What an astounding twelve months it has been.
In the run-up to the anniversary of the EPYC launch, I’ve had the opportunity to talk to several executives at AMD to not only get a refresher of what the year had to offer and what the customer response to this new silicon has been, but also to get a few tantalizing glimpses at what is to come in the future.
The company kicked off the EPYC line with nine 7000-series processors. These chips are built using 14-nanometer architecture and based on the Zen architecture, and consist of four eight-core Zeppelin dies baked into a single multi-chip module.
AMD came out of the gate strong with the EPYC 7000-series, with a broad offering of silicon ranging from 8-core/16-thread parts all the way up to 32-core/64-thread parts (cores are disabled on the Zeppelin dies to achieve desired core count), with speeds spanning 2.7GHz to 3.2GHz, and prices going from $400 to over $2,000 per chip.
With the new EPYC platform, AMD became laser-focused on delivering four things that it believed was lacking in the server market — flexible configurations, an open ecosystem, building platforms optimized for modern workflows, and lowering total cost of ownership.
The EPYC 7000-series processors features list was incredibly disruptive:
- Up to 32 high-performance “Zen” cores
- Single- and dual-socket support
- Eight DDR4 channels per CPU
- Up to 2TB RAM per CPU
- Infinity Fabric for both die-to-die and socket-to-socket interconnect
- 128 PCIe lanes
- Dedicated security subsystem, including Secure Memory Encryption (SME) specifically designed to protect virtual machines and containers from each other
- Integrated chipset
- Socket-compatible with next-gen EPYC processors
The fact that each EPYC processor package can support up to 2TB of DDR4 RAM over eight channels, and has 128 PCIe lanes means that the platform has more than two-and-a-half times the I/O density of a processor such as Intel’s Xeon SP Series.
Finally, a true single-socket server processor
While there’s an awful lot to get excited about with EPYC, one feature that’s turned enterprise customers’ heads is EPYC’s single-socket capability. And not just single-socket, but what AMD called “the industry’s first no-compromise one-socket solutions.”
Prior to EPYC, if you were looking for a fully-featured, high-performance server in a single-socket configuration, you were out of luck.
And this was no small shift. This was a big deal.
A huge deal.
What it meant was that enterprise customers would no longer be funneled into buying unnecessary two-socket servers because of arbitrary and artificial limitations on memory bandwidth and I/O as they are when they buy into Intel server silicon, and instead would be free to choose the number of sockets that suited their workloads and requirements, allowing customers to make a no-compromise choice for underutilized servers, and benefit from lower power consumption and lower capital expenditure.
If your workload only needed a single-socket solution, enterprise customers no longer had to throw a redundant second socket into the system to get the I/O and bandwidth they needed.
The EPYC advantage
Earlier this year I spoke to Raghu Nambiar, AMD’s CVP and CTO of Datacenter Ecosystems and Application Engineering, and Dan Bounds, AMD’s Senior Director for Data Center Products on the subject of SDS — Software Defined Storage — and how EPYC is well placed to lower TOC in ways previously not possible.
Take, for example, how storage and memory are converging in the enterprise space, and EPYC is well placed to take advantage of this. EPYC, with its eight memory channels and 128-lanes of PCI, means that EPYC-powered servers can handle larger memory footprints than those powered by Intel silicon, and there’s also no need to separate PCI switches to be used to take advantage of NVMe storage.
All this allows for greater memory-to-core and cores-to-spindle/drive ratio in the server room.
And choosing to go with a single-socket solution for certain workloads can mean savings of thousands of dollars. For example, VMware vSphere licensing is based on the number of CPUs in a system, and there are special bundles for single-socket systems, where for a one-node system, a single-socket server can deliver $7,610 in savings on VMware licensing costs for vSAN and vSphere, while for a three-node cluster, these savings rise to $22,830.
And these savings are simply a result of switching from dual-socket to single-socket. That’s it. All the savings that come from reduced power consumption and lower cooling costs are separate.
When it comes to virtualization and cloud, EPYC can deliver performance that’s similar or better than Intel-branded silicon along with a 28 percent lower TCO.
Another area where AMD believes that EPYC is changing the server landscape is its support for memory encryption. And not just that, but it offers this without sacrificing performance or requiring that software be recompiled to support the feature. EPYC handles it all in the background, without needing any code changes.
Across several interviews, AMD executives have made it clear that security is of paramount importance for customers, especially in areas such as edge computing and multi-tenancy virtualization. These customers don’t just need it, they require it, and the regulations that spin off as a result of laws such as GDPR could someday demand memory encryption.
EPYC is no slouch when it comes to performance either. Whether you’re spending $400 or $4,000, EPYC is wiping the floor with the Intel Xeon chips, even when factoring in the new Intel Skylake chips on the two-socket front, where the performance advantage ranges from 15 percent for >$4,000 price point (and that gap, while it may seem marginal, is far from it), to a whopping 56 percent at the $500 to $800 price range.
This is an enormous performance gulf.
What really strikes me about EPYC is how rapidly it gained traction in a market that’s usually stifled by inertia. Server customers aren’t the type to adopt something new just for the sake of having something new, so the pace at which EPYC silicon has permeated the market is a clear indicator of the pent-up demand for change — and serious competition to what Intel was offering. The long list of mammoth wins for EPYC range from cloud customers such as Microsoft, Baidu and Packet, to hardware OEMs such as Dell EMC, HPE, Cray, and Cisco.
AMD didn’t sit on its laurels during 2018 though.
In February it expanded the EPYC line by launching eight 3000-series embedded processors, ranging from low-power 4-core/4-thread parts with a TDP of 35 Watts, to monster 16-core/32-thread 100 Watt TDP aimed at some serious heavy data lifting.
And the specs of this silicon are equally impressive:
BGA package (more suited to industries that demand higher reliability, such as medical, transportation, and military)
- Up to 16 “Zen” x86 cores
- Enterprise-grade reliability, availability, and serviceability (RAS)
- Up to 32 megabytes of shared L3 cache
- Maximum boost frequency of 3.1GHz
- Up to four memory channels
- Up to 1TB of memory capacity
- Up to 64 PCIe Gen 3 lanes
- Up to eight 10Gb Ethernet
- Up to 16 SATA or NVMe channels
- Product availability for 10 years
The complete line up was very strong out of the gate, and hit all the wave crests of price and performance points that AMD needed to hit in order to continue to pile the pressure on Intel.
Even when pitted against Intel’s newer Skylake chips, when embedded EPYC parts are compared to Intel hardware — such as the EPYC Embedded 3451 vs. Xeon D-2191 ($800 vs. $2,407) or the EPYC Embedded 3301 vs. Xeon D-2152IT ($450 vs. $438) — the 3000-series EPYC silicon consistently offered more; more in the form of cores, cache, memory capacity, memory speed, PCIe lanes, and more 10GB Ethernet, all for either the same amount of money, or a lot less.
AMD continued with the drumbeat that had become a familiar one throughout the year: More scale. More expansion. More memory bandwidth.
Close partnership between AMD and HPE
One of AMD’s closest server partners is — and always has been — HPE. The two companies work closely together to deliver products to market for a wide array of enterprise workloads. And the latest products to come from this close association are HPE’s new EPYC 7000-series powered dual-socket ProLiant DL385 Gen10 and the upcoming single-socket ProLiant DL325 Gen10 servers.
The ProLiant DL385 Gen10 has been built from the ground up for virtualized and memory-centric workloads, and can offer up to 64 cores, up to 32 DIMMs/4TB memory capacity, along with support for up to 24 NVMe drives, and takes advantage of EPYC’s secure memory encryption (SME) and secure encrypted virtualization (SEV) to deliver critical security functionality such as Secure Root of Trust, Secure Run, and Secure Move.
A customer that’s already making use of ProLiant DL385 Gen10 is the Center for Research Computing (CRC) at the University of Notre Dame. The CRC has several projects on-going that need heavy-duty high-performance computing (HPC), such as the Department of Energy’s Center for Shock Wave-processing of Advanced Reactive Materials (C-SWARM), the Computational Hydraulics Laboratory (CHL), and Terrestrial Ecosystems and Climate Modeling projects.
“The AMD and HPE systems gave us performance that allows larger analytics,” said Paul Brenner, Associate Director of the Notre Dame Center for Research Computing. “A lot of these are memory-bound applications, fluid dynamics specifically, so [the systems] gave us a nice architecture upon which we could perform larger memory, high-performance calculations with good performance between the cores and the memory itself.”
And the ProLiant 385 Gen10 servers have proved themselves to be highly competitive.
“We felt that the quantifiable performance to price ratio was outstanding and that’s why we selected HPE and AMD,” said Brenner. “And that has proved to be true across multiple applications already and in a month or so of production utilization…the ultimate metric for us is being able to have those scientific outputs for all of those different models, faster.”
One of the challenges facing AMD with its re-entry into the server market was making sure that it didn’t repeat the mistakes it made in the past. After speaking to AMD’s Forrest Norrod, SVP and general manager of the Enterprise, Embedded, and Semi-Custom (“EESC”) group, it’s clear that AMD both knows where it went wrong with the server market in the past, and what it takes to win back trust and execute EPYC development in a way that is sustainable.
According to Norrod, AMD is also working on building a roadmap that allows for predictable execution, going so far as to have overlapping design teams working on the development of four generations of architectures, from next-generation 7-nanometer Zen2 (codenamed Rome) out to Zen5.
The company is also committed to its “no compromise” single- and dual-socket server systems, systems with no arbitrary limitations set on I/O, or memory bandwidth and performance as it sees this as the way to allow customers to make a no-compromise choice for underutilized servers, and benefit from lower power consumption and lower capital expenditure.
And just because there’s been so much attention placed on the single-socket market doesn’t mean that AMD is taking its eye off the dual-socket market in the least. Indeed, the message from AMD has been consistent in pointing out that there’s no shortage of workloads — and customer demand — for two-socket servers, and that this will continue to be the case for the foreseeable future.
But what about four-socket EPYC servers and above? Is this an area where AMD wants to expand EPYC into?
When I’ve asked about this, the answer has been a “no.” Not only does this sort of scaling increase the failure surface the more silicon is packed into a server, but there’s also a knock-on effect in terms of increased qualification and validation time for hardware.
AMD is laser-focused on single- and dual-socket, and it seems happy to leave whatever market exists beyond that to the competition.
So, what does the future hold for server customers?
Well, as you can imagine, as open and candid as AMD executives are when it comes to talking about things out in the open, it’s understandable that conversations become more reserved when talking about what the future holds. However, I did get plenty of pointers as to where the market is headed and what we should be keeping an eye on:
- Multi-chip modules will bring greater flexibility, especially when it comes to storage class memory
- Shift from LRDIMM memory to RDIMM, which will bring with it higher-density, better performance, and lower power consumption (Samsung said that soon it will begin sampling 16Gb-based 256GB RDIMMs, which would expand the memory capacity for a 2P server to as much as 8TB)
- More cores with the shift to 7-nanometer architecture and beyond — The next generation 7nm EPYC processor, codenamed “Rome” and built around the Zen2 architecture, is now running in AMD labs and will begin sampling to customers in the second half of this year, and will launch in 2019
- DDR5 memory which promises to be twice as fast as DDR4
- Huge increase in NVMe density, which seems set to double every year for the next few years
- Server landscape to become increasingly heterogeneous, with more reliance on accelerators and GPUs
The server market hasn’t been this exciting for years. Maybe even decades.