Modern Computational Techniques for the HMMER Sequence Analysis

Table 2

Performance comparison among different hardware approaches.

Acceleration hardware type	Hardware accelerator	Accelerated programs	Datasets	Host or base hardware environment	Reported max. speedup
Network Processor	Intel IXP 2850 network processor	Viterbi	Pfam_ls database (7459 models)	2.6 GHz Intel Pentium 4 CPU with 768 MB of SDRAM and 32 MB of QDR SRAM	1.82x [14]
GPGPU	8800 GTX Ultra	hmmsearch	3 GB nr Database (5.5 million sequences) 3 models (77, 209, 456, 789, and 1431 states)	—	38.6x [15]
	4 Tesla C1060s	hmmsearch	5.4 GB Database (10.54 million sequences) 3 models (128, 256 and 507 states)	2.33 GHz AMD Opteron	100+x [16]
Heterogeneous multi-core chip	CELL BE	hmmpfam	100 HMM states and characters	Dual-core 2.4 GHz Opteron with 8 GB RAM	~3.5x [17]
FPGA	Spartan-3 XC3S1500	hmmsearch	244 HMM states and a database consisting of 643,552 sequences	AMD Athlon 64 3500+	31x [18]
	Spartan-3 XC3S1500	hmmpfam	A database consisting of 1,544 HMMs and 1000 protein sequences	AMD Athlon 64 3500+	39x [18]
	Virtex-5 110T	hmmsearch	122,564 query inputs	2.33 GHz Intel Core2 Duo with 4 GB RAM	56.8x [19]
Heterogeneous approach (MPI  +  FPGA)	2 Spartan-3 XC3S1500	hmmsearch	2 HMM models (77 and 236 states) and 2 databases (217,875 and 2,521,679 sequences)	A cluster consists of 10 worker nodes, each with a dual core AMD Opteron 175 processor with 2 GB memory per node	30x [13]