Buy a NeXT Blasting P4 Hardware!

The following described and depicted system was a somehow actual (8/2005) full compatible NeXT based system on Intel sold at NiKo-Internetpräsenz. The company brought down in 2007, but feel free to drop me a mail, I can assist for sure, don't hesitate to ask for details.

Just in advance, take also a closer look under P4-NeXT, the shown system is faster than the depicted one here, but drivers etc. remain the same. Most of the components are brand new, or still available. If not, it is noted as used. When asking for a system, please ask also for availability of single components, I possibly can help. Please note that the operating system was not part of the distribution, but I will send you the necessary divers to be set if needed, or surge the NeXT-Step package.
Enough said, lets go into details. The system makes use of a socket 478 based industry main board, which in turn utilizes a Pentium IV with 3GHz and more. Several components are changed for better reliability. The board provides three different kinds of busses, i.e. ISA, PCI, and AGP up to 8x. For a NeXT the usual components come into play.

The depicted system has some additional stuff inside, just to demonstrate, what is possible at all. For other operating systems, well they possibly exist... The NeXT runs with the 2GB main memory, but can only use 958MB, what is close to 1GB. In the next pictures and tables you will see some typical NeXT benchmarks, to give an idea of the capabilities of the system. You can easily compare this e.g. to the other server system shown on the other page, or an original NeXT.

So lets start with a simple Tensor test under NeXT-Step:

...and a simple Tensor test under OpenStep:
Keep in mind that the Tensor disc times are made under NeXT-Step with the 9GB disc. This IBM one is old, used, very noisy, hot, and compared to the 36GB Seagate Cheetah's real slow... Impressive to see more than 320 times the floating point power compared to an original NeXT, when talking about trigonometrics. The noticeable differences could be the usual spread, but under several tests you must notice that OpenStep is always a bit slower than NeXT-Step! Anyway, the main difference can be seen at the Disk I/O.

OK, before going into more benchmark details some pictures. The case can be in NeXT-Color, what is not very impressive. You will also see that this case made into the P4-NeXT.

So what else you may ask, well a usual PC, or not?

Hmm, not really. The system comes along with a very powerful switching power supply, not one of these cheap ones, holding from 12PM to noon. At the bottom you can see the mounted four EIDE drives with 0,64TBytes, and the noisy 9GB SCSI above. From the front a 120mm blower will cool down the drives to a minimum without being noisy. On the left side at the bottom the usual cards as described above. From top to bottom, MATROX graphic adapter, Adaptec SCSI adapter, 3Com ethernet card, digital TV receiver, spare place for ISA or PCI, SoundBlaster 16, US-Robotics modem.
Have you seen the spare space above the disc drives with the silent blower on the right? Well here you will find the OpenStep drives:

...and mounted into the case, the system is now complete:
You say, spare me the PC details, what's up? OK, the benchmarks. Start with the memory performance first. Take a look at the sustained main memory bandwidth, impressive 900MByte per second under NeXT-Step! And can you see the memory bandwidth of the NeXT, it is the red line, not the pink one. Even if the AMD server with the Athlon looks quite smooth and powerful, the Pentium kicks ass and even blast the IBM Mainframe from 2002 not only with small block sizes.

Having the Disk I/O from Tensor in mind you may want to see how fast the drives really are, well the 9GB NeXT-Step drive on left, the 36GB Cheetah for OpenStep on the right:

Please note, that the window for the Read value can not even hold the number for the Cheetah drive on the right! And remember the small performance difference I stated at the beginning? See NXFactor doing the job under NeXT-Step and OpenStep, and keep in mind, it is exactly the same executable, and that the Dhrystone value is incorrect (I had to tune the loops in the binary, otherwise the machine is too fast, multiply the value with 10):

(NeXT-Step on the left, and OpenStep on the right, as before.)

Still not enough and need more benchmarks? OK, the last comparisons are the well known ByteBench's. First, the base is NeXT-Step, the other OpenStep.

Source File for target machine results: pgms/ByteBenchP4_OpenStep.base
pgms/index.base /tmp/4334.dat

                     INDEX VALUES            
TEST                                        BASELINE     RESULT      INDEX

Dhrystone 2 without register variables     4135425.5  3488795.6        0.8
Dhrystone 2 using register variables       4082058.6  3463669.0        0.8
Arithmetic Test (type = arithoh)          15766610.4 15767198.6        1.0
Arithmetic Test (type = register)           457671.5   438670.1        1.0
Arithmetic Test (type = short)              487874.4   479043.1        1.0
Arithmetic Test (type = int)                457440.5   440107.5        1.0
Arithmetic Test (type = long)               457743.8   440282.9        1.0
Arithmetic Test (type = float)              767524.0   746378.5        1.0
Arithmetic Test (type = double)             767439.4   745535.9        1.0
System Call Overhead Test                   288525.4   300359.4        1.0
Pipe Throughput Test                        443590.5   436887.9        1.0
Pipe-based Context Switching Test            89539.0    92415.0        1.0
Process Creation Test                         2975.4     2294.0        0.8
Execl Throughput Test                         1546.2      372.3        0.2
File Read  (10 seconds)                     835653.0   806724.0        1.0
File Write (10 seconds)                       4566.0    17697.0        3.9
File Copy  (10 seconds)                       4918.0    15305.0        3.1
File Read  (30 seconds)                     803481.0   809081.0        1.0
File Write (30 seconds)                       4275.0    17833.0        4.2
File Copy  (30 seconds)                       4533.0    11783.0        2.6
C Compiler Test                                315.9     1063.0        3.4
Shell scripts (1 concurrent)                   389.3     1752.3        4.5
Shell scripts (2 concurrent)                   181.0      911.7        5.0
Shell scripts (4 concurrent)                    91.3      445.7        4.9
Shell scripts (8 concurrent)                    47.0      221.3        4.7
Dc: sqrt(2) to 99 decimal places             45611.4    19332.7        0.4
Recursion Test--Tower of Hanoi               70217.0    50935.1        0.7
     SUM of 27 items                                                  51.9
     AVERAGE                                                           1.9

As system inherent, the disk i/o differs due to the fact of different disc drives. These would be most probably the same with same drives, but as mentioned above, OpenStep is always slower than NeXT-Step, except the Shell/Task switch. BTW, the compiler was 2.7.2 under NeXT-Step (opensource) and 2.7.2 under OpenStep (OpenStep).

To put the results in respect to other architectures, two last ByteBench's. The first is OpenStep vs. original NeXT:

Source File for target machine results: pgms/ByteBenchNeXTstation.base
pgms/index.base /tmp/5075.dat

                     INDEX VALUES            
TEST                                        BASELINE     RESULT      INDEX

Dhrystone 2 without register variables     3488795.6    25789.2        0.0
Dhrystone 2 using register variables       3463669.0    25811.2        0.0
Arithmetic Test (type = arithoh)          15767198.6    45990.9        0.0
Arithmetic Test (type = register)           438670.1     3026.8        0.0
Arithmetic Test (type = short)              479043.1     3400.9        0.0
Arithmetic Test (type = int)                440107.5     3017.9        0.0
Arithmetic Test (type = long)               440282.9     3012.2        0.0
Arithmetic Test (type = float)              746378.5     2553.3        0.0
Arithmetic Test (type = double)             745535.9     2554.9        0.0
System Call Overhead Test                   300359.4     6277.3        0.0
Pipe Throughput Test                        436887.9     2856.9        0.0
Pipe-based Context Switching Test            92415.0     1211.9        0.0
Process Creation Test                         2294.0       39.0        0.0
Execl Throughput Test                          372.3       29.7        0.1
File Read  (10 seconds)                     806724.0    15742.0        0.0
File Write (10 seconds)                      17697.0      428.0        0.0
File Copy  (10 seconds)                      15305.0      372.0        0.0
File Read  (30 seconds)                     809081.0    16394.0        0.0
File Write (30 seconds)                      17833.0      354.0        0.0
File Copy  (30 seconds)                      11783.0      218.0        0.0
C Compiler Test                               1063.0       25.7        0.0
Shell scripts (1 concurrent)                  1752.3       31.0        0.0
Shell scripts (2 concurrent)                   911.7       16.9        0.0
Shell scripts (4 concurrent)                   445.7        8.6        0.0
Shell scripts (8 concurrent)                   221.3        3.8        0.0
Dc: sqrt(2) to 99 decimal places             19332.7      776.0        0.0
Recursion Test--Tower of Hanoi               50935.1      342.2        0.0
     SUM of 27 items                                                   0.5
     AVERAGE                                                           0.0

You say this is not fair, more than a decade between both? OK, want to compare this NeXT dream against a 2002 IBM Mainframe? Here is the last comparison, file write not included, because the Mainframe had Linux on it with a RAM-disk, so file write is no real file write and therefore skipped:

Source File for target machine results: pgms/ByteBenchS390G6.base
pgms/index.base /tmp/5371.dat

                     INDEX VALUES            
TEST                                        BASELINE     RESULT      INDEX

Dhrystone 2 without register variables     3488795.6   547808.4        0.2
Dhrystone 2 using register variables       3463669.0   547026.2        0.2
Arithmetic Test (type = arithoh)          15767198.6  1293552.8        0.1
Arithmetic Test (type = register)           438670.1   108506.9        0.2
Arithmetic Test (type = short)              479043.1   106554.3        0.2
Arithmetic Test (type = int)                440107.5   108525.2        0.2
Arithmetic Test (type = long)               440282.9   108948.0        0.2
Arithmetic Test (type = float)              746378.5    35100.6        0.0
Arithmetic Test (type = double)             745535.9    44878.0        0.1
System Call Overhead Test                   300359.4   274911.2        0.9
Pipe Throughput Test                        436887.9   279618.8        0.6
Pipe-based Context Switching Test            92415.0    74697.3        0.8
Process Creation Test                         2294.0     3067.7        1.3
Execl Throughput Test                          372.3      562.3        1.5
File Read  (10 seconds)                     806724.0   731069.0        0.9
File Copy  (10 seconds)                      15305.0    11234.0        0.7
File Read  (30 seconds)                     809081.0   732527.0        0.9
File Copy  (30 seconds)                      11783.0     7001.0        0.6
C Compiler Test                               1063.0      358.0        0.3
Shell scripts (1 concurrent)                  1752.3      814.8        0.5
Shell scripts (2 concurrent)                   911.7      478.2        0.5
Shell scripts (4 concurrent)                   445.7      243.9        0.5
Shell scripts (8 concurrent)                   221.3      137.6        0.6
Recursion Test--Tower of Hanoi               50935.1     6929.1        0.1
     SUM of 24 items                                                  12.5
     AVERAGE                                                           0.5
So this NeXT P4 outperforms a 2002 IBM Mainframe by 2 times under this ByteBench!

If you want to have more details, send a mail to:
I will answer ASAP, usually at the weekend.

© Dr. Henry Koplien