AMD new chip | Techspot.com | 3rd-gen Epyc server CPUs

 Nov. 10, 2021

Here is the article.

AMD intros 3rd-gen Epyc server CPUs using 3D V-Cache and up to 768MB of L3 cache

AMD's Milan-X CPUs are drop-in compatible with a BIOS update

By Shawn Knight November 8, 2021, 4:53 PM 

In a nutshell: Codenamed Milan-X, these new chips will feature the same capabilities and features as existing third-gen Epyc processors, albeit with three times the amount of L3 cache. At the top tier, this works out to 768MB of L3 cache and up to 804MB of total cache per socket.

AMD at Computex earlier this year teased next-gen 3D chiplet technology it developed in collaboration with Taiwan Semiconductor Manufacturing Company. During the company’s virtual Accelerated Data Center Premiere event on Monday, AMD announced it would be bringing this new technology to the data center.

When architecting Milan-X, AMD studied how technical computing applications behaved, and found that a large amount of cache was key to obtaining higher performance. This keeps data closer to the cores, thus reducing overall system latency.

Su said Milan-X chips are the fastest server processors for technical computing workloads, offering more than a 50 percent uplift compared to standard Milan processors. What’s more, they support Socket SP3, will be drop-in compatible with existing boards with a simple BIOS update, and take advantage of today’s software with no changes required.

Third-gen AMD Epyc CPUs with 3D V-Cache are scheduled to launch in the first quarter of 2022. Cisco, Dell, Lenovo, HPE and Supermicro have already partnered with AMD and will be offering server solutions utilizing these processors when the time comes.

If you can’t wait that long, Microsoft Azure HPC has virtual machines featuring the new chips available from today in private preview with a broader rollout slated for the coming weeks.

Share value in AMD is up more than 10 percent on the day.

 VitalyT

Is it not more than a marketing trick, shoving absurd amounts of cache into a CPU? I understand that these are the most expensive CPU-s that AMD produces, but that cache size is used by SSD-s, not CPU-s. What's the real benefit here, if any?

• REPLY

November 8, 2021, 5:44 PMTom Yum

VitalyT said

Is it not more than a marketing trick, shoving absurd amounts of cache into a CPU? I understand that these are the most expensive CPU-s that AMD produces, but that cache size is used by SSD-s, not CPU-s. What's the real benefit here, if any?

Well this is aimed at servers, but the answer is memory latency and bandwidth. L3 cache latency is around 9-10ns, whereas the best SSD's are around 30-50 us, ~3000-5000 times slower. DDR4 latency is around 75ns, or ~8 times slower. A modern processor can execute hundreds of instruction in the time it takes for data to become available from the memory, or ~100k instructions in the time taken to get data off the fastest SSD's. For certain workloads being able to avoid those long access times by keeping data closer to the processor will provide significant performance uplifts.