DNA basecalling performance benchmarked, RTX3080 vs RTX1080Ti

[xkm1948]

One of the collaborators recently acquired a RTX3080FE for his work. Before retiring the 1080Ti FE he decided to compare the base calling performance between the two.

This is base calling 4GB of RAW current based DNA graph into DNA basepairs in FASTQ format. Exact same base caller, exact same parameters.

CPU is a 3970X Threadripper 32Core64Threads, 256GB DRAM

The DNA sequencer in context here:

So about 31 minutes for 1080Ti, 6 minutes for 3080

1080Ti has 3584 CUDA cores, 3080 has 8704 CUDA cores. Based on the number, the 5X performance increase definitely is way more than simply increases in CUDA cores themselves.


I have processed over 10TB of data on my 3090 so far. Things are noticeably faster than my 2080Ti stock vs stock.

It was always nice to think that I will just work less time when I have more computing power at hand. What actually ends up happening is I got more work piled up simply because I can push them through analysis faster.

 

[cine.chris]

I eyeball that at about 500% clock-watching time.
That's a huge number, especially if revenue is linked to the analysis.
Wasn't there something about them doing two backend processes per clock cycle with the 3080?

 

[dgianstefani]

Interesting. I'm tempted to use my rig for my study since I use medical and scientific software in my course at Swansea Medical school.

 

[xkm1948]

I eyeball that at about 500% clock-watching time.
That's a huge number, especially if revenue is linked to the analysis.
Wasn't there something about them doing two backend processes per clock cycle with the 3080?

Yeah hence the huge jump in CUDA counts

Interesting. I'm tempted to use my rig for my study since I use medical and scientific software in my course at Swansea Medical school.

Check out Nvidia developer here https://developer.nvidia.com/nvbio

 

[dragontamer5788]

Yeah hence the huge jump in CUDA counts

Not that the 3080 isn't impressive, but NVidia is very misleading about their CUDA-core counts. They changed what a "CUDA core" means between 2xxx and 3xxx.

It used to be 32 CUDA-cores per SM. But now its 64 CUDA-cores per SM, but they still only run 32-threads per clock tick. The threads are now superscalar though, so its a bit like the Pentium vs 486 (Pentium could run two instructions per clock tick, but the 486 could only run one per clock tick). As such, its more like one core thats faster sometimes, when integer and floating point pipelines can stay filled.

Switching to superscalar (going from 1 instruction / clock to 2 instructions/clock) isn't too big a jump in the great scheme of things. In theory, it doubles speed. But in practice, you just end up waiting on RAM more often. Besides, 1 instruction/clock is already kind of ridiculous given what needs to happen (decoding / grabbing registers / figuring out dependencies/ etc. etc.). It requires a tightly pipelined core before you can get to 1 instruction/clock theoretical, so 2 instructions/clock theoretical isn't too much of a difference. You're already taking advantage of instruction-level parallelism to get to 1 instruction/clock.

Back in the 386 days, before pipelines were common, you'd need 12 clock ticks per instruction as you went through every stage of instruction execution. (Decoding, Fetch, Execution, etc. etc.). 1-clock tick per instruction was possible due to pipelines, but you'd reach read/write hazards (which GPUs still have to resolve today). You can only reach 1-instruction (or 2-instructions/clock) if you're read/write hazard free. Same as always.

 

[P4-630]

RTX1080Ti​

It's a GTX 1080 Ti