What's the most powerful non-clustered supercomputer?
 

Seems almost all the top500 supercomputers are clusters of thousands of individual systems, which require parallelized software to run in a clustered configuration.
Wondering what is the most powerful monolithic (aka non-clustered) supercomputer out there today?
Pick your favorite benchmark, just wondering how powerful a single system can be without resorting to clustering.

Is a blade server 1 system or 14 systems?

I started to tell you, but then I noticed that people were pausing outside my office window. They were driving curiously-large black trucks and wearing sunglasses, and there seemed to be a funny bulge inside of their gray suit-coats. So, I decided against it. ;)

Check this out: http://www.networkworld.com/article/...th.html#slide2

I believe they are treated as a cluster of 14 nodes...

None of them.

The problem is the architecture.

All current processors have a single bus to memory (even the multi-core processors still only have a single bus). SOMETIMES that single bus is multiplexed to memory (which is why you have to install DIMMs in pairs). The one apparent exception is the IBM Power line (Power7 appears to have 3 memory buses).

NONE of them have a good memory bus OR a good I/O bus.

To gain speed the only solution currently available is to give each processor (or multi-core) its own private memory, and its own private I/O bus. On top of that you then have to add custom interfaces... specifically a high speed inter-node communication bus (which by the way, can saturate the private I/O bus quite easily). But now you are in the realm of multi-node clustering.

The fastest single core CPU is likely the IBM Power 8... which always comes in multi-core packages used in SMP (an 4,6,8,12 core hyperthreaded units for up to 96 parallel threads). The memory controllers use three buses (two read, one write) to aggregate memory memory speed (40 ns latency). https://en.wikipedia.org/wiki/POWER8

But for supercomputer use, it STILL takes clustering to produce the throughput...

All supercomputers in the top 10 or top 100 lists are heavily clustered systems. There is however a market for non clustered supercomputers, i.e. systems where a single operating system instance is able to manage the whole hardware (CPU and RAM) and a single multithreaded process is able to use all the CPU cores and threads available. At least Oracle, Fujitsu, IBM and possibly still HPE compete here.

An Oracle SPARC M7-16 is a single monolithic computer with 8 TB or RAM, 512 x 4.13GHz cores / 4096 threads. The M7 series architecture is designed to support up to 64 sockets (i.e. 2048 cores / 16384 threads) and 32 TB of RAM, but there is not (yet?) such a machine on the catalog. There are also the Fujitsu M10-4s with up to 1024 x 3.7 GHz cores / 2048 threads and up to 32 TB of RAM and the IBM Power 8 E880 with 192 4 GhZ cores / 1536 threads and up to 32 TB of RAM. I'm not sure about what HPE biggest monolithic server is.

512 cores is no longer "non clustered". It is SMP... but that also requires parallel programming for maximum throughput.

But a single core Power8 beats it quite handily.

I agree these servers low level architecture is providing a distributed infrastructure that can be described as a non uniform cluster of interconnected cores where memory locality matters. There is however quite a big difference between cluster based parallel software where plenty of operating systems and processes need to communicate and cooperate together and traditional multi-threading where everything can be done in a single process using a single memory space.

My understanding is the OP is asking about the latter and not about single core performance.

You could be right. In which case the Power8 is still top.

Even in SMP you have to have multiple processes that need to cooperate rather than "a single process using a single memory space", which I interpreted as a single thread.

A single process using a single memory space doesn't need to be single threaded, it certainly can be heavily multi threaded. All of these servers do support large scale SMT applications. (https://en.wikipedia.org/wiki/Multit...r_architecture).

About which CPU leads, these blogs differ from your claim: One SPARC beats two Power8 and SPARC T7-1 Delivers 1-Chip World Records for SPEC CPU2006 Rate Benchmarks but I must disclose I'm biased, being in the SPARC side since a couple of decades...

Oracle is biased... the processor compared was one of the slower Power8 models in 2006 :) IBM has made the fastest for quite some time. (Power8 can run at 5GHz, and the Z series can reach 5.5).

But it doesn't matter that much anyway - the cost of any of these is sky high...

Companies have been fudging benchmarks forever. Barely worth spit.