June 3,$Advanced Micro Devices (AMD.US)$Chairman and CEO Su Zifeng announced the roadmap for a new cloud-based AI acceleration chip during the opening keynote speech at the Computex 2024 exhibition. The new Instinct MI325X will be launched this year. At the same time, AMD also released the third-generation AI PC chip “Ryzen AI 300 Series,” codenamed “Strix Point,” and the AMD Ryzen 9000 series desktop processor (Granite Ridge).

The MI325X will be launched this year, and the performance is 1.3 times higher than the Nvidia H200

According to the new cloud-based AI accelerator roadmap announced by AMD, AMD will launch the new AI accelerator Instinct MI325X this year, the MI350 in 2025, and the MI400 in 2026.

According to reports, the MI325X will continue the CDNA3 architecture and use the fourth-generation high-bandwidth memory (HBM) HBM3E. The capacity will be greatly increased to 288GB, and the memory bandwidth will also be increased to 6Tb/s. Overall performance will be further improved. Other specifications will basically remain consistent with the MI300X, making it easier for customers to upgrade their products.

Su Zifeng pointed out that the AI performance improvement of the MI325X is the biggest in AMD history. Compared with the competitor Nvidia H200, it will be more than 1.3 times better, and at the same time more cost-effective, the MI325X will be delivered in the fourth quarter of this year.

In addition, AMD will also launch a new generation of MI350 series in 2025. This series of chips will use a 3nm process, integrate 288GB HBM3E memory based on a new architecture, and support FP4/FP6 data formats. The inference calculation speed is 35 times faster than the existing MI300 series chips.

According to supply chain information, the MI350 will enter the liquid cooling generation to provide stronger computing power and better energy efficiency for AI server computing.

In 2026, AMD will also launch a new generation of MI400 series, maintaining the pace of annual updates.

AMD Ryzen AI 300 series released: AI computing power up to 50 TOPS!

As early as early 2023, AMD released the Ryzen 7040 series (codename Pheonix), the world's first x86 processor with an independent NPU AI engine. Based on the newly designed XDNA architecture, the computing power is about 10 TOPS (10 trillion floating point operations per second). In addition, the overall computing power of the CPU and GPU is about 33 TOPS, creating a new era of AI PCs.

At the end of 2023, AMD also launched the Ryzen 8040 series (codename Hawk Point). The NPU AI computing power increased by 60% to about 16 TOPS, and the overall computing power also increased to 39 TOPS.

The latest “Ryzen AI 300 Series,” codenamed Strix Point, uses a new Zen5 PU architecture, and the GPU core has also been upgraded to an RDNA3.5 architecture. The NPU is also a new XDNA2 architecture, claiming to be a “world-class processor for the next generation of AI PC/Copilot+ PCs.”

There are only two models in the Ryzen AI 300 series. “Ryzen AI 9 HX 370” and “Ryzen AI 9 HX 365” are both positioned in the high-end market.

Among them, the “Ryzen AI 9 HX 370” is the top flagship. The CPU part has 12 cores and 24 threads. For the first time in many years, it provides more cores in this position, an increase of up to 1/3 compared to the Ryzen 8040 series. The second-level cache is still 1MB per core, and the total capacity naturally increases to 12MB. The Level 3 cache finally broke the “ban” of 16MB and increased by half to 24MB. The highest main frequency is 5.1 GHz. The GPU section not only upgraded the architecture, but also increased the number of CU units from 12 to 16, named “Radeon 890M”. The NPU computing power was increased to 50 TOPS, more than 2 times more.

The “Ryzen AI 9 365” is also a high-end model. It has 10 cores and 20 threads, a second-level cache of 10MB, and a third-level cache of 24MB. The maximum frequency is 5.0 GHz. The NPU computing power is also 50 TOPS, and the GPU part is also Radeon 890M.

From the comparison chart of AMD's first and second generation XDNA NPU architectures, it can be seen that the overall layout is basically the same, but the scale has been greatly expanded.

AI Tile (first generation called AIE Tile) is also the core AI calculation engine module. It has increased from the previous 20 to 32, plus its own enhancements.

Mem Tile is also known as a local memory module. It has been increased from 5 to 8, which can better match larger scale local scheduling and computation.

In addition, the cross bus used for interconnection was upgraded from an ordinary data fabric to an ubiquitous Infinity Fabric in the Zen/RDNA family, with higher transmission bandwidth and efficiency.

AMD claims that the XDNA2 NPU has increased its computing power by up to 5 times, its ability to multitask in parallel has doubled, and its energy efficiency has also been increased by up to 2 times.

It should be pointed out that the AI computing power mentioned here is increased by 5 times, which means the response speed of Llama's 27 billion parameter model. From launch to obtaining the first token, the Ryzen AI 9 HX 370 reached up to 5 times that of the Ryzen 9 8940HS.

Another very important point is that xDNA2 first introduced a new block FP16 floating point accuracy, that is, bFloat16 and BF16. It's already very common on CPUs and GPUs, and it's the first time on NPU.

The traditional FP8 floating-point format has high performance and insufficient accuracy. The FP16 floating-point format has high accuracy and slightly poor performance, while BF16, which combines the two, can achieve a good balance of accuracy and performance, and is also more flexible.

At the same time, most AI applications use 16-bit accuracy, so with BF16, quantization to 8-bit accuracy is no longer required, reducing conversion steps, and improving execution efficiency.

The Qualcomm Snapdragon X Elite NPU has 45 TOPS of computing power. Intel's upcoming next-generation Core Ultra Lunar Lake NPU also has a computing power of 45 TOPS, and the Ryzen AI 300 series surpassed both in one fell swoop, making it the strongest NPU today.

As for the Apple M4, its NPU's computing power is only 38 TOPS, which is less than the minimum computing power requirement of 40 TOPS for Copilot+ PCs in the Windows camp.

With the drastic increase in computing power, the application of NPU will no longer be limited to some continuous low-load scenarios (such as video conferencing), but will have more possibilities. On the one hand, it can partially replace CPUs and GPUs in more scenarios to perform AI computations with higher energy efficiency, greatly improving the battery life of notebooks.

On the other hand, with a more powerful NPU and a more powerful CPU and GPU, end-side AI can be deployed in more scenarios to further eliminate reliance on the cloud side. The biggest benefit is that privacy leaks and security threats can be avoided.

Of course, no matter how powerful the hardware computing power is, it also requires the implementation and cooperation of ecological applications. As a pioneer of AI PCs, AMD will have more than 150 ISV partner manufacturers within 2024, including world-class manufacturers such as Adobe, Microsoft, and Topaz Labs, as well as famous domestic companies such as Baichuan Intelligence, DingTalk, Muzankong, and Youdao. The future is limitless.

The biggest highlight of the new generation of AI PCs is that with the latest Windows system, it can create a new Copilot+ experience, such as historical memories, real-time video conference recording and translation, and collaborative creation.

According to officially provided data, Ryzen AI 9 HX 370 is far ahead in everyday office, productivity creation, multitasking, graphics, etc. compared to Qualcomm Snapdragon X Elite. In particular, graphics computing performance is 73% higher.

Compared to Intel's current strongest Core Ultra 9 185H, both everyday applications and games have been greatly improved, with an average performance increase of 36%. Just look at how Intel's next generation Lunar Lake performs.

Compared to Apple M3, Ryzen AI 9 HX 370 also has great performance advantages in all aspects, especially in 3D performance, which exceeds 98%.

Ryzen AI 300 series notebooks will be launched one after another starting in July. Currently, there are more than 100 designs, covering major OEM brands such as Acer, Asus, Dell, HP, Lenovo, and MSI.

Asus will announce a series of notebooks equipped with Ryzen AI 300 series processors at the Taipei Computer Show. Among them, the lightweight ones include the 16-inch Zenbook S (Zenbook S), the 14/15/16-inch Vivobook S (Vivobook S), the 16- and 13-inch ProArt P16/X13, and the game books include the 16-inch ROG series and the 14/16 inch Tianxuan series (TUF GAMING A14/A16).

The first three MSI models are all 16-inch large-screen models, including the Summit A16 AI+ for high-end business offices, the Superstar A16 AI+, a lightweight and all-purpose gaming model, and the Medallion A16 AI+, which focuses on ultra-thin business and creation.

AMD Zen5 Ryzen 9000 Series Released: 56% Faster Performance Than i9-14900K

The newly released AMD Ryzen 9000 series desktop processors (Granite Ridge) are based on the Zen5 architecture, and the first batch of products will be launched in July 2024.

Su Zifeng emphasized that the Ryzen 9000 series is the third series of AM5 slots after the Ryzen 7000 “Raphael” and Ryzen 8000 “Hawk Point” series. It will be equipped with two small Zen5 chips, each with a maximum of 8 cores and a maximum of 16 cores and 32 threads, similar to the Ryzen 7000 series. AMD also continues to support SMT (simultaneous multi-threading feature).

AMD first launched four CPU products, all equipped with 2CU RDNA2 display:

Ryzen 9 9950X is equipped with 16 cores, 32 threads, 170W TDP, maximum frequency 5.7 GHz;

Ryzen 9 9900X is equipped with 12 cores, 24 threads, 120W TDP, maximum frequency 5.6 GHz;

Ryzen 7 9700X is equipped with 8 cores, 16 threads, 65W TDP, maximum frequency 5.5 GHz;

Ryzen 5 9600X is equipped with 6 cores, 12 threads, 65W TDP, and a maximum clock frequency of 5.4 GHz.

According to official AMD tests, the IPC performance of the Zen 5 core for PC platforms improved by an average of about 16% compared to Zen 4.

Compared with the Intel Core i9-14900K, the AMD flagship 16-core Ryzen 9 9950X was 4% to 23% faster in gaming performance tests. In productivity performance tests, the 9950X showed a greater advantage, 7% to 56% faster than the Intel Core i9-14900K.

AMD announced that the Ryzen 9000 series will be launched on July 31, and specific pricing details will be announced soon.

Editor/Somer