AMD adds 128-core Bergamo and 3D V-Cache Genoa CPUs to Zen 4 Epyc lineup

Ever since its launch in 2017, AMD's Epyc server CPUs have offered some of the highest core counts for high-performance server chips. Epyc has gone from 32 cores with the original Zen, to 64 cores with Zen 2, and 96 cores with Zen 4, but Epyc is hitting 128 cores today as AMD launches its Zen 4c architecture, which is designed to deliver better core density and power efficiency than Zen 4. Additionally, AMD is also releasing 96-core Epyc chips with 3D V-Cache, the first CPUs to offer 1GB of L3 cache.

Zen 4c is smaller and more efficient than Zen 4, but cuts back on per-core performance

Announced about a year ago, Zen 4c was billed as the ideal cloud server CPU architecture thanks to offering 35% smaller cores and better power efficiency compared to Zen 4. Codenamed Bergamo, Zen 4c Epyc CPUs come with up to 128 cores spread over eight core complex die (or CCD) chiplets with 16 cores each, while Genoa, AMD's regular Zen 4 server CPU, has 12 CCDs with eight cores each. However, both Bergamo and Genoa share the same I/O die (or IOD), which contains hardware like memory controllers.

However, Zen 4c didn't get all of this for free. For starters, Zen 4c will have a significantly reduced boost frequency compared to Genoa, which will mean lower single-threaded performance. Additionally, the total L3 cache is lower on Bergamo because AMD only puts 32MB on each CCD, and since Bergamo has fewer CCDs than Genoa, that means less cache. Bergamo does have more L2 cache than Genoa, but that's tied to core count, and it probably won't offset the lower amount of L3 cache.

3D V-Cache comes to 4th Generation Epyc in Genoa-X

Although 3D V-Cache has been most visible with gaming CPUs like the Ryzen 7 7800X3D, it actually first debuted with 3rd Generation Epyc Milan-X CPUs, which used eight V-Cache chiplets to add 512MB of L3 cache to Milan CPUs. Using the same 3D V-Cache technology as Ryzen 7000X3D CPUs, Genoa-X will offer over 1GB of L3 cache, the first CPU to ever do so.

The most obvious competitor to Genoa-X is Intel's Sapphire Rapids Xeon Max CPU, which uses HBM2 to offer up to 64GB of L4 cache. That puts AMD in second place when it comes to capacity, but Genoa-X also offers 96 cores while Sapphire Rapids only goes up to 56, and that HBM2-powered L4 cache will undoubtedly offer worse latency than Genoa-X's more traditional L3 cache. Both of these CPUs, however, are quite niche and won't be used nearly as much as their lower-capacity counterparts.

Radeon Instinct MI300X is made for AI and large language models

The final product AMD showed off was its brand-new MI300X server GPU, which is a variant of the MI300A APU that replaces the three Zen 4 chiplets with three CDNA 3 chiplets and adds 64GB of HBM3 for a total of 192GB. The MI300X will compete with Nvidia's H100 and Intel's Falcon Shores, which was also supposed to offer an APU/XPU option like the MI300A but has since been canceled.

MI300X begins sampling in the third quarter of this year and production will ramp in the fourth quarter, which means MI300X will likely launch in early to mid 2024.