Skip to main content

Tencent Launches AMD EPYC-Powered "Star Lake" Servers, Claims 35 Percent Performance Improvement

(Image credit: AMD)

Today AMD announced that it had formed a partnership with Tencent, the largest Chinese data center operator, and said it would provide Tencent with Epyc Rome CPUs for Tencent's new "Star Lake server platform" (no relation to Intel). This is yet another win for AMD after winning several supercomputers, and companies that haven't already invested into AMD's server CPUs are certainly looking into it.

Star Lake is Tencent's custom server implementation of Rome, described as "self-designed." This custom design, according to Tencent, improves power efficiency under maximum load by 50% by using "advanced thermosyphon heat dissipation technology." It's not clear what exactly that is beyond a name for a proprietary cooling solution, but the improved cooling obviously impacts performance, as hotter processors become less efficient because they consume more power. AMD did not say which specific Epyc CPU Tencent will use, but it could be the 7H12, which is AMD's highest-performing and most power-hungry Epyc CPU. That would explain why just improving the cooling solution would create such a drastic change. Tencent nor AMD gave a figure for power efficiency improvements overall, but AMD emphasized power efficiency and a "significant TCO (total cost of ownership) advantage," which would include power efficiency.

The performance gains are also impressive, with Tencent claiming a 35% boost overall compared to its previous servers. In one specific workload, in page QPS (queries per second), Tencent noted a 150% performance improvement. Tencent didn't share what drove the massive performance improvement in QPS, but it could be thanks to Rome's extremely high floating point performance; a 32-core Rome CPU has twice the floating point performance of a 32-core Naples CPU.

Most importantly, however, Tencent says Rome can "meet 98% of Tencent cloud application scenarios." This is extremely important for AMD that Rome can be this flexible; Naples (and other Zen-based CPUs) had issues with being much faster than Xeon in some applications but much slower in others, making the prospect of using Naples appealing to those whose needs were more specific.

Rome's ability to handle basically all of Tencent's needs is a big victory. The company hosts China's most expansive data center with 1.1 million servers and continues to deploy more.

Tencent didn't share any figures for the number of chips it will deploy, but AMD will likely at least be Tencent's provider for future server deployments, and perhaps Tencent will even replace some of its existing servers with its Star Lake servers. It certainly cannot be understated how impressive the adoption of Rome has been, especially given the sluggishness of large data centers to adopt new architectures.