevated through at least Q3 2026, with no clear signal of sustained relief. GDDR7, the memory standard RDNA 5 is expected to use, is being consumed heavily by AI inference hardware before it ever reaches consumer graphics cards. That supply pressure doesn’t vanish just because AMD hits a launch date.


What this means practically: RDNA 5 cards may launch with strong performance and painful MSRPs.


GDDR7 memory, expected on RDNA 5 flagship and upper-mid cards, is under supply pressure from AI hardware demand through at least late 2026. If that pressure doesn’t ease, expect RDNA 5 MSRPs to open higher than historical equivalents. Mid-tier cards will likely hit reasonable street pricing two to three months after launch as channel inventory catches up.

 

[El_Capitan]

Any new GPU coming out will have had to pay the absurd AI-caused DRAM tax, and will be more expensive. My hope is that they're holding their GPU's back from release until the AI RAMpocalypse is over, but they're the ones directly benefitting from it, so I doubt it.

 

[Marees]

Any new GPU coming out will have had to pay the absurd AI-caused DRAM tax, and will be more expensive. My hope is that they're holding their GPU's back from release until the AI RAMpocalypse is over, but they're the ones directly benefitting from it, so I doubt it.

they are holding back the "budget" cards

AT2 will launch in the $500 to $1000 range next july / Q3 2027

AT0 above that will probably not release for gamers

unfortunately the "budget" AT3 & AT4 cannot be released because AMD made the "foolish" decision of making them lpddr rather than gddr

that means minimum 32gb vram — for a BUDGET card !!!

the end result is for $500 & below, RDNA 4 will continue service until RDNA 6 (2029 / 2030)

the AT3 & AT4 will release as part of framework AI PCs with 128gb/256gb ram !!!!

 

[LukeTbk]

unfortunately the "budget" AT3 & AT4 cannot be released because AMD made the "foolish" decision of making them lpddr rather than gddr

that means minimum 32gb vram — for a BUDGET card !!!

Or if quite limited bandwith but large vram is quite useful, just not make them budget cards...

B70 are not that powerful for 32GB (and juste 600 gb/s...), sometime even the mini RTX 2000 beat them, they can sell them for a lot of money. The gddr vs lpddr price split will determine how bad-good of a pick it was I feel like, maybe they will need to keep them for the small form factor AI Pc going crazy high... if Moe and compression keep them interesting even at very low bandwith.

 

[Marees]

RDNA 5 to beat Blackwell in CPU bottlenecked situations — MLID

https://www.technetbooks.com/2026/08/amd-rdna-5-graphics-cards-target-gaming.html

 

[WilyKit]

RDNA 5 to beat Blackwell in CPU bottlenecked situations — MLID

https://www.technetbooks.com/2026/08/amd-rdna-5-graphics-cards-target-gaming.html

So almost never gonna beat it

 

[El_Capitan]

I have a rare AMD Ryzen AI Max 395+ with 16GB LPDDR5 8000MHz mini-pc incoming. Not good for LLM's due to the lack of memory, but it'll do well for gaming using the iGPU. I'll have to figure out how much memory to allocate to the 8060S. It'll likely perform somewhere between the RTX 4060 and RTX 4070. It has a dedicated OCuLink port (as well as USB4), so I could always just use an eGPU. It's definitely a better deal than the Steam Machine (and performs better). I'm still wondering whether to install Windows 10 Pro on it (it comes with Win 11 Pro), or Steam OS.

Also, I've been looking into rumors that a 9000G series could possibly come out this year. I doubt it'll have a 8060S, but it's possible. The 9700G with a 8060S, and the 9600G with a 8050S would be nice.

 

[ZeroBarrier]

So almost never gonna beat it

Got to have those positive marketing claims whether they are only theoretical or not.

 

[LukeTbk]

RDNA 5 to beat Blackwell in CPU bottlenecked situations — MLID

https://www.technetbooks.com/2026/08/amd-rdna-5-graphics-cards-target-gaming.html

If blackwell which is almost Lovelace is the bar for RDNA 5.... the 4080 super, a 2022 gpu people called a 4070 will still be near the top of all non-nvidia GPU until spring 2029 ?

or course rdna 5 on TSCM 3 should beat very old by then tsmc 5 family node Blackwell and not just during CPU limited scenario, always and of course. That must be not what MLID said (or misleading), rdna5 will beat blackwell even more in cpu bottleneck situtations or something specific like that.

 

[Marees]

If blackwell which is almost Lovelace is the bar for RDNA 5.... the 4080 super, a 2022 gpu people called a 4070 will still be near the top of all non-nvidia GPU until spring 2029 ?

or course rdna 5 on TSCM 3 should beat very old by then tsmc 5 family node Blackwell and not just during CPU limited scenario, always and of course. That must be not what MLID said (or misleading), rdna5 will beat blackwell even more in cpu bottleneck situtations or something specific like that.

looks like I got my architectures mixed up then

he said AT0 will beat 5090 in CPU bottlenecked situations

 

[LukeTbk]

he said AT0 will beat 5090 in CPU bottlenecked situations

Listening to it, seem to mean beat it all the time and that the 5090 TI would not be enough, because AT0 will be big and on TSMC 3N, the gap with blackwell and future RDNA 5 if it exist cannot make up for such a node upgrade.

There is no way a TSMC 3N AMD gpu need some * for when it would be faster than blackwell, or something went really wrong. It's competition must be TSMC 3 rubin...

He seem way too excited in the video about CDNA/AI workload deciding what RDNA 5 for game will look like cache wise, yes AI predictable stream prefer ultra wide HBM high bandwith to local cache, so it make sense to go that way, but gaming love cache a lot, the GLobal L2 being small per core is not some great news for gaming performance here, at least not necessarily so.

That not how it would report it if it was coming from Nvidia.... and he does not seem to present clearly how normal (what they had before and others have, that gL2, talk about it as if it was new)

 

[kram182]

Listening to it, seem to mean beat it all the time and that the 5090 TI would not be enough, because AT0 will be big and on TSMC 3N, the gap with blackwell and future RDNA 5 if it exist cannot make up for such a node upgrade.

There is no way a TSMC 3N AMD gpu need some * for when it would be faster than blackwell, or something went really wrong. It's competition must be TSMC 3 rubin...

He seem way too excited in the video about CDNA/AI workload deciding what RDNA 5 for game will look like cache wise, yes AI predictable stream prefer ultra wide HBM high bandwith to local cache, so it make sense to go that way, but gaming love cache a lot, the GLobal L2 being small per core is not some great news for gaming performance here, at least not necessarily so.

That not how it would report it if it was coming from Nvidia.... and he does not seem to present clearly how normal (what they had before and others have, that gL2, talk about it as if it was new)

He should stick to leaks - he's an AMD fanboy and an Nvidia hater so him and all his predictions go together and have panned out like icebergs and the Titanic