AMD’s arrival 3Research and development The EPYC-generation processor family, which uses the new Zen 3 core, was highly anticipated. The promise of new microprocessor architecture, connectivity updates and new security options while preserving platform compatibility, is a good metric for an enterprise platform update, but One True Metric is platform performance. The vision of Zen 3 to lead per center in the consumer market again in November has raised expectations for a similar Slam Dunk tournament in the enterprise market, and today we can see these results.
AMD EPYC 7003: 64 cores Milano
The key figures that AMD is touting with the new generation of devices are an increase in raw performance yields of + 19%, due to improvements with the new core design. Moreover, AMD has new security features, improvements for various memory configurations, and updated performance with Infinity Fabric and connectivity.
Anyone looking for the shorthand specifications in the new EPYC 7003 Series, known by its codename Milan, will see great familiarity with the previous generation, but this time around AMD targets several different design points.
The Milan processors will offer up to 64 cores and 128 threads, using AMD’s latest Zen 3 cores. The processor was designed with eight chips of eight cores each, similar to Roma, but this time all eight cores were plugged into the smaller chip, enabling an efficient dual-buffer L3 cache design for lower cache latency overall. All processors will have 128 PCIe 4.0 lanes, eight memory channels, with most models support for dual processor connectivity, and new options are available to improve channel memory. All Milan processors must be compatible with the Rome Series platforms with firmware update.
|AMD EPYC: generation after generation|
|Exact architecture||Zine||Zain 2||Zain 3|
|Primary manufacturing||14 nm||7nm||7nm|
|Max cores / threads||32/64||64/128||64/128|
|The basic complex||4G + 8 MB||4G + 16 MB||8C + 32MB|
|Memory support||8 x DDR4-2666||8 x DDR4-3200||8 x DDR4-3200|
|Memory capacity||2 TB||4 TB||4 TB|
|PCIe||3.0 x 128||4.0 x 128||4.0 x 128|
|protection||Small and medium-sized companies|
|Small and medium-sized companies|
|Small and medium-sized companies|
|Maximum power||180 watts||240 watts *||280 watts|
|* Roma provided 280 watts for HPC’s special mid-cycle|
One of the highlights here is that the new generation of processors will offer 280 watt models for all customers – previous generations only had 240 watt models for everyone and then 280 watts for specific HPC customers, but this time all customers can enable this high performance parts with a new core design.
This becomes clear if we make direct comparisons with the top of the stack:
|2P Top of Stack GA Deals|
|uArch||Zine||Zain 2||Zain 3||Cascade|
|TDP||180 watts||240 watts||280 watts||205 watts|
|L3 cache||64 MB||256 MB||256 MB||37.5 MB|
|PCIe||3.0 x 128||4.0 x 128||4.0 x 128||3.0 x 48|
|DDR4||8 x 2666||8 x 3200||8 x 3200||6 x 2933|
|DRAM cover||2 TB||4 TB||4 TB||1 terabyte|
|price||4,200 dollars||$ 6,950||$ 7890||$ 3950|
AMD’s new top processor is EPYC 7763, a 64-core 280W TDP processor that delivers 2.45GHz base frequency and 3.50GHz boost frequency. AMD claims this processor offers +106% performance in the industry benchmark compared to its best-in-class 2P 28-core processor, Gold 6258R, and + 17% over its previous generation 280 W 7H12 Edition.
Peak performance versus baseline performance
One of AMD’s corners with the new generation of Milan will be targeted performance metrics, as the company not only seeks “peak” numbers, but also takes a broader view of customers who need high performance per center as well, especially for software. This is always a performance for every limited or licensed position. With that in mind, AMD’s “fast” series of processors are now stacked up.
|AMD EPYC 7003 F series processors|
(8 x 32)
|280 watts||4860 dollars|
|EPYC 74F3||24/48||3200||4000||240 watts||2900 dollars|
|EPYC 73F3||16/32||3500||4000||240 watts||$ 3521|
|EPYC 72F3||8/16||3700||4100||180 watts||$ 2468|
These processors have the highest correlated values for anything else in AMD offerings, along with a full 256MB of L3 cache, and in our results we get the best scores on a per-thread basis more than anything we have tested for enterprises across x86 and ARM – more detail in Review. The F series processors will come at a slight premium over the other processors.
AMD EPYC: Italy Tour
The first generation EPYC launched in June 2017. At the time, AMD was essentially a phoenix: breaking out of the ashes of Opteron’s earlier business, with a promise to return to high-performance computing with a new processor design philosophy.
At the time, the traditional enterprise customer base was not convinced at first – AMD’s recent foray into the enterprise space with a new generation of modular variable processor cores, while it was a hit, it had set back as AMD had to prevent itself from bankruptcy. Opteron customers were not left with any updates in sight at the time, and the willingness to jump onto an unknown platform from a company that had stung many in the past was not a positive for many.
At the time, AMD laid out a three-year roadmap, detailing the future generations and the path the company will take to conquer 99% of the giant market share in performance and offerings. These were seen as lofty goals, and many of them sat ready to watch others gamble.
With the launch of the first-generation Napoli, it delivered some impressive performance numbers. It didn’t quite compete across the board, and as with any new platform, there were some initial issues to get started. AMD has kept the initial cycle of a few of its major OEM partners, before slowly expanding the ecosystem. Naples was the first platform to offer condensed PCIe 3.0 and lots of memory support, and the platform initially targeted those storage or intensive PCIe deployments.
The second generation Roma, launched in August 2019 (+26 months), created a lot of fanfare. AMD’s latest Zen 2 core was a competitor in the consumer space, and there were a number of major design changes in the SoC layout (such as the move to flat NUMA design) that encouraged a number of skeptics to begin evaluating the platform. This was the concern AMD had told us that they had to be selective with the OEM platforms in which they would have helped before the official launch. Roma did well, scoring some high-profile wins in supercomputers, but more importantly, it may have shown AMD was able to implement that roadmap in June 2017.
The flatbed SoC architecture, combined with an updated Zen 2 processor core (which has already borrowed elements from Zen 3) and PCIe 4.0, allowed AMD to start competing over performance as well as IO, and AMD’s OEM partners were constantly announcing Roma processors as computing platforms. They often replace Intel 28-core processors for AMD 64-core processor which also has higher memory support and more PCIe offerings. This also allows for computing intensity, and AMD has been in a place where it can help drive software improvements for its platform as well, extract performance, but also move to parity in the edge cases for which its competitors are optimized. All major density meters have evaluated and published AMD-based offerings to their customers, as well as internally. AMD approval label has been around pretty much.
And so today AMD continues that tour in Italy with a trip to Milan, some 19 months after Rome. The basic SoC layout is the same as that of Rome, but we have higher performance on the table, with added security and more configuration options. Superballs have already had the end hardware for six months for their deployments, and AMD is now in a position to help enable more OEM platforms at launch. Milan matches Roma, which certainly helps, but with Milan covering more points of improvement, AMD believes it is better positioned to target more of the market with high-performance processors and high-performance processors per core, more than it has ever been before.
AMD sees the launch of Milan as a third step in its roadmap that was shown back in June 2017, verifying its ability to reliably deliver for its customers but also offering gains above industry standards for its customers.
Next stop on the Italy tour is Genoa, which is set to use AMD’s upcoming Zen 4 precision architecture. AMD has also said that the Zen 5 is in the pipeline.
AMD is launching this new generation of Milanese processors approximately 19 months after the launch of Roma. At the time, we saw the launch of both the Amazon Graviton2 and the Ampere Altra, built on Arm’s Neoverse N1 family of cores.
|Milan Top of Stack Competition|
|Healer||7763||Graviton 2||S 80-33||6258 p|
|TDP||280 watts||?||250 watts||205 watts|
|L3 cache||256 MB||32 MB||32 MB||37.5 MB|
|PCIe||4.0 x 128||?||4.0 x 128||3.0 x 48|
|DDR4||8 x 3200||8 x 3200||8 x 3200||6 x 2933|
|DRAM cover||4 TB||?||4 TB||1 terabyte|
|price||$ 7890||Unavailable||$ 4,050||$ 3950|
From Intel, the company has divided its efforts between large socket and small socket configurations. For large intakes (4+) there’s Cooper Lake, which is a derivative of Skylake for select customers only. For smaller (1–2) socket configurations, Intel is set to launch the 10nm Ice Lake Wallet sometime this year, but it remains muted on the specified dates. To that end, all we have to compare Milan to is the Intel Cascade Lake Xeon Scalable, which was the same platform we compared Rome to.
Have fun times for sure.
In this review, AMD gave us remote access to multiple identical servers with different processor configurations. We focused our efforts on the best-in-the-stack EPYC 7763 processor, a 280-watt 64-core processor, an EPYC 7713, a 225-watt 64-core processor, and an EPYC 7F53, a 280-watt 32-core processor designed for the Halo Milan processor designed to perform each core.
On the next page, we’ll walk through AMD’s Milan processor lineup, its comparison to Roma as well as the comparison to Intel’s current offerings. We then move on to our test systems, discussions about testing SoC architecture (cache, core-to-core, bandwidth), processor power and then to our full benchmarks.
- This page, overview
- Milan therapist offers
- Test bed settings, translator options
- Topology, memory subsystem and latency
- Processor power: Core vs IO
- Specifications: multi-threaded performance
- Specifications: Single thread performance
- The specs: Winning performance per center at 75F3
- SPECjbb MultiJVM: Java performance
- Compilation and calculation of criteria
- Conclusions and concluding observations
These pages can be accessed by clicking on the links, or by using the drop-down menu below.