NOVATTE | High Performance Computing Blog

Benefits of GPU computing systems comparing to CPU-based systems

NOVATTE — Wed, 14 Mar 2012 04:54:39 +0000

Take a look at this nice video which explains advantages of GPU computing systems VS systems based on CPUs only. MythBusters hosts nailed it. This video is a Great start in the World of GPU Computing!

5 reasons to consider switching to Intel E5-2600 series CPUs starting from yesterday

NOVATTE — Fri, 09 Mar 2012 08:08:47 +0000

On Mar 6, 2012 Intel announced new sever CPUs for dual-CPU platform – Sandy Bridge E5-2600-series in which by adding just 2 additional cores it made E5-2600 almost 2 times faster than the previous X5600 generation! How did they do it and why one should consider switching to E5-based system?

Architecture difference between X5600 series and E5-2600 series

Reason 1: increased Bandwidth. The second Quick Path Interconnect (QPI) link was added to the CPU and the speed of QPI itself was increased from 6.4 Giga Transfers per second (GT/s) for X5600 series to 8GT/s for E5-2600 series resulting in huge performance increase through increased intra-CPU communication speed.

Reason 2: more CPU cores. The number of cores increased from 6 to 8, and even though the cores became smaller, the number of instructions per cycle increased from 4 (X5600 series) to 8 (E5-2600 series) in AVX resulting almost 100% performance increase for floating-point operations!

Reason 3: faster memory speed and increase RAM capacity. The number of memory channels increased from 3 to 4 per CPU plus Sandy Bidge can now support DDR3-1600 memory of up to 32GB per module (comparing to DDR3-1333 maximum memory for X5600 series) providing a maximum memory capacity of 768GB RAM per dual-CPU system.

To summarize the differences between X5600 series and E5-2600 series platforms:

	X5600 series platform	*E5-2600 series platform*
Processor	One QPI 6.4GT/s link 6 cores / 12 threads / 12MB cache Turbo 1.0	Two QPI 8.0 GT/s links 8 cores / 16 threads / 20MB cache Turbo 2.0 AVX
Memory	3 channels Up to 1333 MHz Up to 18 DIMMs Up to 288GB per system	4 channels Up to 1600MHz Up to 24 DIMMs Up to 768GB per system LRDIMMs
I/O	Two-chipIOH/ICHUp to 32 lanes of PCI Express 2.0	Integrated I/OUp to 80 lanes of PCI Express 3.0DDIO
Power Management	NM 1.5	NM 2.0

Reason 4: x2 performance increase

Server CPU Theoretical Peak Performance comparison:

	*X5600 series CPUs*		*E5-2600 series CPUs*
	Part Number	Performance, GFLOPS	Part Number	Performance, GFLOPS
Top BinLow bin	X5690 X5680 X5675 X5670 X5660 X5650	83.04 79.92 73.44 70.32 67.2 63.84	E5-2690 E5-2680 E5-2670 E5-2665 E5-2660 E5-2650	185.6 172.8 166.4 153.6 140.8 128

Reason 5: great price / performance

Price-performance comparison

X5600 series VS E5-2600 series	Price Difference, Times	Performance, Difference, Times
X5690 / E5-2690 X5680 / E5-2680 X5675 / E5-2670 X5670 / E5-2665 X5660 / E5-2660 X5650 / E5-2650	1.25 1.05 1.09 1.01 1.10 1.12	2.24 2.16 2.27 2.18 2.10 2.01

In other words for around 10% increase in price one gets more than 100% increase in performance which in tern means that one will need to buy twice as less servers to complete the same task!

How to calculate theoretical peak performance of a CPU-based HPC system

NOVATTE — Thu, 08 Mar 2012 07:54:20 +0000

To calculate theoretical peak performance of the HPC system we first need to calculate theoretical peak performance of one node (server) in GFLOPS and than just to multiply node performance on the number of nodes your HPC system has. HPC world is using the following formulae for node theoretical peak performance:

Node performance in GFLOPS = (CPU speed in GHz) x (number of CPU cores) x (CPU instruction per cycle) x (number of CPUs per node).

Note:
- Intel X5600 series CPUs and AMD 6100/6200 series CPUs have 4 instructions per cycle
- Intel E5-2600 series have 8 instructions per cycle

Example 1: Dual-CPU server based on Intel X5675 (3.06GHz 6-cores) CPUs:
3.06 x 6 x 4 x 2 = 144.88 GFLOPS

Example 2: Dual-CPU server based on Intel E5-2670 (2.6GHz 8-cores) CPUs:
2.6 x 8 x 8 x 2 = 332.8 GFLOPS
(Note that number instructions per cycle for E5-2600 series CPUs is equal to 8 )

Example 3: Dual-CPU server based on AMD 6176 (2.3GHz 12-cores) CPUs:
2.3 x 12 x 4 x 2 = 220.8 GFLOPS

Example 4: Dual-CPU server based on AMD 6274 (2.2GHz 16-cores) CPUs:
2.2 x 16 x 4 x 2 = 281.6 GFLOPS

And now, after we know that our node performance – we just multiply that number to the number of nodes we have in our system to get the theoretical peak performance of a CPU-based HPC system (we’ll use the node from Example 2 above to calculate HPC system performance consisting of 8 nodes):
332.8 GFLOPS x 8 = 2,442.4 GFLOPS = 2.44 TFLOPS

Some words about Efficiency

oxana — Thu, 10 Nov 2011 23:18:28 +0000

It has been a busy and productive period for NOVATTE. Our team just finished deployment of powerful GPU cluster for one of the Singapore Government agencies. The cluster built on new NVIDIA Tesla S2050 and NOVATTE Twin servers has the peak theoretical performance of 16 TFLOPS. The actual performance measured by the LINPACK Benchmark is 12.45 TFLOPS, providing efficiency coefficient of 77.81%. Great result for a cluster based on open source software. To compare, the second largest cluster in the world accelerated by NVIDIA Tesla C2050 GPUs has efficiency of 42.58% (http://www.top500.org/list/2010/06/100). Now NOVATTE HPC cluster is to facilitate efficiency of the customers, saving their time and resources.

Welcome

oxana — Mon, 10 Oct 2011 23:18:05 +0000

Hi There! We are glad you stopped by. We are in the process of creating blog for NOVATTE. This blog will focus on issues, solutions and trends in high-performance computing (HPC) in Asia.

Come back soon to exchange ideas on the topic, get insights and useful technical tips on different aspects of supercomputing.