Never really understood why AMD keeps iGPUs behind. Like when RDNA was a thing, they still used GCN for a while in the iGPU space.

Ryzen 5000 series 3 years after the last discrete GPU, and the Ryzen 7000 series 5 years after the last discrete GPU got a couple models with GCN.

They probably upgrade the "igpus" next time with UDNA. Smart move tbh.

Will the differences between RDNA 3.5 and 4 even fit inside an SoC even on a smaller node? Also I’m guessing a lot of the differences are AI and RT related. SoCs have NPUs and I doubt anyone could actually play games with RT enabled on an SoC.

Finally isn’t RDNA 3.5 + an NPU + media decoding/encoding units close to RDNA4 anyway. I think there needs to be more context that an architecture number difference.

mrnagantNever really understood why AMD keeps iGPUs behind. Like when RDNA was a thing, they still used GCN for a while in the iGPU space.

Ryzen 5000 series 3 years after the last discrete GPU, and the Ryzen 7000 series 5 years after the last discrete GPU got a couple models with GCN.

Product segmentation.

mrnagantNever really understood why AMD keeps iGPUs behind. Like when RDNA was a thing, they still used GCN for a while in the iGPU space.

Ryzen 5000 series 3 years after the last discrete GPU, and the Ryzen 7000 series 5 years after the last discrete GPU got a couple models with GCN.

In this case, I can understand AMD keeping RDNA 3.5 for the new IGPs.
Specifically here, RDNA 3.5 has been highly optimized for power and bandwidth restrained mobile scenarios.
While it might make us feel better to have RDNA 4 (and FSR 4 would be very welcome in mobile!), unless they have the resources free to make an optimized 'RDNA 4.5', it would probably be a step backwards.

mrnagantNever really understood why AMD keeps iGPUs behind.

COST?

Historically, for every Phoenix or Hawk point sold to someone who actually cares about that there are hundreds sold to business or someone else who doesn't care about that. AMD doesn't have unlimited resources, and APU's is clearly not their cash cow.

Even if they did go for RDNA4, what's the point? Putting all that extra work into something that won't do rays in a meaningful anyway because it's too slow to begin with.

SL2Even if they did go for RDNA4, what's the point? Putting all that extra work into something that won't do rays in a meaningful anyway because it's too slow to begin with.

Nonsense. You know that RX 9070 is about 27-30% faster than 7800XT in rasterizing, too? It has ~7% less shaders but equally clocks up to 7% higher and has only slightly more bandwith, so you can assume that that's the difference between RDNA3 and RDNA4.
Now who wouldn't want 890M iGPU or even 8060S to be 30% faster? And who wouldn't wan't FSR4 on mobile? The whole point of Strix Halo (and Medusa Halo) is IGP-performance. It's a shame already that Strix Halo is not RDNA4.

Of course, the reason is cost. With only slightly more shaders, the same IF-cache and memory controller, but upgraded architecture, double the L2-cache and PCIe Gen5, Navi48 has nearly double the amount of transistors of Navi32 and nearly as much as Navi31 and can only stay about as big as Navi32 due to more expensive N4-process and highly increased packing densitiy. So upgrading IGPs to RDNA will make them quite a lot more costly to produce, which isn't an option yet. However, just as I would prefer a 9070XT over a 7900XT I would prefer a 9050S with 32CUs RDNA4 over a 8060S with 40CUs RDNA3.5.

For those speculating as to why the iGPU is RDNA 3.5 and not 4, the answer is simple: die space

I don't undertand why companies keep releasing new products loaded with old tech. RDNA4 will probably have longer-lived driver support which alone is reason to use it. If it's too expensive, reduce the CUs from 16 to 12. Since the CU count can be reduced I don't think die size alone can explain this.

TigerfoxThe whole point of Strix Halo (and Medusa Halo) is IGP-performance. It's a shame already that Strix Halo is not RDNA4.

The article is about Medusa Point, which would be the successor to Strix point. The iGPU performance is nice but it's not the main feature of this range of product.
I haven't seen talks about a "Medusa Halo" other than minor things from MLID, and I don't think this is a credible source at all.

TigerfoxNonsense. You know that RX 9070 is about 27-30% faster than 7800XT in rasterizing, too? It has ~7% less shaders but equally clocks up to 7% higher and has only slightly more bandwith, so you can assume that that's the difference between RDNA3 and RDNA4.

Oh, you want to talk about nonsense? That's cute.

I bring a valid reason and it's nonsense. Meanwhile, you bring up the fact that the 9070 XT does more work than a 7800 XT (which is RDNA 3, not 3.5). You forgot about power draw, how convenient.

Power draw is everything for mobile devices, and the 9070 XT and the 7900 XTX has identical power draw, about the same performance, and about the same efficiency. HW is vastly different and so is manufacturing cost, but that's a given.

So tell me again, how does an RDNA 4 GPU with a 3 % better efficiency than the RDNA 3 GPU, which is negligible, become a better choice than an RDNA 3.5 GPU, IF it means a lot of extra work for AMD?

Whatever comes after RDNA 4 might be a worthy upgrade for mobile, but this is not it.

mrnagantNever really understood why AMD keeps iGPUs behind. Like when RDNA was a thing, they still used GCN for a while in the iGPU space.

Because the newer GPUs are not yet available (in small enough versions) when they start development on the APUs. Any cpu takes several years to get into the market. RDNA4 on it which came out JUST NOW and doesn't even have a smaller version yet. Medusa Point has been mentioned for a few years now, so they could not have started developing it with RDNA4 which was not yet finished.
And it's not possible - well, possible, but not a good idea - to develop both new cpu + gpu both at the same time. So they just develop a new CPU and slap on whatever existing, proven gpu they have.

The better question is, why does it take so much time to release new desktop APUs? It took 3 years to go from Cezanne to Phoenix. Strix Point is already out but there's no news about any new APUs, not even on leaked roadmaps. Will we have to wait until 2027 for a 8600G successor (which will most likely just be a Strix Point port)?

Once again...

FouquinPeople saying RDNA3.5 is outdated don't understand how divergent architectures work. RDNA3.5 is not RDNA3, it's also not RDNA4. It's a focused divergent architecture between the two that is concurrent with both generations. It puts a focus on power and density efficiency that neither of the bigger designs need to bother with to hit their targets.

SL2Meanwhile, you bring up the fact that the 9070 XT does more work than a 7800 XT. You forgot about power draw, how convenient.

I did not compare 7800XT with 9070XT, but with 9070 non-XT, which offers identical computing power as 7800XT (as stated above, 7% less CUs but 7% higher clock) but still is 27-30% ffaster in raster and 43% faster in RT. 7800XT isn't even an efficient RDNA3-card, others have a much better fps-per-watt ratio, but 9070 is 43% more efficient than 7800XT (43% more fps per watt) and 22% more efficient than 9070XT. 9070XT is very much optimized for performance, not efficiency. 9070 is by far the most efficient card compared to 4070 and 7800XT and everything above.

But I agree that this might be not enough to justify upgrading to RDNA4, since it nearly doubles the transistorcount and thus makes costly N4-node necessary. Yet we know nothing about UDNA, so we can only hope it will be more of an upgrade than RDNA4.

ymdhisAny cpu takes several years to get into the market. RDNA4 on it which came out JUST NOW and doesn't even have a smaller version yet. Medusa Point has been mentioned for a few years now, so they could not have started developing it with RDNA4 which was not yet finished.
And it's not possible - well, possible, but not a good idea - to develop both new cpu + gpu both at the same time.

That's not entirely true. Phoenix with RDNA3 came out less than five months after RX 7900XT(X), so they were developed rather close together. It's risky but necessary not to fall behind. AMD has some catching up to do GPU-wise.

ymdhisThe better question is, why does it take so much time to release new desktop APUs? It took 3 years to go from Cezanne to Phoenix. Strix Point is already out but there's no news about any new APUs, not even on leaked roadmaps. Will we have to wait until 2027 for a 8600G successor (which will most likely just be a Strix Point port)

The answer is because desktop-APUs aren't a big market and it's even less important how fast they are. Their only selling point is low price and that's only true for 8400F and 8500G. Even though RAM-OC records are made with Phoenix-APUs right now, there is no real reason to play games on 8700G and invest in fast RAM because 7500F and anything above RX 6400 will be much faster. Only reason is if you really don't have space for anything but the tiniest ITX-cases.

What I don't get is why AMD willingy makes themselves look worse than Intel on everything but halo-laptops by always recycling old APU-generations once or twice. While only the cheapest, lowest-end Intel-laptops come with last generation's CPUs, it is commonplace that in many identical AMD-variants, the refresh of the last generation is used. Compare the last generations of Lenovo Thinkbooks and Thinkpad E- and L-series aswell as HP Probook 400 and Elitebook 600 between Intel and AMD.
While Intel did the same thing this time around - recycling Alder Lake and Raptor Lake as Core 200 without ultra (core i 14000 was a refresh with increased clock and partly corecount atleast), it seems they will be used not nearly as much as the second Phoenix-recyling called AMD Ryzen 200U/H.

They did the right thing by developing a seperate, smaller and cheaper DIE with the same tech as Strix Point with Krakan Point, but now it seems many laptops will be available with two or even three different DIEs.

TigerfoxYet we know nothing about UDNA, so we can only hope it will be more of an upgrade than RDNA4.

We know this:

• Prior to RDNA was GCN. A major difference between the two is that RDNA removed a lot of logic intended for compute so that more transistors could be allocated to gaming tasks.
• Now AMD says this is a problem; RDNA couldn't (until recently) run ROCm and so very few developers that have AMD graphics have access to ROCm and so no one uses it. AMD's stated solution to this is to make ROCm more widely available and to make a new graphics architecture (UDNA) that has more compute-focused resources than RDNA.
• Supporting more features usually means more transistors are needed.

If the PS5 Pro with RDNA 3.5 has FSR4, so will the Medusa Point APU...

Indiana Jones on hand held with ray tracing & upscaling is a milestone.

FouquinOnce again...

IGP fancreatures are the most naive of them all, never realizing how few they are compared to the whole market. It's like talking to a brick wall.

It's the opposite of desktop graphics cards, where pretty much every buyer wants the same things.

TigerfoxI did not compare 7800XT with 9070XT, but with 9070 non-XT, which offers identical computing power as 7800XT (as stated above, 7% less CUs but 7% higher clock) but still is 27-30% ffaster in raster and 43% faster in RT. 7800XT isn't even an efficient RDNA3-card, others have a much better fps-per-watt ratio, but 9070 is 43% more efficient than 7800XT (43% more fps per watt) and 22% more efficient than 9070XT. 9070XT is very much optimized for performance, not efficiency. 9070 is by far the most efficient card compared to 4070 and 7800XT and everything above.

Yeah, I missed the lack of XT there, but like I tried to explain, none of that matters as we don't know where mobile RDNA 3.5 fits into all this. The decimal part of 3.5 doesn't reflect on whether it's closer to RDNA 3 or 4. It's just something different, optimized for mobile (just like Fouquin said)

ymdhisThe better question is, why does it take so much time to release new desktop APUs? It took 3 years to go from Cezanne to Phoenix.

WTF are you even talking about.

866 days between mobile Cezanneand Phoenix, and 1000 days between AM5 Cezanneand Phoenix. That's 4.5 months longer, is that something to whine about? :roll:

AM5 APU's are not a priority, and with soldered APU's becoming more popular in desktops, that won't change. If you want the best graphics performance out of any Strix you will want soldered LPDDR RAM anyway.

Tigerfox[..] It's a shame already that Strix Halo is not RDNA4. [..]

Right, good point. *Another reason so skip*

The 128 GB RAM of Strix Halo are not enough to load even the smallest real (=not the "-Distill-" ones are just Qwen / Llama fine-tunes) DeepSeek-R1 quant either. A Strix Halo successor with 256 GB RAM and (at least) double the bandwidth (>500 GB/s) is what many are waiting for.

SL2Power draw is everything for mobile devices, and the 9070 XT and the 7900 XTX has identical power draw, about the same performance, and about the same efficiency. HW is vastly different and so is manufacturing cost, but that's a given.

Basically right: It's 304W vs 355W and according to some sources, the power efficiency improved by like 7-15%. According to this (would prefer a test of a non-OC card), it is indeed the same performance (well, 1% difference): If we assume the TDP limit was fully utilized: 355W/304W = 1.168 -> 16.8% and 23% in ray-tracing. Both percentages aren't even generational improvements (if we define a generational improvement between 25% and 33% more FPS per Watt).

I see, so this is also Medusa Halo related (Strix Halo successor). Medusa Halo rumor is that it may also come in a 384-bit version (number are scaled up to that):

Strix Halo	256-bit	up to 128 GB RAM	256 GB/s (LPDDR5X-8000) (256 = 8000 [MT/s] * 64 [bit/channel] * 4 [channels] / 8 [byte] / 1000 [MB -> GB])
Medusa Halo	384-bit	up to 192 GB RAM	384 GB/s
APPLE M4 Max	512-bit	up to 128 GB RAM	546 GB/s (LPDDR5X-8533)
APPLE M3 Ultra	1024-bit	256 GB RAM	819 GB/s (LPDDR5-6400)
APPLE M3 Ultra	1024-bit	up to 512 GB RAM	819 GB/s (LPDDR5-6400)

I wish Medusa Halo would be double that of Strix Halo. APPLE is offering up to 512 GB RAM in Mac Studio.

Edit:

10tothemin9voltsRight, good point. *Another reason so skip*

The 128 GB RAM of Strix Halo are not enough to load even the smallest real (=not the "-Distill-" ones are just Qwen / Llama fine-tunes) DeepSeek-R1 quant either. A Strix Halo successor with 256 GB RAM and (at least) double the bandwidth (>500 GB/s) is what many are waiting for.

Let me add, why at least 256 GB RAM: According to this (which was posted after my post, but it's kinda common knowledge by now), recommended DeepSeek-V3 / -R1 quants are 2.42bit (203GB) / 2.71bit (231GB) and above.

M4 max taps out at 128GB right now. So about half of your chart.

10tothemin9voltsI see, so this is also Medusa Halo related (Strix Halo successor). Medusa Halo rumor is that it may also come in a 384-bit version (number are scaled up to that):

Strix Halo	256-bit	up to 128 GB RAM	256 GB/s (LPDDR5X-8000) (256 = 8000 [MT/s] * 64 [bit/channel] * 4 [channels] / 8 [byte] / 1000 [MB -> GB])
Medusa Halo	384-bit	up to 192 GB RAM	384 GB/s
APPLE M4 Max	512-bit	up to 128 GB RAM	546 GB/s (LPDDR5X-8533)
APPLE M3 Ultra	1024-bit	256 GB RAM	819 GB/s (LPDDR5-6400)
APPLE M3 Ultra	1024-bit	up to 512 GB RAM	819 GB/s (LPDDR5-6400)

I wish Medusa Halo would be double that of Strix Halo. APPLE is offering up to 512 GB RAM in Mac Studio.

The base price of the M4 Max in a Macbook is $3200 with 36GB of memory. Asus charges $2800 for the Z13 Flow with Strix Halo and 128GB of RAM. To get an Apple silicon laptop that cheap you'd have to dip into the M4 Pro territory, which I can't find specs for but if it's like the M3 Pro has a 192-bit interface and it has 273 GB/s of bandwidth (which does suggest a 256-bit interface unlike the M3 Pro). No Apple device will have 128GB of RAM for $2800.

Also Strix Halo has "Infinity Cache"; 32MB of GPU cache which allows it to perform as if it had more memory bandwidth.

SquaredThe base price of the M4 Max in a Macbook is $3200 with 36GB of memory. Asus charges $2800 for the Z13 Flow with Strix Halo and 128GB of RAM. To get an Apple silicon laptop that cheap you'd have to dip into the M4 Pro territory, which I can't find specs for but if it's like the M3 Pro has a 192-bit interface and it has 273 GB/s of bandwidth (which does suggest a 256-bit interface unlike the M3 Pro). No Apple device will have 128GB of RAM for $2800.

According to the Wiki page' table in my quoted post, the M4 Pro' 273 GB/s come indeed from 256-bit ("256-bit" is stated in the table):
273 GB/s = 8533 * 64 (bit per channel) * 4 (channels) (64-bit * 4 is where the 256-bit come from) / 8 / 1000.
APPLE's own site also lists 273 GB/s.

MacBook Pro with 48GB RAM (M4 Pro variant, 256-bit) - 3100 $: 273 GB/s.
MacBook Pro with 48GB RAM (M4 Max 16-core variant, 512-bit) - 4000 $: 546 GB/s (= 8533 * 64 * 8 / 8 / 1000).
So, with the same 48GB RAM, the M4 Max variant costs 29% more, but is also double the memory speed.
It also appears that all MacBook Pro M4 Max 48GB, 64GB and 128GB RAM variants have the same memory bandwidth / speeds (to get the same RAM speeds across all RAM size variants, it looks like different densities per LPDDR5(X) chip are used).

Yep, the elephant in the room is the APPLE tax. The 128 GB RAM MacBook Pro costs 5000 bucks.
Basically AMD' Strix Halo and APPLE use the same memory LPDDR5(X) chips, but APPLE's wants 79% (1.79 = 5000 $ / 2800 $) more for the same 128GB RAM capacity.