Update: A typo in this article suggested pure-play foundry 10nm is a match for Intel’s 10nm. The correct comparison (in terms of feature sizes) is 10nm at TSMC/Samsung and 14nm at Intel. ET regrets the error.
Over the past six years, Intel has adroitly retaken and defended its data center business from all challengers. AMD was never able to win significant share for its Bulldozer and Piledriver families, and while companies like Calxeda intended to challenge Intel in high-density servers, such battles never materialized. AMD’s Epyc servers may take market share from Intel over the longer term, but AMD has repeatedly said that it’s ramping Epyc slowly through the end of 2017.
That just leaves Qualcomm, but that “just” may need to be in quotation marks. Early benchmarks on Qualcomm’s Centriq server CPUs (codenamed Falkor) by Cloudflare show the Qualcomm chips offering potent competition for Intel. The information in question comes from a blog post published by Cloudflare. I absolutely recommend giving it a read for the full details of what we’re going to discuss, but we’ll call out some salient results. First off, here’s how the three cores compare:
On paper, they’re pretty similar. Qualcomm’s expertise at building small cores clearly comes in handy on the TDP front, allowing it stack 46 cores in a package at a lower TDP than we’ve seen from Intel or AMD. The difference in nodes (Intel’s 14nm versus Qualcomm’s 10nm) shouldn’t really matter; the 10nm process node deployed at the pure-play foundries is believed to be a rough match for Intel’s own 14nm. Cloudflare puts the three chips head to head in a variety of benchmarks, including OpenSSL, compression tests, Golang (Go crypto, gzip, regexp, strings), LuaJIT, and NGINX.
Since we don’t want to pull all of Cloudflare’s results, we’ve settled on two tests that illustrate overall performance standings: Single-threaded tests (LuaJIT) and multi-threaded tests (gzip).
Benchmarks and graphs by Cloudflare
In LuaJIT, the Intel chips win single-threaded performance comparisons, hands down. This reflects Skylake and Broadwell’s overall single-thread performance in the benchmarks Cloudflare ran — the Intel chips are at their strongest here, thanks to higher innate single-threaded perf and higher target clocks. If you’re doing lightly threaded work, Intel may well be the decisive leader (though we’d ask why you’re buying a 10-12 core CPU in the first place).
In multi-threading workloads, the Qualcomm Cintriq often whacks Intel upside the head. This is, again, not particularly surprising. Intel’s own data suggests that while it has a significant clock advantage in single-threaded, where its CPUs can burst up to 3GHz+, it’s full-core maximum is well below Qualcomm, at 2.1Ghz versus 2.5GHz. And while the Intel chips have Hyper-Threading, which helps, it’s never been the same as having a full second core. Hyper-Threading would typically be expected to improve performance by 20-30 percent, which leaves Cloudflare’s Qualcomm system with a significant core advantage.
Higher core counts have a strength of their own, as seen consumer-side when AMD’s Threadripper took on Intel’s equivalently priced Core i9-7900X. The 7900X might offer higher per-core performance in many tests, but 16 cores versus 10 means that even applications that favor Intel intrinsically can still wind up as wins for AMD.
Why Intel Should Be Worried
Up until now, Intel has faced two lightweight opponents: AMD, which is only just reentering the server market with no near-term expectation of seizing huge amounts of market share (Lisa Su has stated that the firm will slowly ramp server sales through the end of 2018); and smaller ARM manufacturers who might compete in dense server applications, but haven’t brought the fight to Intel’s core markets. As the Cloudflare article notes, there are a lot of gears that have to mesh properly for an ARM vendor to take on the x86 hegemony, from proper software optimizations to RAS feature support. Server vendors are conservative. Companies buying servers are conservative. That’s why AMD has planned a slow ramp into the market, and it’s why we don’t expect to see Cintriq hoovering up a huge percentage of sales its first year out the door. We also don’t know how much the Cintriq systems cost, or how it performs in other types of workloads.
But while these numbers aren’t a knockout blow, Chipzilla will be paying serious attention nonetheless. Data centers, AI, cloud, and machine learning workloads are all critical to Intel’s future, and the x86 manufacturer isn’t going to cede ground to Qualcomm (or AMD) without a furious fight. Hat-tip to Tech Report for finding this story.