AMD launches Ryzen AI 400 desktop processors with up to 50 TOPS NPU and Copilot+ support

[Redaktion]

The AMD Ryzen AI 400 Series, announced at MWC 2026, are officailly the world's first desktop chips designed to support Microsoft Copilot+ PC experiences. Like AMD's recent Ryzen AI mobile platforms, this one also features Zen 5 CPU cores, AMD RDNA 3.5 graphics, and an XDNA 2 NPU.


https://www.notebookcheck.net/AMD-launches-Ryzen-AI-400-desktop-processors-with-up-to-50-TOPS-NPU-and-Copilot-support.1239340.0.html

[Prassel]

So it shows that AMD, like Intel, is "allergic" to desktop APUs ...

Their goodwill for desktop APUs only goes as far as approximately GTX 1050 Ti graphics power. Sad. APUs could completely remove the need for any dGPU if you keep your expectations in gaming at low settings (but not necessarily ugly settings). Add to that the possibility to easily upgrade the chip by swapping in and out the socketed chip - would have been great. Imo as a custom ITX build better than a NUC (because of standard coolers, standard PSUs, etc)

But "they" don't want that. Both AMD and Intel. They don't want that flexibility for customers. Obviously.

[captainobvious]

Very true, neither company is interested in selling top iGPU to customers.
Coincidentally, neither am I interested in buying their monster graphics cards with hundreds of watt power consumption just to play the latest "Silly man-animal of fluid gender" FPS game.

[Now you understand]

For those who still don't know: Ryzen AI 400 is just a rename of Ryzen AI 300.

So it shows that AMD, like Intel, is "allergic" to desktop APUs ...

Their goodwill for desktop APUs only goes as far as approximately GTX 1050 Ti graphics power.

Very true, neither company is interested in selling top iGPU to customers.

Strix Halo' iGPU has access to 256 GB/s memory due to its 256-bit RAM access and its performance is approx on the level of a 4060 Laptop. 256 GB/s equals to quad-channel (4 * 64-bit per RAM channel) in the desktop. In desktop (if we define the word "desktop" as anything that looks like a (big) desktop tower), only the (entry level) Threadripper platform offers quad-channel.

So, there's simply not enough bandwidth.

PS: Panther Lake's Arc B390 iGPU is connected to 128-bit (dual-channel) memory running at LPDDR5X-9600 MT/s, but it has bad 1% FPS lows (YT/Just Josh: ExpertBook Ultra Panther Lake youtu.be/jduWl1J_4lQ?t=630). I wonder what is up with that and if it's caused by the LPDDR5X' high latency when running it at 9600 MT/s. Now I wonder if Strix Halo also has bad 1% FPS lows issue.

[Now you understand]

Here is another with bad 1% FPS lows in Panther Lake's Arc B390: youtu.be/A_jHa0d1aFA?t=319 (YT/Just Josh: "Intel's New Integrated GPU vs 10 Popular Games (Shockingly Playable)")

..we wouldn't recommend it. The 1% lows and .1% lows are just too rough.

[Prassel]

Strix Halo iGPU ... not enough bandwidth.

Uhm ... "Halo" is not the point. It wouldn't fit on a default socket anyways.


I'm talking about the smaller chips, like the socketed 8000G-series, which has up to 12 CUs for the iGPU.

And now the "successor" comes as the socketed 450G-series, which gets 8 CUs only... and even though the same small chip in laptops has 16 CUs (Radeon 890m).

[NVL-S]

Prassel,

Probably not the answer you're looking for but in the entire history of desktop APUs, I can only think of maybe once or twice where graphics performance was a priority and taken seriously.

I think this was Intel's Crystal Well (22nm Haswell processor) and later 14 nm Broadwell with on-package 128MB eDRAM L4 Cache.

You need faster ram but socketed ram is always slower which is why the focus is on laptops because soldered there is more popular there.

[Ryzen AI PRO 400 lineup]

Ryzen AI 9 HX PRO 475
Ryzen AI 9 HX PRO 470
Ryzen AI 9 PRO 465
Ryzen AI 7 PRO 450
Ryzen AI 5 PRO 440
Ryzen AI 5 PRO 435

"In addition to the desktop launch, AMD also announced the Ryzen AI PRO 400 Series mobile processors designed for enterprise-class mobile workstations like laptops and tablets. The company mentions that the top-of-the-line Ryzen AI 9 HX PRO 470, with its 50 TOPS AI power, delivers 30% faster multithreaded performance compared to the competition, along with all-day battery life."

[Now you understand]

Strix Halo iGPU ... not enough bandwidth.

Uhm ... "Halo" is not the point. It wouldn't fit on a default socket anyways.


I'm talking about the smaller chips, like the socketed 8000G-series, which has up to 12 CUs for the iGPU.

And now the "successor" comes as the socketed 450G-series, which gets 8 CUs only... and even though the same small chip in laptops has 16 CUs (Radeon 890m).

Ok, let's calculate the maximum performance you could expect from your desktop PC: A normal/typical desktop PC has its RAM connected to 128-bit (2*64-bit per RAM channel, aka dual-channel) wide memory bus and the dual-channel RAM is typically running at 5600 to maaaybe 6400 MT/s:

Strix Halo (4060 Laptop level of performance): 256 GB/s = 256-bit * 8000 MT/s / 1000 / 8.
Typical dual-channel DDR5 desktop PC: 96 GB/s = 128-bit * 6000 MT/s / 1000 / 8.
256 GB/s / 96 GB/s -> Expect a maximum iGPU performance that is 2.67 times slower than a 4060 Laptop.

Let's compare -- 3dmark.com/search (Time Spy):

• 8700G (780M (12 CUs)): ~3300
• Radeon 880M iGPU (12 CUs): "Average score: 3473"
• 4060 (notebook): "Average score: 10506"

(the factory of roughly 3 lines up)

• And now desktop CPU iGPU perf: "Radeon Graphics" iGPU (Ryzen 7000 & 9000 series) (2 CUs): "Average score: 826"

While 3 times slower than a 4060 Laptop doesn't sound like fast, such an iGPU in a desktop CPU would still be roughly 4 times (3502/826) faster.
(names from: en.wikipedia.org/wiki/RDNA_2#Integrated_graphics_processing_units_(iGPUs))

You are right, the memory bandwidth would allow for a (roughly) 4-time increase in performance vs the current 2CUs iGPU ones and putting such an iGPU into a 9000G series would make sense for those, where a 4-time perf increase is all they need.

AMD is prob not prioritizing a G-series CPU because even a, say, 3500 score in Time Spy isn't that much when compared vs e.g. a RX 6600 (non-XT) ("Average score: 8040") that can be bought (used) cheaply (and undervolted, power-scaled, too). Yes, such a dGPU PC is going to be less slim (or is harder to build slim (there are low profile GPUs, tho)), but is also 2.4 times faster (8040/3300). And the GPU can be upgraded vs in a CPU you'd have to change the whole CPU.
Or the sales of the 8000G series were bad.

Let's take a newer GPU architecture (thanks to TSMC's better node) that has a better perf/W than RDNA2 (and other features):

• RX 6600: 132W TDP and "Average score: 8040" = 60 points per Watt.
• RX 9060 XT: 150W TDP and Average score: 14601 = 97 points per Watt.

(the smallest non-OEM RDNA4 GPU is the RX 9060 XT and it's 150W TDP)

• RTX 4060: 115W TDP and "Average score: 10398" = 90 points per Watt.

Being able to simply upgrade the (low profile) GPU and not needing to buy a new CPU, is understandably a very big thing.

Even if one would prefer AMD, it's NVIDIA that offers the lower TDP card (115W vs 150W).

I have a 16400 Time Spy score GPU, so I'm not in the iGPU CPU market, but I'd still find it nice if there were variants with more than 2 CUs, but only if this wouldn't pull AMD's other important resources much.

[Prassel]

Ok, let's calculate the maximum performance you could expect from your desktop PC

No. You don't have to calculate anything. Performance is roughly known already. The Radeon 890m already exists with 16 CUs (iGPU), but the upcoming desktop equivalent only gets 8 CUs, which obviously is much slower. And a joke to many AMD users, specifically because AMD made itself known by having good iGPUs in the standard socketed desktop CPUs (5700G, 8700G).

I'm also not sure what your textwall wants to prove, as it's mostly offtopic, and btw. looks like it's coming straight from ChatGPT, and you even admitted you have no idea about iGPUs. Not sure why you are even posting here ...

I have a 16400 Time Spy score GPU, so I'm not in the iGPU CPU market

[Prassel]

Probably not the answer you're looking for but in the entire history of desktop APUs, I can only think of maybe once or twice where graphics performance was a priority and taken seriously.

You need faster ram but socketed ram is always slower which is why the focus is on laptops because soldered there is more popular there.

The entire point of a desktop APU is the focus on a better iGPU, which made specific AMD CPUs popular, like the 5000G or 8000G series, coming in a standard, swappable, socketed CPU chip.

Also it's wrong that socketed RAM is slower than soldered RAM. Check benchmarks of soldered 6400 MT/S LPDDR5x versus 6400 MT/s CSODIMM, and you'll see they're practically equal.

[NVL-S]

I don't really consider 6400 MT/s as fast. More like 9600 or at least 8533.

They may have been marketed that way but looking at raw specs they've always been woefully lacking compared to modern budget dgpus and never been able to replace them. It's always been that way.

I've always seen them, as just slightly faster office igpu pcs with a bit better performance at rendering PowerPoint slides. IMO, the G series has always been a very niche market.

Whenever a fast desktop APU was made it was just too overpriced which perpetuated the stereotype that nobody is interested in them.

Doesn't help that the type of memory that CPU and GPU memory require is completely different. CPUs require low latency (which is typically low bandwidth) and GPUs require high bandwidth (which typically is high latency).

On consoles, the high latency GDDR setup isn't a big deal because the OS and games are coded for it but windows isn't. Its a completely different platform.

Anyways, things are changing with LPCAMM2 memory, maybe when that becomes more mainstream things will change for the better...

[Walls of text]

The Radeon 890m already exists with 16 CUs (iGPU), but the upcoming desktop equivalent only gets 8 CUs, which obviously is much slower.

If it makes you feel any better, future laptop mainstream apu zen6 medusa point (990m?) will also have the half the CUs. So at least you don't feel left out this time and it's fair xD.

[Prassel]

If it makes you feel any better, future laptop mainstream apu zen6 medusa point (990m?) will also have the half the CUs. So at least you don't feel left out this time and it's fair xD.

I don't own an AMD computer, but the potential of a standard socketed CPU with decent iGPU is huge. Unfortunately neither AMD nor INTEL seem to like that, probably because it would be too flexible for the customer.

In some way I'm also "relieved" that AMD now plays the same §hitshow as INTEL - which is that they both stopped making iGPUs better for desktop CPUs, and only offer better versions as soldered in laptops or mini-pcs. Now people (especially AMD supporters) can really see both companies use the same tactics and don't care a f... about them.