Ever since Meltdown and Spectre were disclosed, Intel’s various customers have been asking how long it would take for hardware fixes to these problems to ship. The fixes will deploy with Cascade Lake, Intel’s next server platform due later this year, but the company is finally lifting the lid on some of those improvements and security enhancements at Hot Chips this week.
One major concern? Putting back the performance that previous solutions have lost as a result of Meltdown and Spectre. It’s hard to quantify exactly what this looks like, because the impact tends to be extremely workload-dependent. But Intel’s guidance has been in the 5-10 percent range, depending on workload and platform, and with the understanding that older CPUs were sometimes hit harder than newer ones. Intel wasn’t willing to speak to exactly what kind of uplift users should expect, but Lisa Spelman, VP of Intel’s Data Center Group, told Anandtech that the new hardware solutions would have an “impact” on the performance hit from mitigation, and that overall performance would improve at the platform level regardless.
Variant 1 will still require software-level protections, while Variant 2 (that’s the “classic” Spectre attack) will require a mixture of hardware and software protection. Variant 3 (Meltdown) will be blocked in hardware, 3a (discovered by ARM) patched via firmware, with Variant 5 (Foreshadow) also patched in hardware. Firmware fixes can be treated as equivalent to hardware since the new systems that’ll launch with Cascade Lake will include them by default.
Elsewhere, Cascade Lake will keep socket compatibility and core count equivalency when compared with Intel’s first-generation Xeon Scalable Processors. We may see performance improvements from silicon tweaks (if these chips use Intel’s 14nm++ silicon), but any real benefits are expected to be delivered either by support for new AVX-512 instructions (VNNI, or Vector Neural Net Instructions) or by Optane DC Persistent Memory. This type of enhancement is intended for inferencing workloads, not training neural nets. CPUs don’t compete well when it comes to training neural networks in the first place; the architectural mismatch is too great. Inferencing workloads, however, have shown more promise, and Intel is investing heavily in improving its performance in this space.
Overall demand for these parts is expected to be robust, given how rapidly the IT industry is shifting to cloud computing. Any talk of fallout from the initial disclosures has been limited, and there’s no sign that anyone has treated the issue as a reason to jump ship from Intel and adopt its competitors’ hardware. If ARM or AMD are seeing an uptick in adoptions based on these security issues, they have yet to say so.
Now Read: Foreshadow Flaw Cracks Intel’s Security Extensions, Report Claims Intel Core i9-9900K Will Use Solder, and Intel is at a Crossroads