AMD's Ryzen AI 300 Series Mobile APUs Should Be Interesting For Next-Gen Laptops

[phoronix]

Phoronix: AMD's Ryzen AI 300 Series Mobile APUs Should Be Interesting For Next-Gen Laptops

In addition to announcing the AMD Ryzen 9000 series desktop processors powered by Zen 5, Lisa Su at Computex 2024 also announced the AMD Ryzen AI 300 series as the next-generation mobile processors powered by Zen 5 CPU cores while sporting RDNA 3.5 (also referred to as RDNA 3+ and RDNA3 refresh) integrated graphics and an XDNA 2 NPU...

https://www.phoronix.com/news/AMD-Ryzen-AI-300-Series-Zen-5

[arunbupathy]

Hardware lives and dies by the software support. AMD, please open up your Neural engines to tinkerers, fast!

[sophisticles]

The top-end model announced for Computex 2024 is the AMD Ryzen AI 9 HX 370 with 12 cores / 24 threads, a 5.1GHz maximum boost frequency, 36MB cache, 50 TOPS NPU, and Radeon 890M graphics. The AMD Ryzen AI 9 365 was also announced as a 10-core / 20-thread Zen 5 mobile APU with 5.0GHz maximum boost frequency, 34MB cache, 50 TOPS NPU, and Radeon 880M Graphics.

With the AMD XDNA 2 neural processing unit, AMD is talking up big gains with up to 5x the compute capacity and up to 2x the power efficiency. AMD is claiming that their 3rd Gen AMD Ryzen AI will be able to outperform not only the Intel Meteor Lake by wide margins but also the Apple M4, the anticipated performance of Intel Lunar Lake, and also the Qualcomm Snapdragon Elite X.

Of the 3 announced product lines, this to me is the most exciting.

Long before AMD went down the "more cores" road, they were talking up heterogeneous computing, about how they were going to leverage their APU's so that the CPU and GPU worked together as one.

They even advertised a "12 core" APU on their website, 4 CPU cores and 8 GPU cores and they helped add OpenCL acceleration to Open Office.

It's too bad that Lisa Su took then down the road she did, imagine a Ryzen AI 9 HX 370 with software that was able to fully utilize the CPU and GPU the way AMD had claimed they wanted to see happen.

[shinger]

Very impressive. The waiting indeed paid off with even this apu. An igpu with 16 cu good to have for retro gaming to be sure that even ps4 games are fairly playable.

[drakonas777]

It's too bad that Lisa Su took then down the road she did, imagine a Ryzen AI 9 HX 370 with software that was able to fully utilize the CPU and GPU the way AMD had claimed they wanted to see happen.

Focusing on the consumer segment heterogeneous bullshit was one of the main reason why AMD went near bankruptcy. It was a major mistake by Rory Read. A lot of people in this forum do not understand that a geek fetish theory and economical semicon/hardware reality often do not come hand in hand.

[Jedibeeftrix]

looking forward to a thin-n-light 13" laptop sporting one of these!

the Z13 from Lenovo should do nicely.

[uid313]

I think they are trying to scam us by hoping we fall for the AI hype by renaming their line from Ryzen to Ryzen AI.
I don't really care about AI performance since I consider it a niche use case, it's like AVX-512, I don't care, it's not useful for me. I care more about performance for compilation and gaming.

Also it seems they are trying to scam us by showing us benchmarks where compare their CPU against other CPUs with a different TDP. I don't care how Ryzen performs against Snapdragon X Elite, Apple M4 or Intel Meteor Lake at a different TDP, it tells me nothing.

How much performance does it have per watt? ... oh, it got silent.

[Jumbotron]

The statement that this APU will outperform an Apple M4 is Intel level bullshit.

Firstly: This APU will never achieve that claim without good software, compiler support. firmware, driver and vendor support. All of which AMD historically sucks at going all the way back to 3DNow and through the Fusion HSA era into today with ROCm and XDNA. Apple will always and forever more have the better software, firmware and driver support and optimizations simply because they do it all in house . Which means better and more sustained performance over the time that those services are needed from hardware and into the future as Apple’s support of their own hardware is years longer than anything in the x86 world much less the Android world.

Secondly: This APU will never achieve those top line performance numbers at the TDP of an Apple Silicon chip. Full stop. This is why Microsoft went with Qualcomm and not AMD nor Intel for their Surface clone of the MacBook Pro.

The “Compute Industrial Complex” has spoken. They chose Intel’s CXL protocol for providing NUMA for CPUs to all other hardware inside the computer and out to various racks. They chose Intel’s OneAPI to program and communicate with various accelerators inside the computer and out to the various racks through the UXL framework as a counter to Nvidia’s CUDA. They did NOT choose AMD’s ROCm. Where they did choose AMD they chose AMD’s Infinity Fabric now called UAlink to tie together GPUs in a NUMA fashion from GPU pod to GPU pod in similar fashion to what HP’s Gen-Z interconnect had in mind for racks before they and everyone else gave over their various protocols to Intel’s CXL standard body.

But that’s it for AMD concerning widespread industry support.

[Jumbotron]

Focusing on the consumer segment heterogeneous bullshit was one of the main reason why AMD went near bankruptcy. It was a major mistake by Rory Read. A lot of people in this forum do not understand that a geek fetish theory and economical semicon/hardware reality often do not come hand in hand.

And yet every Apple Silicon chip from the M1 to the M4 is precisely what AMD had at one time with their Fusion APUs, particularly Carrizo and its ultimate and final refinement with Bristol Ridge. And partly because Apple perfected what AMD started they are a 3 trillion dollar valuated company with more actual liquid cash on hand than either Intel or AMD. So much for “geek fetish theory and economical semicon/hardware reality often do not come hand in hand.”

[agd5f]

Hardware lives and dies by the software support. AMD, please open up your Neural engines to tinkerers, fast!

It's been open source for a while. We are working on getting it uptream as well.

https://github.com/amd/xdna-driver