IBM recently introduced three new Linux-based (LC) Power Systems targeting the hottest workload environments. These POWER8-based, enhanced models are configured to satisfy system cost, performance and processing demands of Big Data, cognitive, GPU, dense computing and memory intensive, high throughput processing. When compared to a Dell system, the newly announced IBM S822LC for HPC achieved 2.5 times the performance with costs of hardware and maintenance 52% lower! Let’s review the details.
IBM’s LC family servers are designed and cost optimized for “scale-out” multi-server cloud and cluster configured environments to satisfy customer preferences for clouds over expanding on-premise data centers.
IBM’s new lineup of LC models includes:
- The S822LC for Big Data
- The S822LC for Commercial Computing
- An S822 LC for High Performance with a new version POWER8 chip and a very high speed link between the CPU and onboard GPUs.
Other family members include:
- An “entry level” S812LC targeting customers with new memory intensive, Big Data workloads.
- An S821LC with 2 POWER8 Sockets (processors) in a 1U form factor for computing in dense database, virtualization and container environments.
We created this table to highlight key features of the different models:
Model # of CPUs #sockets Max Cores # GPUs Max threads
S812LC
|
1
|
1
|
8 or 10
|
--
|
80
|
S821LC
|
2
|
2
|
16 or 20
|
1
|
160
|
S822LC for
Big Data
|
2
|
2
|
20
|
2
|
160
|
S822LC for
Commercial
|
2
|
2
|
20
|
1
|
160
|
S822LC for
High Perf.
|
2
|
2
|
20
|
4
|
160
|
Ten-core systems have a 2.92 GHZ version of POWER8, while the 8-core systems have a 3.32 GHz chip. All include what IBM calls a 9x5, 3-year warranty with next day service.
IBM’s website[1] has
additional details on other system characteristics that may be important to existing
or planned applications.
Some Key Considerations
Complementing the scale-out systems are scale-up systems, IBM
E870 and IBM E880. These may be more appropriate for some applications. We do
not discuss those here.
The S822LC for High Performance system has characteristics
worth mentioning. There is the water cooling option which allows a turbo high
speed mode to be used extensively. Also, it uses a new version Power8 chip with
a special link to the GPUs in the system significantly speeding up the
connection between the GPU and the CPU. IBM reports the old GPU-to-CPU connection
speed via a PCIe link was 32GB/sec. The new NVLink clocks out at 80 GB/sec.
This leads us to a discussion of system performance.
Performance Background
IBM is very clear they believe the X86 has hit a barrier
regarding Moore’s law predictions of future performance improvements. Moore’s
law relates to technological performance enhancements over time. See this note.[2] As long
as the law applies, price/performance improvements were possible. However,
physics is invalidating the law for some existing technology. IBM (and others)
believe system architecture changes, not raw hardware speeds are the more
likely source of necessary future performance improvements[3].
Building on this philosophy, IBM is making changes and
adding interfaces to Power Systems to drive greater performance. A recent
example is CAPI, which we have written about elsewhere, drives large
improvements in applications using in-memory databases, support of many more
threads per core than comparable X86 system and allows more to be done, faster.
Adding Graphical Processing Units (GPU) NVLink technology to S822LC for High
Performance are other examples of improvement.
Of course, such improvements can only benefit those applications
able to take advantage of them. IBM identified those applications (emerging and
existing) to gain market advantage by providing systems optimized (cost, price
and performance) for such applications. This is the strategy to design and
optimize systems for significant market segments.
Performance Data
IBM has released performance and price performance data matching
the latest Power Systems to comparable Intel Systems. Details appear at this
URL.[4]
Summarizing IBM’s results, the best performing Power8
system, IBM S822LC for HPC, achieved 2.5 times the performance of a comparable
Dell system with 52% lower hardware and maintenance costs. The S822LC for Big Data
managed 40% better performance than a Comparable HP system with 31% lower hardware
and maintenance costs. It appears that with comparable hardware and number of
cores, Power8 systems will outperform Intel-based systems and also have a price
performance advantage.
There are caveats about these results. The benchmarks are
not industry standard. They are not supported by the TPC or Spec. IBM has made
the effort to be transparent by documenting what they did. In the past, when we
investigated IBM benchmarks of this type; we found them to be honest and
accurate. We believe someone could repeat the benchmarks and get the same
results. Having said that, it still remains a fact that any vendor-run
benchmark will be suspect in the minds of some.
The IBM results are very useful to make a potential
purchaser aware of potentially significant advantages of Power Systems. We
recommend potential customers examine Power Systems to determine the potential for
benefit in their environment.
Other Considerations
Intel holds the dominant position in the generic server
market. We believe customers benefit from competition in an open market. We
therefore support other options whether ARM-based or from AMD.
POWER8 provides a realistic alternative. We hope it flourishes.
We find the growth of the OpenPOWER Foundation to over 260 companies encouraging.
Note, we are not saying to blindly choose a non-Intel alternative. We do
believe sensible customers should carefully evaluate all options to determine
the best architecture for their business.
IBM Power Systems possess some significant advantages for specific application types and to leverage new technologies, e.g. Big Data, analytics, AI-Cognitive Computing (Watson), etc. where customers are now investing. Take a look, and decide for yourselves.
[2]
An interesting article about Moore’s
law (actually more an observation than a law) and its current state is in
Wikipedia. Our opinion is that it supports IBM’s position. See https://en.wikipedia.org/wiki/Moore%27s_law.
[3]
For another amusing look at Moore’s law see http://www.theregister.co.uk/2016/09/22/the_evolution_of_moores_law_suggests_hardware_is_eating_software/