IBM announced Power8 yesterday, ahead of a more official launch this coming Monday. Notably, in a marked contrast from previous Power generations, the core is fully licensable (like ARM) and the platform has the support of a broad ecosystem - most notably Mellanox, Canonical, Google, and Nvidia. Tyan will be providing less expensive whitebox Power8 systems later on, but the machines available Monday are IBM, and will carry an IBM pricetag (which is significantly less than the list prices of entry level systems from other high-end RISC vendors, but still higher than you'd likely pay for a 1-2 socket Supermicro whitebox.) The cheapest machine will have a list price of about US$8k.
The Power8 core is an extremely ambitious design; it can decode and issue six non-branch and two branch instructions per cycle, from as many as eight simultaneous threads. The processor as currently implemented has twelve cores and tops out around 4GHz. While industry-standard benchmarks likely won't show up until next week, I expect Power8 to be significantly faster than Intel's Ivy Bridge-EX at database and multithreaded workloads, and comparable or slightly faster on single-threaded compute loads. Single-threaded performance has become harder to measure lately, as certain subtests on the SPEC suite appear to no longer reflect CPU performance accurately (such as libquantum), but I would be surprised if Intel had any meaningful single-threaded advantage over P8. While Power8 is likely quite a bit faster than Ivy-EX, it's also thirstier; Ivy-EX high-end parts have a TDP of 155W, while the Power8 datasheet says that the P8 SCM has a TDP of 190W.
Current Oracle and Fujitsu SPARCs are likely in an even worse position against P8, as they are already slower than Ivy-EX at some loads, as measured by SPECint[1][2][4]. The good news for those platforms is that both companies have strong roadmaps with major evolutions in 2015; Fujitsu, for instance, has roadmapped a 24-core, 96-thread SPARC64 at 4.5GHz[3], which should be a good match for Power8 on multithreaded loads.
Interestingly, P8 is being marketed strongly toward Linux rather than legacy AIX or iSeries workloads. During the announcement presentation yesterday, Linux was mentioned dozens of times, while I only caught one mention of AIX and iSeries. Of the five new machines, two are Linux-only, and it's a safe bet that third-party Power8 whiteboxes will not be capable of running AIX or iSeries. While AIX and iSeries will no doubt be supported and developed into the far future, it's probable that those platforms are picking up relatively few numbers of new customers, and the focus on Linux is an attempt to focus on a wider audience while the proprietary UNIX market shrinks.
In other vendor developments, we have bits from a number of companies. Nvidia and IBM are going to be working to ship GPU/P8 integrated systems in Q4 of this year; Nvidia's new NVlink technology (available in the next-generation GPU codenamed "Pascal") is designed for high-speed integration with Power processors. Google is investigating Power8 for use in its own datacenters, and is contributing to software and firmware for the platform. A Mainland Chinese company - Suzhou PowerCore - has licensed the Power8 core and is planning to build locally-produced Power8 compatible processors for the PRC server market.
Overall, Power8 is an immensely ambitious processor surrounded by a new development model and broad industry support, and is likely to be the fastest general-purpose processor in the world. While questions remain about whether it can turn around the overall decline in the high-end RISC market, I think it has a very good shot - it's a new business model, with new partners, looking for new workloads in new price ranges.
Notes
[1]: 8 socket Xeon E7-8890 @ 2.8GHz, 120 cores, 240 threads, vs 8 socket SPARC T5 @ 3.6GHz, 128 cores, 1024 threads
[2]: 4 socket Xeon E7-4890 @ 2.8GHz, 60 cores, 120 threads, vs 4 socket Fujitsu SPARC64 X+ @ 3.7GHz, 64 cores, 128 threads
[3]: http://www.fujitsu.com/global/services/computing/server/sparc/key-reports/roadmap/
[4]: Note: The exact performance of T5 relative to Xeon remains ambiguous. T5 tends to perform quite well at Java benchmarks, but the SPEC numbers are worse than Ivy-EX; SAP SD-2, while being difficult to interpret due to being somewhat multidimensional, shows Ivy-EX running a higher number of users (albeit at a slightly higher response time) than T5. TPC-H is similar - a quad-socket Ivy-EX has a higher throughput than T5-4, but at a higher load time (but T5 is equipped with a far larger storage:database ratio.) I expect P8 to significantly outperform T5 across the board, but as always, judge by the benchmark closest to your workload.
The Power8 core is an extremely ambitious design; it can decode and issue six non-branch and two branch instructions per cycle, from as many as eight simultaneous threads. The processor as currently implemented has twelve cores and tops out around 4GHz. While industry-standard benchmarks likely won't show up until next week, I expect Power8 to be significantly faster than Intel's Ivy Bridge-EX at database and multithreaded workloads, and comparable or slightly faster on single-threaded compute loads. Single-threaded performance has become harder to measure lately, as certain subtests on the SPEC suite appear to no longer reflect CPU performance accurately (such as libquantum), but I would be surprised if Intel had any meaningful single-threaded advantage over P8. While Power8 is likely quite a bit faster than Ivy-EX, it's also thirstier; Ivy-EX high-end parts have a TDP of 155W, while the Power8 datasheet says that the P8 SCM has a TDP of 190W.
Current Oracle and Fujitsu SPARCs are likely in an even worse position against P8, as they are already slower than Ivy-EX at some loads, as measured by SPECint[1][2][4]. The good news for those platforms is that both companies have strong roadmaps with major evolutions in 2015; Fujitsu, for instance, has roadmapped a 24-core, 96-thread SPARC64 at 4.5GHz[3], which should be a good match for Power8 on multithreaded loads.
Interestingly, P8 is being marketed strongly toward Linux rather than legacy AIX or iSeries workloads. During the announcement presentation yesterday, Linux was mentioned dozens of times, while I only caught one mention of AIX and iSeries. Of the five new machines, two are Linux-only, and it's a safe bet that third-party Power8 whiteboxes will not be capable of running AIX or iSeries. While AIX and iSeries will no doubt be supported and developed into the far future, it's probable that those platforms are picking up relatively few numbers of new customers, and the focus on Linux is an attempt to focus on a wider audience while the proprietary UNIX market shrinks.
In other vendor developments, we have bits from a number of companies. Nvidia and IBM are going to be working to ship GPU/P8 integrated systems in Q4 of this year; Nvidia's new NVlink technology (available in the next-generation GPU codenamed "Pascal") is designed for high-speed integration with Power processors. Google is investigating Power8 for use in its own datacenters, and is contributing to software and firmware for the platform. A Mainland Chinese company - Suzhou PowerCore - has licensed the Power8 core and is planning to build locally-produced Power8 compatible processors for the PRC server market.
Overall, Power8 is an immensely ambitious processor surrounded by a new development model and broad industry support, and is likely to be the fastest general-purpose processor in the world. While questions remain about whether it can turn around the overall decline in the high-end RISC market, I think it has a very good shot - it's a new business model, with new partners, looking for new workloads in new price ranges.
Notes
[1]: 8 socket Xeon E7-8890 @ 2.8GHz, 120 cores, 240 threads, vs 8 socket SPARC T5 @ 3.6GHz, 128 cores, 1024 threads
[2]: 4 socket Xeon E7-4890 @ 2.8GHz, 60 cores, 120 threads, vs 4 socket Fujitsu SPARC64 X+ @ 3.7GHz, 64 cores, 128 threads
[3]: http://www.fujitsu.com/global/services/computing/server/sparc/key-reports/roadmap/
[4]: Note: The exact performance of T5 relative to Xeon remains ambiguous. T5 tends to perform quite well at Java benchmarks, but the SPEC numbers are worse than Ivy-EX; SAP SD-2, while being difficult to interpret due to being somewhat multidimensional, shows Ivy-EX running a higher number of users (albeit at a slightly higher response time) than T5. TPC-H is similar - a quad-socket Ivy-EX has a higher throughput than T5-4, but at a higher load time (but T5 is equipped with a far larger storage:database ratio.) I expect P8 to significantly outperform T5 across the board, but as always, judge by the benchmark closest to your workload.
...
-
