AMD EPYC 7002 Server Processors

AMD made big strides when it launched the Epyc range of server processors in June 2017. Having had practically no presence in the server space since the demise of the Opteron range of chips more than 10 years previously, Epyc, built on the all-new Zen architecture, offered up to 32 cores and 64 threads per socket, class-leading memory bandwidth, and a fully-featured expansion ecosystem providing 128 PCI Gen 3 lanes.

Evaluated over a series of industry-standard server benchmarks, 1st Gen Epyc enjoyed a compelling performance and total cost of ownership advantage over incumbent Intel Xeon processors. In the intervening two years, however, Intel, which still commands a significant 95 per cent of total x86 server CPU sales, has launched its own series of infrastructure upgrades to CPUs, memory support, SSDs, and networking.

AMD's server plans were not just limited to the original Epyc, codenamed Naples. At its inception, the company detailed follow-up architectures with codenames of Rome (Epyc 2nd Gen) and Milan (Epyc 3rd Gen), to arrive in 2019 and 2021, respectively. Importantly for server longevity and ease of upgrade, subsequent Epyc chips would also use the original's SP3 socket form factor.

Epyc 7002 Series processors debuted in August 2019. They are drop-in upgrades for installations using the first-gen motherboards, and they arrive to market armed with a number of impressive improvements. Two key attributes enable better performance that, in certain cases, is up to 4x higher than Epyc 7001 Series - 7nm production and the Zen 2 architecture.

Epyc 7002 Series' Weapons: 7nm

AMD moves the 14nm process production down to 7nm in the transition between Epyc CPU series. This is a big deal because halving the transistor size enables far greater density, lower power, or a combination of both, and AMD uses these characteristics to good effect by doubling the number of cores and threads, from 32/64 to 64/128, for the range-topping chips in each Epyc generation. Think about that for a second - AMD has (at least) doubled the potential of each chip by smart engineering and close collaboration with foundry partner TSMC who builds the cores.

As before, AMD adopts a multi-chip design approach when building out high-core Epyc chips. Each die houses eight execution cores and 16 threads, so getting to that 64C128T number requires eight dies. Enabling such a chip to be built at an economically palatable level shows the indirect power of moving to 7nm. Rival Intel, meanwhile, has just started mass production of its 10nm process, meaning that AMD has a process advantage that will stay a while.

Let's be clear. If AMD had not moved to a 7nm process, a 64C128T CPU would have been burdened with an untenable power budget incompatible with existing server infrastructure. Epyc 7002 series only works in its present form because of the advantages inherent in 7nm production.

In a clever move designed to keep costs in check and improve the chip's internal bandwidth and performance, AMD separates the execution cores from the supporting I/O logic. All core-housing dies feed into a central I/O die that accommodates the six-channel memory controller, 128 PCIe lanes, and all other attendant hardware. Rather than use costly, leading-edge 7nm wafers for this purpose, Epyc's I/O die is built on the more mature, affordable 14nm process. In other words, AMD puts the right silicon in the right areas.

Epyc 7002 Series' Weapons: Zen 2 architecture

The above commentary illustrates how AMD got to 64 cores and 128 threads in a single CPU package. But that wasn't enough, because the server team wanted to improve each core's performance in a wide range of present and evolving server workloads. Enter Zen 2.

The latest AMD CPU architecture, also present on the desktop Ryzen 3rd Gen chips, improves the all-important instructions-per-clock (IPC) metric by approximately 15 per cent compared to Zen. How? Zen 2 employs an improved branch predictor, beefs up the caches, particularly the L3, which is doubled from 8MB to 16MB per 8-core complex.

There's also an improved central core, enhanced by having a third address generation unit, faster connections between memory, and more bandwidth into the chip via six-channel DDR4-2933 instead of DDR4-2666. Last but not least, AMD upgrades the chipset connectivity from PCIe 3.0 to PCIe 4.0, which doubles bandwidth for high-speed devices such as graphics cards and networking adapters. The total number of lanes is kept at 128, however, which is certainly more than sufficient for a server setup.

Helping out in floating-point-intensive tasks, AMD also increases the AVX vector width from 128 bits to 256 bits. This change, again, doubles performance in highly-vectorised workloads that are becoming more common in the server space. Doubling the cores and threads and doubling the AVX capability provides up to 4x the amount of floating-point oomph as the first generation.

Also, like first-gen Epyc, two chips can be mated together in what is known as a 2P setup. The two processors are connected by 64 PCIe lanes each - leaving 64 for each processor, ostensibly for expansion components such as graphics, NVMe SSDs, and 100Gbit networking.

EYPC 7002 Models

Product Stack Analysis

AMD has launched 19 Epyc 7002 Series CPUs, complementing the 14 drop-in Epyc 7001 Series and eight Epyc 3000 Series embedded chips already in the market.

All of these processors will run in a single-socket system, but those designated with a P suffix cannot run alongside a second chip on a 2P motherboard. This is why they're cheaper, but otherwise identical, than their two-way counterparts.

AMD addresses the entire stack of the server/workstation market by scaling from eight cores through to 64. As each die is home to eight cores and 32MB of cache, AMD tries to use fully-intact dies rather than have cores deactivated per complex.

There's still plenty of choice as one navigates up the pack, and as before, there are multiple Epyc processors harnessing identical core-and-thread counts, albeit differentiated on frequency and TDP. Speaking to power, the TDPs do rise as more cores and threads are added, as expected. The range-topping Epyc 7742, for example, consumes 225W, compared to 180W for the first-gen Epyc 7601.

Further maximising implementation potential within a wide range of servers, each chip has a configurable TDP that modulates frequency according to power. The top-line 225W TDP parts are not designed to run lower than specification, unlike the others, and there's an extra 15W TDP on tap for those servers with excellent cooling and airflow.

Summary

AMD has made significant strides in improving its already-decent line of Epyc server processors. The second generation, codenamed Rome, offers up to twice the cores, up to 4x the floating-point performance, and an extra 15 per cent general-purpose oomph at the same frequencies as the original Epyc.

Making excellent strides in every area, AMD is now the performance leader in the x86 CPU server space. Epyc 2 ought to increase the firm's market-share significantly in 2020.

As always, Scan Computers retails a wide selection of AMD Epyc 7002 Series processors