Intel Xeon Refresh: New Cascade Lake Refresh CPUs up to 60 Percent Cheaper Per Core

[harlan4096]

Intel takes an axe to pricing

Intel introduced its new lineup of Cascade Lake Refresh Xeon server parts today and they bring along remarkable price reductions compared to the existing Xeon lineup, with the top-line changes to the company's flagship models amounting to a ~60% reduction in per-core pricing in exchange for the inability to scale beyond two sockets. Paired with Intel's other recent price adjustments, a 28-core SKU that used to weigh in at $13,012 now retails for a mere $3,950.

Intel also introduced more subtle changes to its mid-range processors with dual-socket models that feature more cores, cache, and higher frequencies at the same price points as previous-gen models that support more sockets, along with several new workload-optimized parts.

The Cascade Lake Refresh lineup initially came as a bit of a surprise given Intel's pending release of its Cooper Lake server platform that is slated to begin production in the first half of 2020, but the company says the refresh processors offer a more direct upgrade path for existing Xeon Scalable customers and will require little re-validation work, and, more importantly, that Cooper Lake remains on schedule.

Overall, Intel says the Cascade Lake Refresh processors offer an average of 36% more performance and 42% better performance-per-dollar compared to the first-gen Xeon Gold lineup, serving as an obvious counterblow to AMD's EPYC Rome data center processors. First, let's take a look at the most noticeable adjustments on the high-end parts.

Cascade Lake Refresh Brings up to 60% Gen-on-Gen Price Reduction Per Core

Intel added two new processors to the upper end of the Gold series, boosting the mid-range lineup from a maximum of 24 cores to 28, but the two flagship refresh models (denoted with an "R" suffix) pop off the page.

The Gold 6258R has the same specifications, including core count, base/boost frequencies, and 205W TDP as the current Xeon Platinum 8280 flagship, but Intel has reduced pricing from a comparatively eye-watering $10,009 for the 8280 to a mere $3,950 for the 6258R. That's a reduction of roughly 60%.

We also see a similar 57% reduction for the Gold 6248R, which comes with the same accommodations as its comparable sibling, but also comes with an additional 100 MHz base and boost frequency. That equates to a faster processor at a significantly lower price point.

The tradeoff? All Refresh models lose support for servers with four or more sockets (4S+), but the 4S(+) server market is much smaller than the standard dual-socket server market, so these price trimmings should benefit the majority of Intel's data center customers. The new Cascade Lake Refresh processors are based on the same 14nm process and architecture as the existing Xeon Scalable models, so they support the full roster of features like DL Boost and AVX-512, but Intel says the existing processors will remain on the market for customers interested in 4S(+) systems.

The new tier of processors will be helpful in the battle with AMD's EPYC Rome processors, which feature higher core counts at lower price points, along with increased PCIe connectivity (up to 162 lanes in some dual-socket configurations), not to mention the benefits of the PCIe 4.0 interface that is twice as fast as the PCIe 3.0 interface on Intel's server parts.

Key parts of the PCIe 4.0 ecosystem, like the first wave of enterprise-class SSDs, are finally headed to market, giving AMD a connectivity advantage, particularly in the single-socket server space where it wields 128 lanes of PCIe 4.0 connectivity compared to Intel's maximum of 48 lanes of PCIe 3.0. As such, these adjustments will improve Intel's price-to-performance ratio for not only the vast majority of its customers with dual-socket servers, but also for those that invest in Intel's own bustling single-socket server ecosystem.

Intel also recently removed its "M" tier of server processors, which commanded a $3,003 premium over standard models as a fee for stepping up to 2TB of memory capacity per chip. That support now comes free of charge and equates to a sizeable price cut in its own right because it doubles the memory capacity for standard models. AMD still leads in this area with support for up to 4TB of memory per socket, which comes free of charge. Intel also offers "L" models that support 4.5 TB of memory, but these chips now cost an additional $3,003 instead of the previous $7,987 markup over the base models. Intel tells us it made this move to lower the bar for customers interested in Optane DC Persistent Memory DIMMs, but that also means that a Platinum 8280L that retailed for $13,012 now has an equivalent processor that retails for a mere $3,950, albeit at the loss of support for 4S(+) configurations.

Mid-Range Gold Series

Intel's new mid-range Gold Series models offer a mixture of increased core counts, cache, and frequencies across the breadth of the portfolio. Many of these processors come at the same price point as their predecessors, but with beefier accommodations.

For instance, the Gold 6346R comes with four more cores than the Gold 6246, but also lands with the same $3,286 price point, which equates to a healthy 25% reduction in price-per-core. Intel has also long bifurcated its stack by adding more L3 cache in tandem with cores, topping out at the maximum available amount of cache for each die based on core counts. Apparently Intel has shied away from that practice, at least in key sections of its portfolio, and provides a big increase in L3 cache over the previous-gen chips.

These trends reverberate down the stack with significant increases in L3 cache capacity in nearly every model, implying that Intel uses its 28-core XCC silicon (38.5MB of L3 cache) for some products where it previously used 18-core HCC die (24.75MB L3 cache). You could view this in two ways: Either Intel is selling considerably more silicon area at a lower cost than it did previously, which would lead to lower margins, or the new SKUs allow the company to harvest die with defects in the 4S(+)-enabling UPI connections that it couldn't use otherwise, thus realizing some cost benefit. It could even be a mixture of the two, based upon demand.

The addition of more cores in the respective price bands leads to higher TDP envelopes at each price range, and we would typically expect that to equate to a lower base frequency so the company can squeeze within the confines of the TDP rating despite the heightened thermal output. However, Intel has mostly increased base frequencies, with the exception of two SKUs (5218R and 6240R) that decline 200 MHz (but you get four and six extra cores, respectively, in exchange). That speaks to the power of Intel's incessant optimizations to its 14nm node.

Intel also did a solid job of increasing boost frequencies, with increases from 100 MHz to 500 MHz (dependent on SKU), with the lone exception being a 100 MHz decline for the 6246R.

Intel says these alterations benefit customers that run virtualized clouds, hyper-converged infrastructures, and network function virtualization. Of course, the lower pricing will certainly be welcome for nearly any workload. Again, Intel's previous-gen Gold 6238 and 6240 also came as "L" models that commanded an additional $3,003 for 2TB of memory support, but now that feature comes standard on the refresh models, making these chips more of a gen-on-gen savings than apparent in the table above.
...

Continue Reading