There is no one thing as "HPC". It is a lot of normal computer components joined together to form a big "HPC" system. First, identify which kind of scaling you need - are you limited by CPU or memory if you run the code on your PC?

On a side note, please don't think about buying HPC hardware before actually write the code. You can build the code to solve real problems on your normal PC with a multi core CPU and GPU. Then you can benckmark your parallel code to determine what you acutally need.

I could help you out if you can give me more details of what you actually want to achieve.

The one factor that distinguishes a typical HPC installation from what we would typically call a 'cloud' system is the throughpout. HPC installations are designed to provide maximum throughput and this requires very specialized hardware and networking infrastructure (InfiniBand). Cloud installations in contrast make use of widely available commodity hardware (e.g. Raspberry Pis). Unless the tasks you want to do will require an enormous amount of throughput for inter-node messaging within your cluster, then I'd suggest orienting your plans towards the 'cloud' model aka using commodity hardware and taking advantage of an open source cloud platform like OpenStack.

If you don’t need to leave it running between computations, this seems like a reasonable EC2/GCP use case.

University scientific computing labs cost millions of dollars; anything you’re going to build and run at home will be a small-scale replica.

> anything you’re going to build and run at home will be a small-scale replica.

If you’re OK with the performance of a top-tanking 2010 supercomputer, and the task is a good fit for GPUs (=the numbers being crunched are single-precision floats, not too much bandwidth is required, the “hot” set of the data is within a few GB i.e. fits in VRAM), you can build one at home spending very reasonable amount of cash.

Here’s a top 500 list for June 2010: https://www.top500.org/lists/2010/06/ As you see, the system at #10 delivers 433 TFlops. That’s just 40 modern GeForce 1080Ti GPUs. What once costed many millions of dollars and consumed 4 megawatts of electricity, now costs $40k and consumes little over 10KW.

And because cryptocurrency craziness, many-GPU motherboards, cables, PSUs and other components are available on the market and aren’t too expensive. For example, Asus offers B250 mining motherboard with 19 PCIe slots for $140.

Depends on what exactly you gonna crunch.

If you're OK running on a GPUs, you essentially need a cryptocurrency mining rig, there lots of articles about them.

If you need CPU performance look for used servers on eBay, either complete ones or separate components.

You can get very far with just a normal desktop computer with a bunch of CPUs. E.g. http://veekaybee.github.io/2017/03/20/hadoop-or-laptop/

A low cost HPC number cruncher implies you seek the highest performance at its lowest price. The energy used over its lifetime will be the main factor. The capital cost of the processor will probably be less important than the price for the bulk memory.

First you should identify the price performance per watt of the systems (processor+memory+network) over their lifetime in your use case. This turns out to be very difficult. The few benchmarks you find online can not be easily compared. The interesting choices are usually just released and have no benchmarks yet. More problematic, benchmarks seldom list the energy and cost of a system

every year I try to design the lowest price/performance/watt system I can find, I spend a few weeks on this. Even with that much effort I have not been able to establish if a cluster of raspberry pi's is faster and cheaper than an AMD EPYC with several cpu's. Establishing the cost is even harder, if you want to manufacture enough of the systems to benefit from scale. (not a problem in a homebrew system)

Currently I guestimate that the best price/performance/watt system is some mass produced $1-$3 ARM (soon RISC-V) with two DRAM chips and a fast network connection like an FPGA with high speed SERDES link fabric switch. Price per core plus memory must be way below $10. Raspberry pi's are no contenders, they only have 200-300 Mbps network for 4 cores to share, that should be several gigabits/s to be competitive.

A custom build AMD EPYC with several GPU's networked together can turn out to be faster and cheaper. The retail price needs to be lowered and you need to build a similiar network fabric switch.

You can find better performance if you tailor the processor design to the task. So a system with the best price/performance/watt will be different for different software.

Even cheaper will be systems where you can rebalance part of the compute recources between different programs. The system will can less efficient because the hardware reconfigurability has has a high overhead, but that is offset by being more efficient for one program. You could for example balance transistors between integer, floating point, cache or network-on-chip (NoC).

We are currently making prototypes of our own design of a reconfgurable manycore processor with NoC fabric. Around $9 per core in an FPGA. The entry system costs around $80 and $500.

We plan to build this as an ASIC, the price will then drop to $1 per core including DRAM.

Even cheaper will be if we not slice the wafer into 22000 chips but leave it whole. You get over 100.000 cores with little memory with petabits/s network for less than $6000. Half the wafer is reconfigurable logic that can be reprogrammed at runtime as GPU, TPU, CPU or any other custom optimisation. The energy cost of the wafer can be zero if you use the wafer as a water heating element and lower if you only run it on solar PV during the day. (If you share two wafers with a user on the night time side of the earth, you both can have 24 hour computing on $0,02 per kWh of solar PV). We can make a 180nm $500 version and a $1 version but they would not be as good as the 28nm $6000 version. A future 7nm wafer scale version might never be cheaper than the 28nm version, we will have to wait and see.

An silicon optical network on the wafer (also $6000) would allow two or more wafers to be networked at several terabits/s. This overcomes its small memory problem.

Think of it, you want to get rid of any overhead in the system, like pcb, chip package, connectors, cables. You want to put everything on a single large chip, the whole wafer. Because we have different CMOS technology for DRAM or processors, we wind up with needing one wafer.

We are confident in 30 years we can grow the wafer(or 3D block) and the solar panel from CO2.

Can you elaborate on this? I am a programmer by day, amateur scientist by night and I am interested i this exact process. Where can I read more about this? Do you have a corporate landing page for me to start at?

You can find my email in my profile.

Which exact process you want me to elaborate on? I can gather some papers to read more on this. https://scholar.google.nl/scholar?q=siliconsqueak

Our corporate landing page is a mess this month, with broken links http://morphle.com

Thank you so much for your insightful and educational response.