SpikeBench Results

On this page you will find results of SpikeBench Benchmark executed on different hardware platforms (All copyrights/trademarks and company/product/service names are property of their respective owners)

SpikeBench "Server/WS Benchmark" 1.04 Results (CPU)

Test Description: 1 million multi-Compartment Neurons, 210 million synapses (2 receptors each)

Platform Name: Exec. CPU Name CPU Sockets NUMA Nodes Cores per CPU Total Cores CPU Speed RAM Speed Simulation Creation Time in msec (lower is better) Simulation running speed real-time factor (higher is better) Comments
On-Prem Build ("Bare Metal") Bare-metal Dual Intel Xeon E5-2699 v3
("Haswell EP")
2 2 36 72 2300 MHz

(All-core turbo locked to 3300 MHz)
DDR4 2666 MHz (out of spec) 35859ms 0.059x Bare-metal system, no hypervisor installed, boot directly to Windows. All-core turbo forced to 33x using CPU errata (on the cost of much higher TDP)
Amazon r5.24xlarge VM Dual Intel Xeon Platinum 8175M
("Skylake SP")
2 2 48 96 2500 MHz

(All-core Turbo goes to 2700 MHz during simulation)
N/A 51176 ms 0.092x
Amazon r5a.24xlarge VM Dual AMD EPYC 7571
("Naples")
2 6 48 96 2500 MHz N/A 104235ms 0.079x
Amazon x1.32xlarge VM Quad Intel Xeon E7-8800 v3
("Haswell-EX")
4 4 64 128 2300 MHz

(All-core Turbo goes to 2700 MHz during simulation)
N/A 90485ms 0.064x* For work stealing, threads were allowed to cross NUMA boundaries with -crossnumaboundaries. This helps utilize all CPUs but probably with sub-optimal performance. Further investigation is needed (possibly related to memory availability per NUMA node and/or way CPU node affinity is reported by the OS)
Google n1-ultramem-160 VM Quad Intel Xeon E7-8880 v4 (Broadwell EX) 4 4 40 160 2200 MHz N/A 92719ms 0.083x
Google n1-highmem-96 VM Dual Intel Xeon Platinum SP*** 2 2 48 96 2000 MHz N/A 101829ms 0.065x ** Note: Early-access SKU with B1 stepping and multiplier-locked to 2000 MHz
Microsoft Azure Standard M128 VM Quad Intel Xeon E7-8890 v3
("Haswell-EX")
4 4 64 128 2500 MHz N/A 37735ms 0.062x* For work stealing, threads were allowed to cross NUMA boundaries with -crossnumaboundaries. This helps utilize all CPUs but probably with sub-optimal performance. Further investigation is needed (possibly related to memory availability per NUMA node and/or way CPU node affinity is reported by the OS)

SpikeBench "Small Benchmark" 0.93 Results (CPU)

Test Description: 32768 Multi-Compartment Neurons, 1.8 million Synapses (2 receptors each)

Platform Name: Num. CPUs Num. Cores CPU Speed RAM Speed Avg. Simulation Performance Avg. Total Memory Bandwidth
Dual Intel Xeon E5-2699 v3
("Haswell EP")
2 36 2300 MHz DDR4 2400 MHz 2.690x 76.1 GiB/s
Dual Intel Xeon E5-2697 v2
("Ivy Bridge EP")
2 24 2700 MHz DDR3 2133 MHz 1.710x 55.1 GiB/s
Dual Intel Xeon E5-2687W
("Sandy Bridge EP")
2 16 3100 MHz DDR3 2133 MHz 1.170x 49.1 GiB/s
Lenovo W541
Core i7-4810MQ
("Haswell")
1 4 2800 MHz DDR3 1600 MHz 0.601x 46.04 GiB/s
(est.)
Apple Macbook Pro Retina 15" (2011)
Core i7-3720QM
("Ivy Bridge")
1 4 2600 MHz DDR3 1600 MHz 0.294x**
(** CPU throttles after ~20 sec)
19.2 GiB/s
(est.)


SpikeBench "Small Benchmark" 0.88a Results

Test Description: 32768 Multi-Compartment Neurons, 1.8 million Synapses (2 receptors each)

Platform Name: Num. CPUs Num. Cores CPU Speed RAM Speed Avg. Simulation Performance Avg. Total Memory Bandwidth
Dual Intel Xeon E5-2687W
("Sandy Bridge EP")
2 16 3100 MHz DDR3 2133 MHz 0.720x 39.10 GiB/s
Sony Vaio SVZ131
Intel Ivy Bridge
Core i7-3612QM
1 4 2100 MHz DDR3 1600 MHz 0.215x 16.5 GiB/s
(est.)
Apple Macbook Pro Retina 15"
Intel Ivy Bridge
Core i7-3720QM
1 4 2600 MHz DDR3 1600 MHz 0.230x**
(** CPU throttles after ~20 sec)
16.0 GiB/s
(est.)

* Due to slim thin design of the laptop and the significant thermal design power (TDP) of the CPU (45W) and the fan cooling strategy optimized for silence, after ~20 seconds, system will undergo heavy CPU throttling and reduction of performance by approx. 15-20%. After ~1 minute, fan will ramp up to full speed and throttling will stop, indicating that the laptop has adequate thermal headroom but due to the prioritization of low noise cooling it is not able to use all of its thermal removal capacity immediately.

SpikeBench "Small Benchmark" 0.67 Results

Test Description: 32768 Multi-Compartment Neurons, 1.8 million Synapses (2 receptors each)

Platform Name: Num. CPUs Num. Cores CPU Speed RAM Speed Avg. Simulation Performance Avg. Total Memory Bandwidth
Dual Intel Xeon E5-2687W
("Sandy Bridge EP")
2 16 3100 MHz DDR3 2133 MHz 0.525x 55.12 GiB/s
Intel Ivy Bridge
Core i7-3820QM
1 4 2700 MHz DDR3 1600 MHz 0.165x 16.2 GiB/s
(est.)
Sony Vaio SVZ131
Intel Ivy Bridge
Core i7-3612QM
1 4 2100 MHz DDR3 1600 MHz 0.161x 16.5 GiB/s
(est.)
Sony Vaio Z21
Intel Sandy Bridge
Core i7-2620M
1 2 2700 MHz DDR3 1333 MHz 0.098x
(* CPU throttles)
9.8 GiB/s
(est.)
Intel Core 2 Quad
Q6600
1 4 2666 MHz (OC) DDR2 720 MHz (OC) 0.026x 2.57 GiB/s
(est.)
Intel Core 2 Duo
"Merom"
1 2 1800 MHz DDR2 666 MHz 0.011x 1.29 GiB/s
(est.)

* Due to slim thin design of the laptop and the significant thermal design power (TDP) of the CPU when compared to the laptop size (35 Watts), one of the CPU cores cannot remain in the highest P state without resorting to thermal throttling. Performance without throttling is ~15-20% better on average.