AMD to Stick to RDNA 3+ To Power Processor iGPUs Till 2027 At Least

[btarunr]

The RDNA 3+ graphics architecture will power integrated graphics solutions of AMD processors for the foreseeable future, a reliable source with AMD leaks says. The company is planning to debut RDNA 3+, a feature update to RDNA 3, with the upcoming Ryzen "Strix Point" mobile processor. A scaled-up version will power the "Strix Halo" processor meant for notebooks with powerful integrated graphics. Given that AMD is able to scale between a certain number of compute units for its "Strix Point" and "Strix Halo" processors, it could stick with the graphics architecture for iGPUs with its upcoming processor microarchitectures even 3 years into the future.

This isn't new for AMD, the company's Vega graphics architecture debuted in 2017, but powered the iGPUs of its "Cezanne" mobile processor that came out as recently as 2021. The RDNA 3-based iGPU powering the current "Hawk Point" processor trades blows with the Arc "Alchemist" Xe-LPG iGPU powering Intel's "Meteor Lake" processor. Intel is expected to make a generational jump in iGPU performance with its upcoming "Lunar Lake" processor that debuts the Xe2 "Battlemage" graphics architecture for its iGPU, to which AMD is responding with a new iGPU based on the updated RDNA 3+.



View at TechPowerUp Main Site | Source

 

[Daven]

Seeing as there are still products with Vega/GCN, I would be surprised if we don’t see older iterations of RDNA for some time to come.

 

[Denver]

Please don't tell me the source is a forum post...

But anyway, memory has to evolve before it's worth major uArch transitions on the iGPU side, it makes sense. Could it also have something to do with avoiding competition with future portable consoles from Sony and Microsoft? Maybe.

 

[MaMoo]

Please don't tell me the source is a forum post...

View attachment 343000


But anyway, memory has to evolve before it's worth major uArch transitions on the iGPU side, it makes sense. Could it also have something to do with avoiding competition with future portable consoles from Sony and Microsoft? Maybe.

I heard it has to do with Microsoft wanting some AI NPU in every CPU, which took away room for more on-die cache and a newer iGPU. I might be wrong though, although I can't see how in principle, more room could be dedicated to NPUs without some room being sacrificed.

 

[Calenhad]

Please don't tell me the source is a forum post...

But anyway, memory has to evolve before it's worth major uArch transitions on the iGPU side, it makes sense. Could it also have something to do with avoiding competition with future portable consoles from Sony and Microsoft? Maybe.

Nuclear display. Better use proper protective gear while gaming.

 

[Nhonho]

AMD needs to improve the image quality of its AV1 encoder to a level as good as Intel and Nvidia encoders.

 

[AnarchoPrimitiv]

It's a shame this AI lunacy had basically made HBM unobtanium, otherwise with the way 3D packaging has entered the consumer realm, I would imagine if they were able to integrate some HBM in with the APU (4-8GB, that could also be dynamically utilized as an L3/L4 cache for the CPU cores), we could probably get some desktop APUs that could legitimately rival the better "entry" level dGPUs.

I've been anticipating the day we get APUs that can provide 60fps @ 1080p with High/Ultra settings a little under a decade and I feel like we're finally within striking distance.....that HBM would definitely make that goal achievable now.

 

[Chomiq]

Please don't tell me the source is a forum post...

View attachment 343000


But anyway, memory has to evolve before it's worth major uArch transitions on the iGPU side, it makes sense. Could it also have something to do with avoiding competition with future portable consoles from Sony and Microsoft? Maybe.

Weeb avatar, seems legit.

 

[LabRat 891]

Nuclear display. Better use proper protective gear while gaming.

This is a personal blast from the past, for me.
Never thought I'd re-hear/read the Gamer-ADDled rantings of a Jr. High classmate....

Lame story short:
Classmate and I in 'Reading class' both chose Doom 3 as the topic for our over-the-top 'forward-looking' writing-presentation assignment.
He, went quite over-the-top; mentioning things like a Nuclear Display that grabs you out of your seat, and fun stuff like that.

It's a shame this AI lunacy had basically made HBM unobtanium, otherwise with the way 3D packaging has entered the consumer realm, I would imagine if they were able to integrate some HBM in with the APU (4-8GB, that could also be dynamically utilized as an L3/L4 cache for the CPU cores), we could probably get some desktop APUs that could legitimately rival the better "entry" level dGPUs.

I've been anticipating the day we get APUs that can provide 60fps @ 1080p with High/Ultra settings a little under a decade and I feel like we're finally within striking distance.....that HBM would definitely make that goal achievable now.

IIRC, this is how MI3## Accelerators work, and some EPYC. 110% agreed.

 

[Avro Arrow]

This makes sense as the vast majority of PCs that use IGPs aren't used for gaming and even the old Vega IGPs are just fine for 2D use. There's little point in making an investment into an improvement on a part that few will take advantage of. AFAIC, an IGP is just a video adapter. Sure, you technically can game on it (and get pretty decent performance in some titles at 720p or 1080p) but considering that they get beat by the cheapest and crappiest video cards on the market (RX 6500 XT anyone?), I don't see the point of using an IGP for gaming.

 

[Minus Infinity]

This makes sense as the vast majority of PCs that use IGPs aren't used for gaming and even the old Vega IGPs are just fine for 2D use. There's little point in making an investment into an improvement on a part that few will take advantage of. AFAIC, an IGP is just a video adapter. Sure, you technically can game on it (and get pretty decent performance in some titles at 720p or 1080p) but considering that they get beat by the cheapest and crappiest video cards on the market (RX 6500 XT anyone?), I don't see the point of using an IGP for gaming.

And what about Halo which is going to kill the need for mid-tier or lower dGPU for laptop? By the time AMD move beyond RDNA3, even Intel will be on Celestial and already even Mediocre Lake's GPU is beating or matching RDNA3 in Hawk Point and this year we get Battlemage in Lunar Lake.

 

[stimpy88]

If this is true, then AMD is getting either arrogant or too comfortable. They should stay hungry and aggressive, as you never know what is cooking in other companies' labs that could suddenly leave you in the dust and bite your ass.

Or maybe Raja and his incompetence has produced a dead-end architecture that AMD is struggling to scale, and they need to come up with something completely new and need more time.

 

[watzupken]

I feel RDNA 3 is still very performant, likely even in another couple of years. If AMD wants to boost performance, than can probably squeeze in 16 CUs instead of 12 now. It may be struggling with some games, but I think we need to take a step back to understand that this is an iGPU. If the expectation is to have run games at higher graphic settings, then one is better off using a laptop with dedicated graphics.

 

[Minus Infinity]

I feel RDNA 3 is still very performant, likely even in another couple of years. If AMD wants to boost performance, than can probably squeeze in 16 CUs instead of 12 now. It may be struggling with some games, but I think we need to take a step back to understand that this is an iGPU. If the expectation is to have run games at higher graphic settings, then one is better off using a laptop with dedicated graphics.

Strix Point is getting 16CU's but I doubt that will be anywhere near enough to beat Battlemage in Lunar Lake. Halo will have 32/40CU but will struggle with bandwidth in laptops despite being 256bit.

 

[ARF]

But. Why ?
In 2027 they will have RDNA 4 and RDNA 5 ready.
RDNA 4 is this summer/autumn, RDNA 5 is in 2 years.

 

[Minus Infinity]

But. Why ?
In 2027 they will have RDNA 4 and RDNA 5 ready.
RDNA 4 is this summer/autumn, RDNA 5 is in 2 years.

Indeed why? I don't believe they won't leverage RDNA4 before 2026 let alone 2027. I would bet Strix Point's successor would move on from RDNA3.

 

[Nhonho]

I hope Intel releases APUs with iGPUs that vastly outperform AMD's iGPUs, to AMD realizes that times are different now.

And AMD urgently needs to improve the image quality of its AV1 encoder because it has already been surpassed even by Intel in this regard.

 

[Lew Zealand]

I hope Intel releases APUs with iGPUs that vastly outperform AMD's iGPUs, to AMD realizes that times are different now.

No chance as iGPU performance is already heavily constrained by DDR5 memory bandwidth. There's a reason iGPU core count is where it is right now and Intel already got burned with their first-gen DDR4 Skylake iGPU where they added 50% more cores but little more bandwidth and it was 10-20% faster. That's not a good enough return for the die size cost (which is going to "AI" NPUs now anyway) and I doubt there will be any more notable gains until that is somehow fixed. There are noises about quad channel RAM but I'll believe that when I see it.

 

[stimpy88]

It's about time we went 256bit dual-channel memory on desktops.

 

[Nhonho]

No chance as iGPU performance is already heavily constrained by DDR5 memory bandwidth. There's a reason iGPU core count is where it is right now and Intel already got burned with their first-gen DDR4 Skylake iGPU where they added 50% more cores but little more bandwidth and it was 10-20% faster. That's not a good enough return for the die size cost (which is going to "AI" NPUs now anyway) and I doubt there will be any more notable gains until that is somehow fixed. There are noises about quad channel RAM but I'll believe that when I see it.

One way to overcome the low data throughput in an APU is to place more cache memory in the APUs, although cache memory greatly increases the size of the APU/CPU die (in mm²).

It's about time we went 256bit dual-channel memory on desktops.

A 256-bit memory access on a desktop is only necessary for APUs, which have an iGPU, or for CPUs that have more than 8 cores.

 

[stimpy88]

A 256-bit memory access on a desktop is only necessary for APUs, which have an iGPU, or for CPUs that have more than 8 cores.

You're wrong. But the second part of your answer should tell you just how so when 8 cores is the minimum in 2024.

 

[LabRat 891]

It's about time we went 256bit dual-channel memory on desktops.

Agreed; not just for APU/iGPU.

One way to overcome the low data throughput in an APU is to place more cache memory in the APUs, although cache memory greatly increases the size of the APU/CPU die (in mm²).

AMD puts HBM stacks in Instinct MI300A 'APUs'. I'm genuinely irritated we're not seeing it trickle down.
An X3D-equipped (consumer-facing) APU would be nice, too.

A 256-bit memory access on a desktop is only necessary for APUs, which have an iGPU, or for CPUs that have more than 8 cores.

Disagree. Part of LGA1366's and LGA2011s' 'long legs' were its huge memory bandwidth.
Currently, we're seeing more bandwidth at the cost of latency. DDR4 and 5 are much higher latency vs. DDR3 Triple/Quad-channel.

 

[AnotherReader]

Strix Point is getting 16CU's but I doubt that will be anywhere near enough to beat Battlemage in Lunar Lake. Halo will have 32/40CU but will struggle with bandwidth in laptops despite being 256bit.

AMD has the solution for this since RDNA2: dedicated last level cache for the GPU. Let's see if they actually equip Strix Halo with one.

 

[Avro Arrow]

And what about Halo which is going to kill the need for mid-tier or lower dGPU for laptop?

What about it? Do you honestly believe that a company the size of AMD is going to needlessly invest millions of dollars to rush out an unfinished product because of one game that may or may not be played on IGPs?

Over 90% of desktops that use IGPs sold today are not used for gaming and the single-digit percentage that do want to game on an IGP will still get one hell of an uplift compared to what they had before regardless of what that is. It's not like people will go nuts for an AMD-based computer just because they put RDNA4 in it. Hell, the overwhelming majority of desktops today would be just fine with RDNA2 IGPs, let alone RDNA3.

Most PCs that use IGPs are office PCs sold by Dell or HP. I'm pretty sure that the companies that buy them consider how well they run Halo to be near or at the bottom when it comes to their list of priorities.

Remember that as bad as the (RDNA2) RX 6500 XT is, it still kicks the butt of every "high-end" desktop IGP out there, let alone the even weaker mobile solutions.

By the time AMD move beyond RDNA3, even Intel will be on Celestial and already even Mediocre Lake's GPU is beating or matching RDNA3 in Hawk Point and this year we get Battlemage in Lunar Lake.

I don't think that Arc is a credible threat to Radeon or GeForce (yet). Let's face it, Intel has included an IGP in every CPU they've produced since Sandy Bridge (except the models ending with "F"). They're not new at this and yet they STILL haven't managed to make anything that was able to challenge Radeon-Fusion APUs over that time period despite having more money than even nVidia at the time.

I think that it's a lot better a decision to stick with RDNA3 while Intel still has nothing that can compete with it (and they don't). This allows the people over at ATi more time to ensure that the RDNA4 models come out nice and polished instead of needlessly rushing something out that isn't ready.

I guarantee you that if Intel rushes their new IGPs out to try and beat AMD to the punch, it will be an absolute disaster for them. Just look at the state of Arc. Sure, their drivers have improved dramatically but improving dramatically from completely broken sounds a lot better than it actually is.