Supercomputer Cluster Build - Chapter 2 - Compute Nodes (x4)

CPU : Intel 7210 Xeon Phi - Knights Landing (64 CORES @ 1.3 GHz)
MCDRAM : 16 GB (400 GB/s) integrated on CPU chip, like GPU memory
SIMD Extensions (SSE) : AVX-512, FMA

the first generation of this CPU was a coprocessor like a GPU
this is the 2nd generation Xeon Phi architecture called Knights Landing
it has all the characteristics of a GPU but its
an
x86 64-bit, OS booting, compiled C code running

each CPU has 64 cores
each core has 2 VPUs (Vector Processing Units)
each VPU has registers 512 bits wide! (512/8 = 64 Bytes)
single precision floating point number is 4 Bytes (1 VPU can fit 16)
double precision floating point number is 8 Bytes (1 VPU can fit 8)
each VPU can execute 1operation on the entire register every clock cycle

for single precision FLOPS (Floating Point Operations Per Second) of:
16*2*64* (1.3 GHz)
2662 G FLOPS
2,662,000,000,000 calculations per second (add multiply subtract, anything in the instruction set)

This page from Intel has official FLOPS values of all their processors
https://www.intel.com/content/www/us/...
Including Zeon Phi
https://www.intel.com/content/dam/sup...
showing exactly the number I calculate here! So now you can do it too! WHOOO!

I have 4 of these CPUs MUHHUUHAHAHAHAHA!

but a lot of times the CPU FLOPS doesn't matter, its the memory bandwidth that limits performance
this is especially true for Finite Difference Time Domain (FDTD) Algorithms
which read and write the entire memory domain every time step
the CPU is sitting there waiting for the memory
means you can add extra local calculations per cell
sub-grid modeling, nonlinear material response, smoothing, without losing any time
More on FDTD later


-----------------------------------------------------------------------------------
The Xeon Phi Bible
https://www.google.com/books/edition/...

MCDRAM On Chip Memory like a GPU!
https://en.wikipedia.org/wiki/MCDRAM
https://colfaxresearch.com/knl-mcdram/
https://www.alcf.anl.gov/files/HC27.2...

Xeon Phi Instruction Set
https://link.springer.com/chapter/10....

Info from Intel
https://ark.intel.com/content/www/us/...

Some of the worlds best supercomputers (like Cori at NERSC) use Xeon Phi Knights Landing CPUs
https://www.nersc.gov/assets/Uploads/...
this one reaches (30,000,000 GFLOPS !!!)
---------------------------------------------------------------------------------------------------------------

Each node has one Xeon Phi CPU and:

RAM : 200 GB, 6-channel DDR4-2133 (100 GB/s), 32 GB each module

System Drive : SEKC SATA III SSD 128 GB (400 MB/s)
Operating System : Linux - CentOS 8

Server : Intel HNS7200AP Compute Module
Motherboard : Intel S7200AP Server Board
CPU Cooling: Super Crazy Turbo Server Fans

PCIe 3.0 lanes : 2x x16

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

Checkout the next chapters in this Supercomputer Cluster Series for more!