Intel 5th Gen Xeon “Emerald Rapids” CPUs Detailed: Up To 40% Faster Than Sapphire Rapids, Die-Shot Pictured

Intel 5th Gen Xeon "Emerald Rapids" CPUs Detailed: Up To 40% Faster Than Sapphire Rapids, First Die-Shot Pictured 1

A brand new Intel 5th Gen Xeon "Emerald Rapids" slide deck has emerged which gives us a bit more information on the chip layout and performance of the next-gen data center CPUs.

The slides come from an Intel "Data-Centric Processor Roadmap" presentation which was spotted by @InstLatX64. The deck looks very unfinished since it's missing several key data points but we do get to learn some additional information on the 5th Gen Xeon lineup which is codenamed Emerald Rapids.

So starting with the details, we first have our first close-up die shot of the 5th Gen Intel Xeon CPU which details its various architectural blocks. The Emerald Rapids CPUs are based on the P-Core architecture under the Raptor Cove ISA. The flagship chip, Xeon 8592+, incorporates a total of 64 cores, 128 threads, and a massive 480 MB L3 cache pool. As you can see, Intel has gone from a 4-tile design in 4th Gen Sapphire Rapids to a 2-tile design in 5th Gen Emerald Rapids CPUs.

Each chiplet is connected through a modular die fabric which sits between the two core and cache arrays. Each array features a total of 35 cores of which 3 cores are disabled per die. Each tile has 2 memory controllers supporting DDR5 DIMMs with up to 5600 MT/s speeds, three PCIe controllers (6 total), 2 UPI (4 total), and two accelerator engines (4 total). According to Intel, the Intel Emerald Rapids CPUs will offer:

Increased Performance & Performance/Watt

Increased Core Counts with larger shared L3 cache

Compute Express Link (CXL) Type 1,2, & 3

Increased Intel Ultra Path Interconnect Speeds (UPI)

Up To 80 PCIe Gen5 Lanes

Intel Accelerator Engines

Increased Memory Speeds

Some of the features to expect within the 5th Gen Xeon "Emerald Rapids" CPUs include:

Up To 3x larger LLC cache

Increased Memory Speed

Higher CPU Core Count (Up To 64 Cores)

CXL Type 3 high bandwidth interface

Workload optimizations (Intel AE)

Optimized Power Mode

Up To 17% General Purpose Perf/Wat Increase

Offload CPU cores with Intel Accelerator Engines for Power Efficiency

Huge performance boost across inference and training

Intel AMX for built-in AI Acceleration

Out-of-box deployed with Optimized SW stacks

The Accelerator Engines embedded within the 5th Gen Xeon "Emerald Rapids" CPUs offer the latest AMX (Advanced Matrix Extensions) which are backed by a solid suite of AI software that utilizes optimized open-source frameworks and tools. Intel also claims that the next-gen Xeon chips will offer increased TCO & Perf/$ gains over 4th Gen Xeon while offering customers a drop-in upgrade path since the CPUs share the same socket and platform as the 4th Gen SKUs.

That brings us to the next topic which is performance. In all benchmarks, Intel compares the 5th Gen Xeon Platinum 8592+ (Emerald Rapids) 64 Core CPU against the 4th Gen Xeon Platinum 8480+ (Sapphire Rapids) 56 Core CPU. The performance increments in both Data Center specific and AI tasks are mentioned below:

5th Gen Xeon 8592+ vs 4th Gen Xeon 8480+ Benchmarks:

Web (Server-side Java Throughput) = +20% Gain

HPC (LAMMPS-Copper) = +30% Gain

Media (Transcode FFMPEG FPS) = +20% Gain

AI Natural Langauge Processing = +40% Gain

AI Recommendation Systems = +40% Gain

Now do keep in mind that these are official benchmarks and one should wait for 3rd party results for a more clear picture. While the gains are decent for a refresh, it remains to be seen if these will be enough to tackle AMD's EPYC 9004/8004 lineup based on the Zen 4 and Zen 4C architecture.

The AMD EPYC family has been smashing records in terms of performance and efficiency and we are just a day away from AMD's Advancing AI event where the company will further expand its server, HPC, and AI portfolio with next-gen accelerators. Meanwhile, Intel's 5th Gen Emerald Rapids CPUs are expected to launch on the 14th of December.

News Sources: InstLatX64, Momomo_US