SHARING AMERICA'S TECH NEWS FROM THE VALLEY TO THE ALLEY
Ever since AMD launched its low-power Brazos architecture in 2011, there’s been chatter over whether or not we’d see a low-power server version of the architecture. AMD declined to go that route with Brazos, but now, with its second-generation APU, the company is going to make a move for server products. Kabini is coming to the server space as the AMD Kyoto. AMD is already talking up its win with HP, as part of Project Moonshot, with the implication that there are other announcements coming in the near future.
How will the new X2150 (APU) or X1150 (CPU) compare against Intel’s S1200? AMD is clearly confident that benchmarks will favor them, and based on what we know of Atom, that’s probably true. Power consumption for the X2150 is somewhat higher than the S1200 (11W TDP, compared to 8.5W for Intel), but the performance difference should be substantial. The chart below should explain why:
Atom’s single-threaded performance has always been abysmal, even compared to Kabini. The S1260 doesn’t have a clock speed advantage — the S1260 and X2150 are both 2GHz parts — and it’s only dual-core/quad-thread, vs. AMD’s full quad-core setup. That’s not a comparison Intel’s 32nm Atom core is ever going to win, particularly not given the substantially improved FPU that Kabini is sporting.
The benefit of up to 32GB of RAM per socket and the additional GPU cores is a bit uncertain. Kabini’s 128 GPU cores are definitely a benefit in tests like Luxmark, CLBenchmark, or in tests like Musemage, which leverages OpenCL — TechReport’s review of the laptop version of the APU show the chip’s OpenCL capabilities giving it a significant boost in these benchmarks. In other tests, the benefit was significantly smaller. Clearly, the GCN-derived Radeon cores can be useful — but how many server workloads are being optimized for OpenCL code?
Similarly, 32GB of RAM per socket seems like a bit of overkill given that Kabini has just one channel of DDR3-1600 RAM. To date, most of the small-core servers from ARM and Intel have focused on smaller amounts of memory, seemingly without detriment.
You might think that these parts were a shoo-in for a SeaMicro server, but AMD was oddly reticent to confirm or deny that possibility. SeaMicro’s SM15000 server is a compute system offered with either 64 AMD Opterons, 64 Intel Xeon’s, or 256 Atom processors linked via the company’s Freedom Fabric ASIC. When we asked if the SM15000 would be updated to add a Kabini option to the portfolio, AMD had no comment. Presumably, we will see these APUs in a SeaMicro system at some point.
The best thing about AMD’s Kyoto is that the company is moving to stake a claim in a nascent market, rather than scurrying to play catch-up with ARM and Intel. Low-power servers are still a wide-open game. This category isn’t going to be a major money-maker for Sunnyvale any time in the near future, but it’s important for the company to demonstrate that it can adroitly address new markets as they happen rather than coming to market 18-24 months after Chipzilla. Everyone is new to this market — Calxeda, which made waves with initial ARM announcements several years ago, only recently began fielding its first generation of Cortex-A9 based products.
Kyoto will offer higher single-threaded performance than current in-order Atom hardware or ARM equivalents, but should still be capable of hitting similar rack densities. The GPU cores don’t seem to be very important — server workloads with OpenCL optimizations are few and far between, which is likely why AMD offers a version of the chip with the GPU disabled and a lower total TDP. Avoton, which is based on the Silvermont (Bay Trail) core and launches later this year, could make trouble for AMD’s plans in this segment. For now, though, Sunnyvale has a first-mover advantage.
Hopefully the company has plans to make use of it.
Thank you. TiA.