Titan supercomputer (render) - source: Wikipedia
Distributed supercomputers like BOINC and F@H are sometimes compared to classic supercomputers like Titan (currently regarded as the 4th most powerful supercomputer in the world). Supercomputers are designed to have low latency and high memory bandwith and comparisons are a bit of apples vs oranges type. However, low latency and high memory bandwith might not be important for some types of research. In such cases comparisons might be more apples to apples.
I must admit that I have failed in an attempt to compare above three architectures / networks. Compute power is often depicted as number of operations on single precision (SP or FP32) numbers per second called FLOPS. There are however several ways to measure FLOPS. There are also other types of numbers like double precision (DP or FP64) and integer and some calculations use more one than other types of numbers.
- Titan supercomputer was benchmarked with LINPACK and has known number and type of CPUs and GPUs. Titan was designed to run GPU optimized programs and CPUs to offer ‘feeding’ GPUs.
- Folding at Home (F@H) project relies on FP32 computation at large part and computations can be done either using CPUs or GPUs, but using CPUs seems like a waste of resources.
- BOINC - the big problem. There are dozens of different projects run on BOINC platform. Some can be executed only on CPUs, some on both, CPUs and GPUs; some benefit from high FP32 hardware capabilities and others from high FP64 capabilities (MilkyWay@Home). Credits are rewarded in strange ways - there is no consistency between projects. As I have not enough hardware to probe all or most of the projects and finding proper data stats available online is not easy and time consuming, I had to take several short cuts and rough estimations. Therefore, take below numbers with a pinch of salt.
Methodology
Based on efficiency of NVIDIA GTX 1080 card in particular projects and quite old now XEON X5650 processor I have calculated (estimated) equivalent hardware networks for F@H and BOINC networks. GTX 1080 is a very popular high end consumer card and X5650 (12 threads) is comparable with AMD Opteron 6274 (16 threads) used in Titan. As for F@H I have reduced computational network to n cards and two CPU cores per card (although one is enough to support GPU, lets leave another one for OS).
I have assumed a cost of $800 for GTX 1080 and $400 for a X5650 based computer
- F@H – I have assumed one GTX1080 can deliver 700k PPD (Point Per Day)
- BOINC – I have reviewed most (but not all) Gridcoin whitelisted projects. I have estimated equivalent number of GTX1080 GPUs or X5650 CPUs based on their performance. The rest is gestimation. Also, in way I have omitted MilkyWay@Home (treated as SP project).
- CPU cores = CPU threads
Some points:
As for BOINC numbers I’ve done some very rough estimations.
F@H claims around 88 PetaFLOPS AND 42 PetaFLOPS while BOINC 25 PetaFLOPS and Titan 17.6 AND 27 PetaFLOPS.
K20X card has much higher FP64 performance than GTX 1080.
‘FP32 – PetaFLOPS’ is a sum of manufacturer performance of a single cards; CPU and FP64 capabilities are omitted.
PetaFLOPS adj (0.24) – calculated capabilities are reduced by 0.24 factor to much LINPACK benchmark of the Titan supercomputer.
Cost – GTX 1080 cost assumed at $800 and X5650 based computer at $480, while for Titan it’s an official cost at time of building. X5650 CPU was over $1000 a few years ago, but now more powerful complete PC can be bought under $1000. Maybe I should find better numbers for cost estimations.
In real life situations, GPU efficiency often can reach as little as 10 or 20 % of potential capabilities;
F@H seems to report (quite well) theoretical maximum speed of the F@H network
Supplement
Some market researchers claim around 3 million graphic cards were bought in 2017 by cryptocurrency miners. As some were slower than GTX 1080, assuming they are an equivalent of one million GTX 1080 cards, F@H and BOINC combined capabilities are only at around 1.5%... Would Gridcoin or Curecoin become popular... what impact on mining difficulty, and involved scientific projects would be?
Links to related articles I've published before:
Counting FLOPSs in a FLOPS - part 1
Counting FLOPSs in a FLOPS – part 2
Time and Work Based Model for GridCoin Reward Mechanism