AMD have arrived at Computex 2024 with two major product announcements amidst a mixed year for the red team. Ryzen is unsurprisingly at the forefront, reaching its fifth major architectural revision while also introducing new capabilities to the ecosystem, but Radeon gets a small mention at the end so stick around for that.
Zen 5 & AM5: More performance, platform stability
After two years, a host of processors and quite a few motherboard chipsets, AMD are introducing the eventual replacement of Ryzen 7000-series CPUs and 600-series motherboards. It's time to say hello to the Zen 5 architecture, scheduled to appear in stores this July.
Zen 5 is another major evolution in the CPU core architecture introduced back in 2017. On the surface it might seem similar - core counts and cache reserves are largely the same as previous gen. counterparts - but under the hood there have been major improvements to branch prediction, wider pipelines and vectors, and a deeper window size to expand parallelism; all told, it's about a 2x the instruction and data bandwidth. AMD believe that this will account for an approximately 16% IPC uplift across the stack even as operating frequencies are improving.
This new architecture will be at the heart of AMD's Ryzen 9000-series, and will continue to leverage PCIe 5.0 and DDR 5 technologies well into the second half of the decade. The series will be headed by the Ryzen 9 9950X, a model with key specs similar to its 7000-series counterpart but boasting a higher 5.7GHz Boost clock. Out of the box AMD believe that it beats Intel's current best across a wide range of benchmarks including gaming.
The CPU is joined by three other models, but X3D designs don't currently feature on the product stack. Once again the Ryzen 9 9900X, Ryzen 7 9700X and Ryzen 5 9600X are similar to their 7000-series counterparts, but they're advertised as far more efficient designs. Now rated with 120W, 65W and 65W TDP despite modest clock speed improvements, they require less meaty cooling or can leverage PBO and greater thermal headroom for improved performance with dynamic frequency boosts.
Launching alongside the Ryzen 9000 Series will be X870 and X870E motherboards. These will continue to use the AM5 socket and be compatible with Ryzen 7000-series processors, but boast expanded capabilities including USB 4.0 as standard on all models, PCIe Gen 5 on Graphics and NVMe storage, and higher AMD EXPO memory clock support.
Furthermore, AMD have clarified that AM5 support will continue through to at least 2027. This echoes long-term AM4 support and offers consumers considerable confidence in their investment in Ryzen as a platform going forward.
Speaking of long-term support, AMD are also announcing two new Ryzen 5000 Series processors for AM4 platforms. The Ryzen 9 5900XT increases the core count of the 5900-class processor to 16 from 12; 8c/16t Ryzen 7 5800XT processors meanwhile offer improved max Boost operating frequencies to up to 4.8GHz.
Pricing for the new Ryzen 9000 Series processors, 5000 Series processors and X800-series motherboards wasn't shared at the time of writing.
Introducing Ryzen AI 300 Series Processors
Earlier this year Microsoft announced Copilot+, a new suite of AI-accelerated features that are set to be a core part of Windows beginning with previews later this year. The much maligned Recall - a means of recording user action history on a PC locally for later search and analysis - is just a part of a featureset that also includes real-time language translation and content creation tools, each of which require hardware capable of substantially accelerating AI workloads.
Desktop PCs and gaming laptops have discrete GPUs often capable of the 40 TOPS minimum necessary to meet Microsoft's requirements for a Copilot+ system; laptops, embedded systems and mobile hardware tend not to. Enter the NPU, an AI co-processor/accelerator developed by Microsoft and their hardware partners including AMD.
AMD previously had NPUs capable of 10 and later 16TOPS. Today they're announcing the Ryzen AI 300 Series codenamed 'Strix', capable of up to 50 TOPS alongside exceptional conventional CPU performance and graphics capabilities.
Ryzen AI 300 Series processors leverage up-to 12 Zen 5 CPU cores with SMT, an XDNA 2 NPU for up to 50 TOPS, and up-to 16 RDNA 3.5 Compute Units. They're launching in the top tier Ryzen 9 segment with two processors, the Ryzen AI 9HX 370 and Ryzen AI 9 365.
This new product lineup also ushers some new branding specific to Ryzen AI processors. It leapfrogs the first two generations and goes straight to the 3rd gen, hence 300-series; meanwhile it's notable that 'HX' branding now simply defines the processor's position at the top of the Series product stack.
AMD strongly tout the potential of their NPU, a 5x performance and 2x power efficiency improvement over the prior generation. That puts it at the top of the NPU 8-bit TOPS performance charts, higher than the anticipated maximum for Intel's upcoming Lunar Lake architecture. More significantly, AMD's XDNA 2 architecture incorporates a Block FP16 mode on the NPU, allowing it to accelerate 16-bit workloads at full 8-bit performance levels without quantization.
While AI will be getting plenty of attention, AMD are also boasting of 'console-like' gaming performance with these APU-power laptops. While that might seem a stretch - gaming optimisations are hard after all - a sport of legacy gaming on titles a few years old could easily be on the cards.
AMD plan to have 150+ AI development software partners by the end of 2024, and 100+ platforms for Ryzen AI from partners including ASUS, Lenovo and MSI (who will announce supported laptops at Computex 2024) by the end of the year.
And finally, a word from Radeon
Despite AI's shadow looming over the convention, AMD's Radeon GPU hardly got a mention at Computex 2024. The red team were proud to reveal a new PRO workstation GPU, the Radeon PRO W7900 Dual Slot, but it was curious to see lack of RDNA 4 announcements given how important the potential for discrete AI acceleration is to the gaming industry at this time.
The Radeon PRO W7900 Dual Slot meanwhile is an interesting model that seems tailor-made for LLM inferencing on models with large memory footprints. It boasts a 48GB frame buffer, 16GB more than their competitors RTX Ada 5000, and rests at a substantially lower price point. No prizes therefore for guessing that AMD highlight its price/perf credentials.
As a Dual Slot certified solution AMD are also keen to point it its compatibility with 4-card configurations within standard workstation chassis, helping organisations assemble scalable multi-GPU AI desktop solutions.
Finally, AMD are adding TensorFlow to ROCm with ROCm 6.1 for Radeon GPUs, expanding the list of supported AI frameworks. ROCm will also receive beta support on WSL 2 for the first time, opening up the environment to more developers.