C (Intel Core Duo + OS X/Darwin), 248 bytes

#include <unistd.h>
#include <mach/mach_time.h>
#define M mach_absolute_time()
main(){char*s="JJJIHGFGHIJJJ@BDFHJJJHFDB@JJJJ@JJJJBBBBBBB";uint64_t i,t,y=1;for(;*s;s++){
for(i=40;i;i--){for(t=M+(*s&15)*9090909;t>M;)y*=7;usleep((11-(*s&15))*9091);}}}

This code is about as portable as the Great Pyramid of Cheops. Sorry about that. The values returned from mach_absolute_time() are hardware-dependent, but on my machine the value increments about once per nanosecond.

Here's the result:

There are two graphs because the processor has two cores. I set the maximum CPU load to about 90% because the process is liable to switch between cores whenever I call usleep(). With a 100% load, the process is chained to one core and the results are illegible (see this, for example)

Python, 143

from time import*
while 1:
 sleep((ord('00012345654321000~~~D:6300036:D~~~000~~~000DDDD~~~~~'[int(time())%52])-48)*0.001);x=10**5
 while x:x-=1

Each character of the string corresponds to one second of activity, from the ASCII character 0 (max load) through to ~ (very light load). The program runs on a time-synchronised loop, so you can run multiple instances for nicer results.

I used Python 2.7.6 on OS X with an Intel Core i7, but it should work on other computers with a bit of tweaking (adjust the 0.001). The screenshot below was taken with significant background activity.

Update - I was able to produce a clearer graph with time()/10 and a lower update frequency:

And finally, here's a more golfed version (123 bytes) and its result:

from time import*
while 1:
 sleep((ord('002464200~~A5005A~~00~~00DDD~~'[int(time()/2)%30])-48)*0.001);x=10**5
 while x:x-=1

Ruby, 150 characters

a=(0..15).map{|i|[0.9-3*i*=0.02,i]}
[9,*a[0,11],*(z=a.reverse)[5,11],11,*z,*a,2,11,6,*[0.2]*9].map{|x,y|c=Time.now
1until Time.now-c>x/3
sleep y||x%3}

This isn't all that short so far, but in my opinion the output's rather nice, so I figured I'd post this anyway. As with most other solutions, you may have to pin the Ruby process to a certain core by prefixing it with taskset -c $core.

The code is a simple combination of spinning/sleeping for a certain amount of time, which should make it somewhat portable. Smooth gradients are created by varying the ratio of spin/sleep time.

Lowering the CPU sampling frequency makes the edges look a bit better:

By adding a few more letters to the alphabet (AILMNUVW are somewhat recognizable), we can also write some other words:

These pictures were generated with the following code:

def gradient num_samples, direction, base = 0.3, increment = 0.02, scale = 1
    range = [*0..num_samples]

    samples = case direction
        when :up then range.reverse
        when :down then range
        when :updown then range.reverse + range
        when :downup then range + range.reverse
    end

    samples.map{|i|
        i *= increment
        [base - scale * i, i]
    }
end

# letters are defined as a series of pairs of (spin-time, sleep-time)
# with the time in seconds
THIN_A = gradient(15, :updown, 0.2, 0.2/15)
A = gradient(15, :updown)
I = 2,0
L = 1.5,0, [[0.1,0.2]]*9
M = 2,0, gradient(9, :downup), 2,0
N = 1,0, gradient(9, :down), 2,0
U = 1,0, gradient(9, :downup, 0.1, 0.03, 0.1), 1,0
V = 0.5,0, gradient(12, :downup, 0.25, 0.02), 0.5,0
W = 0.5,0, [gradient(12, :downup, 0.25, 0.02)]*2, 0.5,0

[A,I,L,M,N,U,V,W].map{|i|
    # add 2 second pause after each letter
    i + [0,2]
}.flatten.each_slice(2){|x,y|
    # spin, then sleep
    c = Time.now
    1 until Time.now-c > x
    sleep y
}

Words that can be written with the implemented letters can be found with

grep -E '^[aijlmnuvw]+$' /usr/share/dict/words 

Python, on Intel Pentium 4 3.0Ghz, 180 166 145 141 138 bytes

Call with taskset -c 0 python cpu_graph_drawer.py.

taskset is needed to restrict the process to use only one CPU/core (hyperthreading in my case.)

from time import*;c=clock
a=[(3,.8),(3,5),(4,5),(1.3,5),(1.3,0)]
a.extend([(.1,.2)]*10)
for x,y in a:
    t=c()
    while c()-t<x:pass
    sleep(y)

Result isn't that great.

Java 8, 482 characters

Every character in the String means number of threads, that will be utilized. Image taken on Intel Core i3 (2 cores / 4 threads).

import java.util.concurrent.ExecutorService;
import java.util.concurrent.Executors;
public class Mail{
public static void main(String[] args) throws InterruptedException{
Thread.sleep(15000);
for(char c:"123432234321000012343210000444000044441111111".toCharArray()){
ExecutorService executorService = Executors.newScheduledThreadPool(4);
for(int i=1;i<c-48;i++)executorService.execute(()->{while(!Thread.interrupted());});
Thread.sleep(1500);
executorService.shutdownNow();
}}}