Intel Xeon Scalable Processors

Intel has revamped its entire server- and workstation-specific line of Xeon processors for mid-2017. The new family is now known as the Intel Xeon Scalable Processors, formerly codenamed Purley, and is based on the newer Skylake-SP architecture. This TekSpek explains the key differences and improvements over the previous generation.

Platinum, Gold, Silver, Bronze

Simplifying the model numbering a great deal by removing the previous naming structure - Xeon E5-2699 v4, for example - the new roster of chips use metals to signify the prowess and potential performance. In order of positioning there are now Platinum, Gold, Silver, and Bronze categories, and Xeon processors within each category have inter-family traits that are not necessarily shared by other, lesser chips.

Point is, Intel is going away from pure speeds and feeds as the differentiator between Xeon chips, and is more focussed on what the chip delivers from a platform ecosystem rather than model numbering being solely reliant on MHz or cores. Intel is therefore matching up the usage scenarios today and tomorrow with feature sets specific to a particular family of Xeons. Let's explain this further as we drill down into each category.

Xeon Platinum - the best of the best

The Platinum range of Xeon chips start off this generation with an 81xx model numbering system, with the higher numbers indicating more performance and subsequently higher cost. The Xeon Platinum 8180 is the fastest model, housing up to 28 cores and, with hyperthreading being present, up to 56 threads. A further 15 chips reduce the core count and frequency for users whose performance and budget demands aren't nearly so great.

These chips offer speeds of up to 3.8GHz in a Turbo state and can be configured on motherboards with 2, 4 or 8 sockets, meaning that one board can house a massive 448 threads. Note that Intel has dropped the previous Xeon E7 branding constituting the very best of the best Xeons, and these now dutifully fall under the Xeon Platinum line.

Other than faster speeds, Xeon Platinum chips offer key advantages over the previous generation. Intel has understood that many emerging datacenter workloads require a careful mix of sheer compute power, enhanced memory bandwidth and improved I/O capabilities. Addressing these needs, Xeon Platinum ups the memory-channel count to six, from four on the previous generation, and also mandates for faster and larger memory support. The upshot is significant performance boosts for a number of nascent applications. Compute wise, the Skylake-SP architecture is approximately 10 per cent faster than the previous-generation Broadwell over a wide range of tasks, though it can be much quicker depending upon workload.

And one aspect where Xeon Platinum CPUs are much faster than ever before rests with applications that take advantage of SIMD processing. Intel now includes AVX-512 support and two FMAs per core. This means that the speed-up is huge for certain programs that pack their compute into vector-optimised code, and Intel believes that focussing power and performance towards vector extensions is a wise move for the datacenter market.

On the I/O front, Intel has increased the performance and speed of the internal interconnects in order to keep pace with the greater performance on tap. There are now 48 PCIe lanes and three super-fast UPI connections, up from 40 and two on the previous generation, respectively. There is little point in bolstering the core potential and then hamstringing it with poor chip bandwidth.

Xeon Gold, Silver and Bronze

Coming in at a more agreeable price point, Xeon Scalable Gold and Silver CPUs still retain much of the goodness of their Platinum cousins. The Gold series offers up to 22 cores, the same six-channel DDR4-2666 memory support and the same performance guts that entail AVX-512 instructions and two FMAs per core. What they can't do, however, is run more than four CPUs on one motherboard, removing them somewhat from mission-critical applications that rely on massive power alone. These Gold chips can be thought of as Platinum without the cherry on top, albeit priced more moderately for markets that don't need scale-up performance.

Intel breaks the Gold line into the 61xx range, with all the features and benefits described above, but also to a cheaper 51xx range that tops out at 14 cores, drops the UPI links to two, reduces memory speed to DDR4-2400, drops one FMA per core, but still retains those six memory channels for huge bandwidth. Intelligent cuts to serve multiple markets. As something very new, a few Xeon Platinum and Gold CPUs have a built-in Omni-Path connector for hooking up to various accelerators used in the high-performance computing space.

Moving on to Silver, Intel targets these at users who need moderate performance at an attractive price. The Xeon Silver 41xx line of chips have up to 12 cores and 24 threads, offer all the benefits of Turbo Boost and AVX-512 technology, but still take heed of the need for lots of memory bandwidth by being able to access the same six channels as Platinum and Gold CPUs. The have two-socket support, as well, so are a good fit for a wide variety of applications.

Xeon Bronze 31xx, meanwhile, brings with it up to eight cores yet is the only range not to feature hyperthreading. Think of these as replacements for the incumbent Xeon E3 line, albeit with a better core and improved security features. These chips also reduce the memory speed to DDR4-2133 and, with a common theme amongst all, keep the six memory channels intact. This feature alone makes them worthy of upgrades from previous generations, given how memory-hungry most datacenter applications are.

All told, there are over 50 new processors, split between the four families, with pricing ranging from $213 through to over $13,000, per chip. All use the same Socket 3467 form factor, meaning users can upgrade as and when necessary without changing the motherboard.

Performance

Staving off the performance threat from a slew of AMD Epyc processors that now play in the same market as Xeons, Intel says the combination of improved core architecture, more cores, a new mesh topology, more memory bandwidth, AVX-512 support, and a host of other, smaller improvements lead to the Xeon Scalable Family being up to 65 per cent faster than the previous generation.

In fact, comparing a two-processor Xeon system from one just over 10 years ago, Intel reckons that performance has improved by a whopping 41x, and the jump in performance between the previous-generation Broadwell-based Xeons and the new Scalable Family is the biggest for a while. It bodes well for successive Xeons using the same compute/memory bandwidth/IO philosophy as these Scalable Processors.

Intel is also keen to point out that improving the Xeon range of CPUs is more than just adding more grunt to the cores. That works, of course, though it is the sum of platform-wide improvements that really lay down the foundation for the gains seen this time around.

The chip giant has made a concerted move away from thinking about servers from a purely CPU point of view. Emerging workloads need different performance parameters - some rely on massive CPU horsepower, some love lots of memory bandwidth, and others work best with lots of specific I/O speed. Intel hopes that by bifurcating the Xeon Scalable Family range of chips into families that excel in certain areas it has the ever-changing market covered.

Improved in every meaningful way from their predecessors, the Xeon Scalable Family represent the pinnacle of Intel engineering. Scan Computers is a premier Intel partner and are stocking a wide range Intel Xeon scalabale processros