Rechercher dans ce blog

Thursday, March 18, 2021

AMD Breaks Server Records With Its Third-Generation Epyc Processor - Forbes

proc.indah.link

Since the introduction of the first Epyc server processor with the original Zen CPU core in 2017, AMD has made steady progress in performance, capability, and market share. With the release of the third-generation Epyc based on the third generation Zen core, AMD has kicked up the competition a notch by offering twice the performance of Intel’s existing Cascade Lake Xeon processors based on various enterprise, cloud, and high-performance computing (HPC) workloads and benchmarks. The application performance improvements in the third generation Epyc are due to a 19% average increase in the instructions executed per clock cycle (referred to as IPC) of the Zen 3 CPU core over its predecessor (based on the mean average of 28 performance measures). AMD calls this new processor family based on the Zen 3 CPU cores Epyc 7003.

It’s hard to image that just over 4 years ago, AMD had only a sliver of the server market and was not considered competitive on performance with Intel. Today, thanks to the Zen CPU architecture and the multi-die (chiplet) packaging strategy, AMD is the performance leader, gaining market share and mind share.

AMD Epyc – the new server performance platform

What we are starting to see is that AMD’s Epyc is becoming a preferred platform for HPC and AI platforms. Last year Nvidia upgraded its DGX AI server product to the second-generation Epyc processors from Intel Xeon processors. Epyc offers more PCIe lanes than Intel’s Xeon, allowing a higher GPU to CPU ratio, which preferable for HPC workloads that rely on GPUs for number crunching. Even with AMD’s second generation Epyc, AMD already had more cores per socket than Intel and was using the latest PCIe technology - PCIe gen 4. As a result, the Epyc platform has much higher chip-to-chip performance.

While the Zen 3 core is manufactured in a roughly similar 7nm process as Zen 2, AMD made architectural changes to the CPU cores, the core complex, and the L3 cache to achieve that 19% performance boost. In particular, the changes to the L3 improve performance on large data workloads. In the Epyc 7003 processors, all eight cores in the CPU complex have direct access to the large 32MB L3, compared with only four cores sharing 16MB L3 caches in Epyc 7002.

MORE FOR YOU

With more cores and improved core performance, AMD claims the Epyc 7003 is roughly twice the performance of Intel’s current generation of Xeon processors. AMD demonstrated benchmarks on enterprise, cloud, and HPC workloads. AMD compared its 64-core processors against the Intel Xeon Gold 6258R with 28 cores (formerly known as Cascade Lake). AMD is ahead on cores per socket using Epyc’s chiplet architecture and the use of TSMC’s 7nm process for the CPU chiplets. Intel’s present Xeon processor is a monolithic die made in the company’s 14nm process technology, which limits the number of CPU cores Intel can put on a single die.

The benchmark comparisons included the Epyc 7763, running the SPEC floating point benchmark, which was 106% faster on HPC workloads – an important market for solving critical business, engineering, and scientific problems. AMD put special emphasis on HPC in the product launch due to its leadership performance.

AMD’s commitment to HPC was evident throughout the presentation and the ultimate performance computer will be the Frontier supercomputer at Oak Ridge National Labs which will use this third generation Epyc processor combined with next-generation AMD Instinct compute processors to deliver 1.5 exaflops of performance. Microsoft Azure and Google Cloud are both offering Epyc instances focused on HPC workloads, as well as HPE Apollo and Cray product lines for HPC.

Lenovo alone set 20 new world records with the Epyc 7003-based SR995, including setting a new world record for virtual machine performance. With so many performance virtual machines, Epyc does well in hyperconverged infrastructure. Epyc is also well suited in other enterprise workloads with a strong performance lead in relational database performance benchmarks and data analytics benchmarks.

n addition to higher performance, AMD improved security with new Secure Encrypted Virtualization-Secure Nested Paging (SEV-SNP) function – which provides memory protection as it can’t be manipulated by a guest hypervisor. AMD calls it AMD Infinity Guard. Secure encrypted virtualization makes the Epyc 7003 perfect for confidential computing, an initiative by cloud vendors that eliminates the data security vulnerability by protecting data in use - that is, during processing or runtime. Microsoft Azure will offer Epyc instances with confidential computing as well as using data encoded in transit and at rest. A secure processor on the Epyc I/O die will also perform hardware validation at boot and monitors runtime execution.

Even with all the improvements in Epyc 7003, it still can drop into the same platforms as the second generation EPYC 7002, with a BIOS upgrade. This is a huge plus for OEMs that already have system designed to use Epyc 7002, including Nvidia.

Partners

Epyc processors have become the top choice for many high-performance workloads and has attracted key systems partners. At the launch, Lenovo, HPE, Gigabyte, and Dell EMC all spoke. Cloud service partners representatives speaking at the launch included those from Google Cloud, Microsoft Azure, TenCent Cloud. The 1st and 2nd generation Epyc have over 200 cloud instances available. AMD expects that with the 3rd generation it will have over 400 Epyc cloud instances by the end of 2021.

With the third-generation Epyc, other partners such as Gigabyte and Qualcomm are using it as a platform for a Peta-OP AI server rack. The technical advantages of AMD’s Epyc processors makes it a great platform for accelerated computing.

Selection

AMD has also been filling up its product line card. There is now a total of 19 versions of the Epyc 7003 processors ranging from eight cores, up to 64 cores, with many choices of core counts and performance options in between. AMD will continue offering a few second-generation parts that have a specific price/power/performance/core count characteristic, even though it lacks the IPC improvements. Most of the processors support dual-socket configurations, but AMD also offers single socket parts in 16, 24, 32, and 64 cores.

Summary

AMD is going all out to maintain performance leadership against Intel in the server market. While AMD launched Epyc 7003 against the 14nm Cascade Lakes and not the forthcoming 10nm Ice Lake Xeons, AMD still holds a solid lead in performance. Intel will close the gap later this year when Ice Lake ships, but for now, AMD has a tremendous advantage on performance. AMD’s 4th generation Epyc processor in 5nm process should ship in 2022. Five years ago, Intel had a virtual monopoly on server processors, but now we have a real competitive market. And everyone wins when there’s competition.

The Link Lonk


March 18, 2021 at 01:53PM
https://ift.tt/3cQJePw

AMD Breaks Server Records With Its Third-Generation Epyc Processor - Forbes

https://ift.tt/2ZDueh5
AMD

No comments:

Post a Comment

Featured Post

Intel Falls on Latest Server Chip Delay; Rival AMD Gains - Yahoo Finance

proc.indah.link (Bloomberg) -- Intel Corp. fell after saying a new version of its Xeon server chip line will go into production in 2022, r...

Popular Posts