Possible Clustering discrepancy

[sanyalab]

Hello,

I ran easy-linclust with the same dataset, once using only CPU's and once using GPU's. I get 2 different results. Below are the bsub commands. I am on MMseqs2 Version: 18.8cc5c

CPU ONLY:

bsub -q normal -J JOB-I98 -n 28 -R "span[hosts=1]" -R "rusage[mem=7000]" -W 12:00 -P blastall -o `pwd`/I98C90_CPU.out.%J -e `pwd`/I98C90_CPU.err.%J \
"mmseqs easy-linclust `pwd`/sampled.fasta `pwd`/CPU_I98C90 $TMPDIR --dbtype 2 --min-seq-id 0.98 -c 0.9 --cov-mode 1 \
--cluster-mode 2 --kmer-per-seq-scale 0.3 --threads 28 --remove-tmp-files 1 --adjust-kmer-len 1 --alignment-mode 3 \
--mask 1 --similarity-type 2"

GPU ONLY:

bsub -q gpu -gpu "num=1:gmodel=NVIDIAL40S" -J JOB-I98 -n 28 -R "span[hosts=1]" -R "rusage[mem=7000]" -W 12:00 -P blastall -o `pwd`/I98C90_GPU.out.%J -e `pwd`/I98C90_GPU.err.%J \
"mmseqs easy-linclust `pwd`/sampled.fasta `pwd`/GPU_I98C90 $TMPDIR --dbtype 2 --min-seq-id 0.98 -c 0.9 --cov-mode 1 \
--cluster-mode 2 --kmer-per-seq-scale 0.3 --createdb-mode 2 --gpu 1 --threads 28 --remove-tmp-files 1 --adjust-kmer-len 1 --alignment-mode 3 \
--mask 1 --similarity-type 2"

I started the clustering on 1 million FLNC sequences (DNA). The number of representative sequences using CPU ONLY = 40261. The number of representative sequences using GPU ONLY = 314732. Thats a huge discrepancy. Additionally, in the GPU only run the *_rep_seq.fasta file only has the ids. No sequence. The CPU ONLY mode is OK. Please advice if my parameters are incorrect and what might be happening. Which one should I believe?

Thanks
Abhijit