AMD Blitzes Computex With AI, From Mobile To Desktops And Data Centers

In the resounding buzz that AI stirred at Computex 2024 in Taiwan, literally all of the major semiconductor players had something to say, though not all had new technologies that will shape this burgeoning new landscape. AMD, on the other hand, stepped out with a full arsenal of new chip technologies targeted at AI-fueled computing from the client edge to gaming desktops and workstation, and the AI-infused data center. Here’s the 5,000 ft view.

Ryzen AI 300 Series Brings A 50 TOPS NPU And Big Compute For Copilot + PCs

AMD announced a new generation of its Strix Point Ryzen laptop processors, the Ryzen AI 300 Series, specifically designed to optimally handle AI PC workloads, content creation and productivity for thin and light laptop designs. With a new 50 TOPS rated NPU on board, this series ushers in a rebranding of AMD’s mobile chips, but with a continuation of the HX suffix designating top-tier performance.

Ryzen AI 300 Series processors are supported by AMD’s XDNA 2 NPU architecture, offering 50 TOPS of AI processing throughput. This engine is still based on Xilinx FPGA technology and this family of processors will surpass Copilot+ AI PC requirements, with a 3X performance lift over AMD’s previous gen mobile CPUs. Built on AMD’s, new Zen 5 and Zen 5c CPU architectures, these mobile CPUs are equipped with 10 or 12 high-performance CPU cores, with support for 20 and 24 threads, respectively, and up to 50% more on-chip L3 cache memory, versus AMD’s previous gen Zen 4-based chips.

It’s going to be very interesting to watch the performance compares, as machines powered by AMD’s new Ryzen AI 300 Series come to market. I saw Acer, ASUS, HP, Lenovo and MSI machines on display at Computex, and AMD claims the platform takes both Intel Meteor Lake and Qualcomm Snapdragon X Elite behind the woodshed, for anything from AI-enhanced productivity to multitasking and graphics. Where the rubber meets the road, however, will also be battery life. We’ll know more when the product launches in July this year.

Zen 5 Based Ryzen 9000 Desktop Processors Promise Big Gains

Code named Granite Ridge, AMD also announced new Zen 5 architecture based Ryzen 9000 Series processors for the desktop. This is a new CPU core architecture, with more details forthcoming from AMD, along with improved branch prediction, wider data pipelines and a wider window size for better parallelism. These chips will also reportedly offer 2X the AI performance and AVX512 throughput, for vector processing requirements in applications like scientific simulation, financial analytics, machine learning and 3D modeling.

The company is claiming approximately a 16% average IPC uplift for PCs over its Zen 4-based Ryzen desktop chips, with larger gains of up to 23% in gaming, 3D rendering and content creation. Lightly threaded workloads like web browsing will see a 15% lift as well, which speaks to Zen 5’s strong single threaded throughput for bursty responsiveness. The chips will be socket compatible with current AMD AM5 motherboards, though the company announced new X870/X870e motherboard chipsets with USB 4.0 and PCI 5.0 support standard on all models.

Ryzen 9000 Series processors will range in core counts from Ryzen 5 products at 6-cores and 12 threads to Ryzen 7 8-core / 16-thread CPUs and Ryzen 9 products with 12 and 16-core configs and 24 and 32-thread support, respectively. I’ll be very keen to see how the new Ryzen 9 9950X 16-core top dog chip performs, and will be eager for the day when a 3D V-Cache enabled Ryzen 9 9950X3D is announced, though that’s just a bit of informed speculation on my part at this point. Regardless, we can expect Ryzen 9000 Series desktop processors to arrive in July this year as well.

AMD Flexes Zen 5 Data Center CPUs, Instinct MI325 GPU And Radeon Pro Workstation Graphics

Not leaving big iron platforms out of the conversation, AMD also announced new Zen 5-based Epyc data center CPUs, code named Turin. The company is ratcheting up core density per socket competition with a 192-core 384-thread chip that’s socket compatible with its current-gen Genoa platform. Word is AMD’s Zen 5C architecture in Turin packs sixteen cores into a single Core Chiplet Die, which should substantially improve multi-core throughput and multithreaded performance.

AMD notes Epyc Turing processors will be available in the second half of this year, with big claims of AI throughput performance advantages over Intel’s currently in-market 5th Gen Xeon 8592+ chips, in two socket server configurations. The truly interesting match-up, however, will be versus Intel’s forthcoming Granite Rapids Xeon 6 products with up to 128 Performance Cores, when they arrive later this year in Q3.

On the AI accelerator front, AMD also unveiled its MI325X follow-up to its popular MI300X that the company notes is quickly ramping up to a $1B product, reportedly the fastest ramp AMD’s history. This is no surprise perhaps, as AI demand has skyrocketed, and this product, along with Intel’s Gaudi accelerators, are one of the few alternatives to Nvidia GPUs. At any rate, MI325X is based on the same AMD CDNA 3 GPU architecture, but now sports 288GB of HBM3 memory (vs 192 on board MI300X) clocked higher as well, for 6TB/s of peak bandwidth. This will serve to boost Large Language Model throughput significantly, as AI workloads, in general, are extremely memory and bandwidth intensive. AMD notes a CNDA 4 MI350 will arrive in 2025 and CDNA 6 is slated for 2026, marking an aggressive annual cadence of new data center GPU architecture releases. In the meantime, MI325X will reportedly arrive later this year.

Among its various unveilings at Computex, the last but not least bit of hardware news AMD announced was the arrival of its a new version of its Radeon Pro W7900 Series workstation AI and 3D rendering GPU accelerators, with a more compact Radeon Pro W7900 Dual Slot card, strapped with 48GB of GDDR 6 ECC memory, versus the original Radeon Pro W7900 triple-slot design. AMD is claiming a 52% higher performance per dollar metric for the card with a $3,499 MSRP, over NVIDIA’s Ada 5000 at $4,000 and Ada 6000 at $6,800, in benchmark apps like SPECviewperf. The card comes with DisplayPort 2.1 support for up to 12K60 output with compression, or 8K60 uncompressed. The cards can hit 8K120 and drive high-refresh rate and multi-display 4K configurations as well. The Radeon Pro W7900 Dual Slot card will be available on 6/19 this month.

All in, AMD hit every major compute platform in the company’s offering with fresh updates, at Computex 2024. I’d expect the month of July specifically to be lit up with deep dives and performance compares on these new platforms for mobile, desktop and the data center, with Zen 5 and AMD’s new 50 TOPS XDNA2 NPU taking center stage in the battle for AI processing supremacy. And at HotTech, we’ll be tracking where all the chips fall in terms of performance and value proposition.