Update introslides

intro-to-hpc/docs/01-intro-supercomputing.md

@@ -4,6 +4,8 @@ event:  CSC Summer School in High-Performance Computing 2024
 lang:   en
 ---
 
+# This lecture
+
 # What is high-performance computing?
 
 - Utilizing computing power that is much larger than available in a typical desktop computer
@@ -16,24 +18,18 @@ lang:   en
 <!-- Copyright CSC -->
  ![](img/cray.png){.center width=30%}
 
+# What is high-performance computing?
+
+<!-- Copyright CSC -->
+ ![](img/LUMI_supercomputer.png){.center width=40%}
+
 # Top 500 list
 
 <!-- Source: top500.org, Copyright 1993-2023 TOP500.org (c) -->
  ![](img/top_500.png){.center width=48%}
 
 # What are supercomputers used for? {.section}
 
-# General use cases
-
-- Simulations of very different scales
-    - From subatomic particles to cosmic scales
-- Problems with very large datasets
-- Complex computational problems
-- Problems that are hard to experiment on
-    - Simulations with decade-long timescales
-- Many of these are very time consuming, or even<br>
-  impossible to solve on a standard computer
-
 # Application areas
 
 - Fundamental sciences such as particle physics and cosmology
@@ -44,7 +40,18 @@ lang:   en
 - Engineering, e.g. manufacturing and infrastructure
 - etc.
 
-# Climate change
+# General use cases
+
+- Problems that are hard to experiment on
+    - Simulations with decade-long timescales
+- Simulations of very different scales
+    - From subatomic particles to cosmic scales
+- Problems with very large datasets
+- Complex computational problems
+- Many of these are very time consuming, or even<br>
+  impossible to solve on a standard computer
+
+# Climate and weather research
 
 <div class=column>
 - Simulating ice sheets, air pollutants, sea-level rise etc.
@@ -68,18 +75,6 @@ lang:   en
 ![](img/covid.png){.center width=100%}
 </div>
 
-# Gravitational waves
-
-<div class=column>
-- Computational modeling of sources of gravitational waves
-- Identifying a phase transition of the Higgs boson “turning on” (10 picoseconds after Big Bang)
-- Large simulations with over ten thousand CPU cores
-- Experimental data from ESA's LISA satellite (Launch date 2037)
-</div>
-<div class=column>
-![](img/grav.png){.center width=90%}
-</div>
-
 #  Topological superconductors
 
 <div class=column>
@@ -97,6 +92,7 @@ lang:   en
 <div class=column>
 - Web-scale Finnish language data together with very deep neural networks utilizing GPUs
 - Open GPT-3 model ("Finnish ChatGPT")
+- The large language model "Poro"
 </div>
 <div class=column>
 <!-- Source: Adobe Stock, CC BY-SA 3.0 -->
@@ -122,8 +118,8 @@ lang:   en
 
 - Modern (super)computers rely on parallel processing
 - **Multiple** CPU cores & accelerators (GPUs)
-    - `#`1 system has `~`600 000 cores and `~`40 000 GPUs
-    - `#`2 system (CPU-only) has `~`8 000 000 cores
+    - `#`1 system has `~`600 000 CPU cores and `~`40 000 GPUs
+    - `#`5 system (CPU-only) has `~`8 000 000 cores
 - Vectorization
     - A single instruction can process multiple data (SIMD)
 - Instruction level parallelism and pipelining
@@ -140,7 +136,6 @@ lang:   en
     - Latency `~`1 µs, bandwidth `~`100 GB / s
 - A node can contain several multicore CPUs and several GPUs
 - Memory within the node is directly usable by all CPU cores
-- GPUs have their own local memory
 
 # Supercomputer autopsy – Lumi
 
@@ -166,7 +161,7 @@ lang:   en
 - Cloud infrastructure is run on top of normal HPC system:
     - Shared memory nodes connected by network
 - User obtains **virtual** machines
-- Several providers offer also bare metal instances
+- Several providers also offer bare metal instances
 - Infrastructure as a service (IaaS)
     - User has full freedom (and responsibility) of operating system and the whole software environment
 - Platform as a service (PaaS)
@@ -176,7 +171,7 @@ lang:   en
 
 - Suitability of cloud computing for HPC depends heavily on application
 - Virtualization adds overhead especially for the networking
-- Bare metal cloud with high-speed interconnects can provide similar performance as traditional cluster
+- Bare metal cloud resources with high-speed interconnects can provide similar performance as traditional cluster
 - Moving data out from the cloud can be time-consuming (and have a monetary cost)
 - Cost-effectiveness of cloud depends heavily on the use case
 
@@ -185,8 +180,8 @@ lang:   en
 # Containers
 
 - Problem: Moving applications between different systems
-    - Library dependencies
-    - Using different library versions
+    - Different software stacks
+    - Different library versions, compilers
     - Different operating systems
 - Security
     - Isolating an application
@@ -230,5 +225,5 @@ lang:   en
 - Performance of supercomputers has increased exponentially for a long time
 - However, there are still challenges in continuing onwards from exascale supercomputers ($> 1 \times 10^{18}$ flop/s)
     - Power consumption: current `#`1 energy efficient system requires `~`20 MW for exascale performances
-    - Cost & Maintaining: Global chip shortage
+    - Manufacturing: Transistor sizes being extremely small 
     - Application scalability: how to program for 100 000 GPUs / 100 000 000 cores?