Move Over Intel, AMD’s Got New Silicon Too!

There’s been a lot of talk about Intel’s Lunar Lake Core Ultra processors this week, but now it’s time to look at AMD’s brand new offerings aimed at both Intel and NVIDIA.  AMD is showing up Intel’s Granite Rapids Xeon chips in the server room, unveiling details on their fifth generation EPYC processors that go by the name of Turin.  The top of the line Epyc 9965 sports 192 cores capable of boosting up to 3.7 GHz, or the 64 core 9575F which can hit 5GHz if speed benefits your applications more than a high core count.  The higher core count Turin processors are using the 3nm Zen 5C architecture which focuses more on throughput than processing power.   The 4nm F series sacrifice insane core counts for higher frequencies, so AMD will have an EPYC for every situation.  The high core counts come with a cost, the Zen5c series devours between 320-500W of power, while the Zen5 tends towards somewhat lower power consumption … with some exceptions.

For CoPilot fans, the Ryzen AI PRO 300 series runs the gamut from the 12C/24T Ryzen AI 9 HX PRO 375, with 24 MB of L3 cache and a boost clock of up to 5.1 GHz to the 6C/12T Ryzen AI 7 HX PRO 360 with a mere 16 MB of L3 cache and boost of 5.0 GHz.  The Ryzen AI Pro series incorporates a dedicated NPU boasting 55 tera operations per second.  Overall they are very similar to the already launched Ryzen AI 300 series but the PRO series incorporates a variety of enterprise focused features, from vPRO to locally run AI-enhanced threat detection and data processing to avoid sending data to the cloud.   AMD claims that these chips will outperform Intel’s Meteor Lake chips by up to 40%, however keep in mind those chips are sub-28W while AMD’s Zen 5 based Ryzen AI Pro chips will run from 15-54W.  That may matter when it comes to performance per watt.

The other announcement concerns AMD’s competitor to NVIDIA’s AI Accelerators, with the new MI325X AI and the planned MI355X AI chips.  The MI325X AI is based on the new CDNA3 GPU architecture and sports 153 billion transistors and uses eight 32 GB stacks of HBM3e.  This architecture is slightly larger than the CPUs above, using a mix of TSMC’s 5 nm and 6 nm lines.. The MI325X AI has 19,456 stream processors and 1,216 matrix cores which are spread across 304 CUS.  The clocks will peak at 2.1 GHz and will provide up to 2.61 PFLOPs of eight-bit precision (FP8) performance 1.3 PFLOPs of half-precision FP16 operations.  That seems lower than what NVIDIA’s H200 provides, but AMD claims that their new chip outperforms it by up to 40% in certain benchmarks.

Still no work on that Ryzen 9 8900X3D yet though.