Wang2200.org

Here are some figures on how fast Wang BASIC running on a 2200 CPU was for various tasks. The numbers below are for the 2200 A/B/C/S/T CPU; the 2200VP was a much faster machine.

Here are some quick links to the sections on this page.

• Statement Timings
• Memory Write Bandwidth
• Memory Move Bandwidth
• Sieve of Eratosthenes
• Wang 2200VP Performance
• 2200SVP Performance vs Other Machines

Statement Timings (link)

From System 2200 Hardware Pocket Guide, pp.36-37, and from the 2200 service manual, pp. 1-19 through 1-22, except as noted:

CENTRAL PROCESSOR SPECIFICATIONS

Average execution times determined using Random Number Arguments with 13 digits of precision. Speeds are faster in calculations with arguments of less precision.

Average Execution Times within Expressions	Time	Conditions/Remarks
add	0.8 ms
subtract	0.8 ms
multiply	3.9 ms
divide	7.4 ms
square root	46.4 ms
x^y	45.4 ms
e^x	25.3 ms
ln(x)	23.2 ms
integer	0.24 ms
absolute value	0.02 ms
sign	0.25 ms
sine	38.3 ms
cosine	38.9 ms
tangent	78.5 ms
arctangent	72.5 ms
random number [RND(1)]	25 ms	timing by web page author
read/write cycle	1.6 usec
2200 A/B Instructions	Instruction Optimum Time	Conditions/Remarks
REM	0.384 ms
DEFFN'1:RETURN	3.360 ms
SELECT PRINT 005	0.984 ms
SELECT PRINT 005(64)	1.200 ms
GOTO	2.640 ms
IF THEN (NUMERIC)	4.320 ms	condition not met
IF THEN (NUMERIC)	4.840 ms	condition met
IF R1$=HEX THEN NNNN	2.640 ms
IF R1$=A1$ THEN NNNN	3.120 ms
PRINTUSING, 1, A	8.892 ms	display 12 characters
NEXT	4.380 ms	on CRT
LET A=B	3.240 ms
LET B=A(1)	5.280 ms
LET B=A(255)	5.400 ms
LET B=123456	2.880 ms
LET X=1	2.880 ms
LET X=Y	3.190 ms
LET X=Y(3)	5.610 ms	Y is dimensioned to 12
LET X=Y(J)	6.810 ms	J=3
LET X=Y(3,4)	7.010 ms	DIM Y(6,8)
LET X=Y(J,K)	9.210 ms	J=3, K=4
LET X=Y(J+K)	9.310 ms	DIM Y(12) J=3 K=4
LET X=Y(J*K)	10.810 ms	DIM Y(12) J=3 K=4
LET B$='0123456789ABCDEF'	2.460 ms
LET A$=B$	2.700 ms
LET A$=STR(B$,1,16)	5.376 ms
LET A$=STR(B$,8,1)	5.400 ms
LET A$='1'	2.635 ms
LET A$=B$	3.010 ms	DIM A$1,B$1
LET A$=B$	3.900 ms	DIM A$1,B$64
LET A$=B$	4.760 ms	DIM A$64,B$1
LET A$=B$	5.700 ms	DIM A$64,B$64
LET A$=STR(B$,1,1)	7.385 ms
LET STR(A$,1,1)=STR(B$,1,1)	8.310 ms
LET STR(A$,1,64)=STR(B$,1,64)	9.735 ms
LET STR(A$,1,30)=STR(B$,32)	8.160 ms
LET STR(A$,32,1)=STR(B$,32,1)	8.560 ms
LET STR(A$,64,1)=STR(B$,64,1)	8.535 ms
LET STR(A$,64)=STR(B$,64)	6.560 ms
LET STR(A$,1,64)=STR(B$,1,1)	8.435 ms
2200 B INSTRUCTIONS	Time	Conditions/Remarks
B=LEN(A$)	4.140 ms
B=VAL(A$)	3.960 ms
ADD(A$,01)	2.970 ms
ADDC(A$,01)	3.000 ms
ADD(A$,B$)	3.696 ms
AND(A$,FF)	6.240 ms
BIN(A$)=64	2.760 ms
BIN(STR(A$,16,1))=64	5.232 ms
BOOL E(A$,55)	5.040 ms
BOOL F(A$,B$)	6.960 ms
CONVERT A TO A$, (#####)	4.896 ms
CONVERT X*2 TO A$, (+##.##)	8.360 ms
CONVERT X TO STR(A$,3,8), (-#.#####)	7.710 ms
CONVERT A$ TO Z	3.710 ms
CONVERT STR(A$,3,8) TO Z	6.235 ms
CONVERT X TO A$, (####.#####)	5.560 ms
INIT (80)A$	2.400 ms
B=NUM(A$)	4.800 ms	A$=10 digits
B=NUM(A$)	4.560 ms	A$=1 digit
ON B GOTO NNNN	4.800 ms	B=1 GOTO condition
ON B GOTO NNNN	2.520 ms	B=5 fall through condition
PACK(####)A$ FROM A	4.320 ms
PACK(####.####)A$ FROM A	4.704 ms
ROTATE(A$,1)	2.970 ms
ROTATE(A$,7)	3.816 ms
UNPACK(####) A$ TO B	3.960 ms

Memory Write Bandwidth (link)

How fast can the Wang hammer memory? Here are some programs that try to measure that.

In the limit, one nibble of memory can be written each microcycle (1.6 us). In practice, the actual throughput is much lower due to looping overheads.

10 DIM A$(64)64: REM 4KB
20 FOR I=1 TO 256
30 INIT ("$") A$()
40 NEXT I

This program writes 1 MB of memory in 14 seconds, for a fill rate of 73 KB/sec. If the microcode had no overhead, it would be able to execute this in 3.4 seconds, so it is running at 25% efficiency.

Memory Move Bandwidth (link)

Memory moves consist of reading a string of bytes from one location and writing them to another location. This test attempts to measure how fast that can be done.

10 DIM A$(64)64: REM 4KB
20 FOR I=1 TO 256
30 MAT COPY A$()<1,4096> TO A$()<1,4096>
40 NEXT I

This reads 1 MB and writes that 1 MB back on itself in 26 seconds, for a copy rate of 39.4 KB/sec. The microarchitecture can either read a byte or write a nibble each microcycle, so in theory this transfer could have been done in 5.03 seconds. Thus it is actually operating at 20% efficiency in this case.

Sieve Benchmark (link)

Back in September of 1981, Byte magazine presented a few small BASIC programs for testing system performance. One of the more famous ones was the Sieve of Eratosthenes. The benchmark was widely (and rightly) criticized for its shortcomings but it was also widely used. The benchmark uses 8192 flags to indicate which numbers have been proven not prime. Wang BASIC can't address strings that long directly nor can it store 8192 integers for use as flags (because all numbers use 8 byte floating point representation). Instead, Wang BASIC has to use a very awkward mechanism for accessing the Nth byte of a string array.

Without further ado, here is one direct mapping of the standard benchmark into Wang BASIC:

20 REM A$() contains flags indicating which numbers have divisors or not
30 REM Because we skip all even numbers, including 2, the first flag, at
40 REM array offset 1, corresponds to 3. the second flag is 5, etc.
50 DIM A$(128)64:REM prime flags
60 DIM D$(1)1:REM one byte array
70 DIM F$(1)1:REM constant "F"
80 INIT ("T") A$():INIT ("F") F$()
90 C=0:REM number of primes
100 FOR I=1 TO 8192
110 MAT COPY A$()<I,1> TO D$():IF D$(1)="F" THEN 180
120 P=I+I+1
130 IF I+P>8192 THEN 170
140 FOR K=I+P TO 8192 STEP P
150 MAT COPY F$() TO A$()<K,1>
160 NEXT K
170 C=C+1
180 NEXT I
190 PRINT C;"primes"

Wang BASIC running on a 2200T CPU takes about 5:13 (313 seconds).

Wang 2200VP Performance (link)

The second generation CPU that powered the VP and MVP families was much faster than the first generation for a few reasons. First, the microarchitecture of the CPU was simply more powerful. It could readily compute 16 bit binary sums in a single instruction, while the 2200T could process only four bits per instruction. The instruction cycle time was also much faster: most operated in 600 ns on the 2200VP, while each 2200T instruction was 1600 ns. Third, when Wang BASIC was developed originally, priority was placed on making the interpreter small, instead of fast.

Wang Systems Newsletter, No. 4 has timing information for some elementary operations under BASIC-2 running on a 2200VP CPU vs. the same operations under Wang BASIC running on a 2200T CPU.

Comparing BASIC-2 running on the 2200VP CPU, the Memory Write Bandwidth program ran in 2.4 seconds, about 5.8 times faster than Wang BASIC on the 2200T CPU.

Likewise, the Memory Move Bandwidth program ran in 3.2 seconds on the 2200VP, about 8.1 times faster than the 2200T CPU.

Finally, the Sieve benchmark ran in 23.5 seconds, about 13 times faster than the 2200T performance. Taking advantage of the improved functionality of BASIC-2, this time could be reduced further, but that isn't the point of this comparison.

2200SVP Performance vs Other Machines (link)

Page 25 of the December, 1981 issue of Computerworld magazine had a rundown on how the Wang 2200SVP ($14,600 as configured) compared to a number of other contemporary machines. The publisher will not allow me to post a scan of this two page article, so I'm just making a table of the results.

In the table, task C-1 represented a scientific workload, which involved solving 50 linear equations with 50 unknowns. Presumably they used the built-in MAT INV function, vs rolling their own Gauss-Jordan elimination program, as they undoubtedly had to do for some of the other machines.

Task C-3 represented an Accounts Receivable workload, which is a combination of disk access and processing. The time reported in the table for the SVP was using files on floppy to make it fair to the other machines which didn't have hard drives. Running the test on the SVP's hard drive resulted in a performance of 29.1 seconds on the task.

It seems that this article is highly related to this benchmark report published by The Association of Computer Users.