AMD’s 2017 so far has been action-packed, and we haven’t even entered the second-half of the year yet. Back in March, the company released its most ambitious desktop processor series in about a decade, called Ryzen. With it, AMD delivered incredible performance gains over its predecessor and suddenly became competitive with Intel again. So competitive, actually, that it apparently forced Intel’s hand to pull some of its future launches ahead.
On the desktop, Ryzen’s biggest perks include its overall performance for its price, as well as very good power-efficiency (especially against the Bulldozer generation). Overall, that’s great for desktop guys, but desktop was just the starting point for AMD. Zen has a little somethign for everyone, but AMD really built this chip from the ground up to cater to its datacenter customers, and to be a little blunt, AMD could really use some love from that market.
Fortunately, the company has the goods to secure that love, although even despite that, it’s going to be an uphill battle to claw back share from such an Intel-dominant segment. Again, just like with desktop Ryzen, enterprise Zen, called EPYC, offers a ton of performance for the price, as well as a plethora of other important enterprise features.
I am going to follow-up to this post after returning home from seeing the EPYC launch in person, so for this post, I’ll knock out some of the most important specs and rollout plans. There is also some additional info about AMD’s other projects such as Threadripper and workstation Vega FE, from our Computex coverage.
First and foremost, EPYC is designed primarily for server customers, but there is a chance that high-end workstation folk will be able to take advantage of it, as well, thanks to AMD’s commitment to 1-socket machines. Whether you go 1-socket or 2-socket, you’ll have 128 PCIe lanes to take advantage of, up to 2TB of memory per CPU, an 8-channel memory controller, a dedicated security subsystem, and notably, an integrated chipset. AMD also commits to the socket design: second-gen EPYC will work in the first-gen platform.
Launch EPYC processors are part of the 7000 series, with 9 models in total. The bottom of the line should start out at around ~$600, while the top-end model could peak at around ~$4,000. The top dogs include 32 cores and 64 threads, and start out at 2.0 or 2.2GHz, but burst to 3.0 and 3.2GHz, respectively.
Higher clocked parts can be had at the cost of cores, so which models you choose will really depend on your overall needs. Some might find it more beneficial to have the quicker single-threaded performance, while others might better use the additional cores.
As mentioned above, all EPYC 7000 series processors include 128 PCIe lanes, but that number doesn’t double on two-socket platforms. This is due to the fact that 64 lanes from each processor will be dedicated to the Infinity Fabric bus, allowing the two chips to negotiate with each other to the point where it should be fast enough that the physical separation of the two chips is largely negated.
On the performance front, AMD says that a dual configuration of its EPYC 7601 (128 threads) beats Intel’s dual E5-2699A V4 (88 threads) by 47% in the SPECint_rate_base2006 compiler test. Shift that to floating-point, and the gain becomes 75%.
With an eight-channel memory controller, anyone is going to be naturally intrigued by the possibilities there. Well, AMD says that in the STREAM Triad test, EPYC 7601 is 2.5x better (bandwidth-wise) over the aforementioned competition (E5-2699A V4).
More on the 1-socket solutions; AMD is going to be offering three specific SKUs for customers wanting that solution, with 16, 24, and 32 core models available. Every single 7000 series EPYC processor supports memory speeds of up to DDR4-2666.
The reason AMD cares about 1-socket systems so much is that it gives regular users a ton of power without the need for a second CPU. In many cases, a single EPYC 1-socket system can replace an aging 2-socket one. 32 cores and 64 threads can sure help with that. And again, despite having just one CPU, no PCIe lanes are lost. It probably goes without saying, but 128 is a huge number of lanes to take advantage of and can result in a (seriously) I/O packed desktop-sized system.
One of the best features of EPYC is that it’s filled to the brim with security features, and every single one of those features are found on every single EPYC 7000 series chip; it doesn’t matter if it’s the bottom of the rung. AMD made its reasons for this clear: it doesn’t believe that performance should dictate a customer’s need for security. I wholeheartedly agree.
Again, I’ll talk about more EPYC specifics when I return home, but this should act as a good primer for what AMD’s delivering. It’s my understanding that EPYC will become available very soon for customer consumption (perhaps even this week), and truly, it couldn’t have gotten here soon enough. It’s going to be a very interesting Q3 for enterprise, and for great reasons. EPYC enjoyed incredible hype leading up to this launch, and by all angles, it’s looking like that hype was justified.