Browser wars again? The latest from Redmond

[Zero999]

Yes, I would say DOS and Windows 3.x/9x was MS's worst software: very buggy, no security and extremely unreliable.

What planet is that?
DOS and Windows 3.11 were stupidly reliable. Production systems based on them would stay operating until the power supply failed.

If they were run in isolation, with carefully written third party software, then fine, they could be reliable.

The problem was one program crashing could easily crash the entire OS which would eventually crash anyway, if programs are being opened, closed and opened again, as in an ordinary desktop environment, because the memory management was so poor, the RAM would eventually fill up, resulting in information being swapped to disk.

[rsjsouza]

Yes, I would say DOS and Windows 3.x/9x was MS's worst software: very buggy, no security and extremely unreliable.

What planet is that?
DOS and Windows 3.11 were stupidly reliable. Production systems based on them would stay operating until the power supply failed.

If they were run in isolation, with carefully written third party software, then fine, they could be reliable.

The problem was one program crashing could easily crash the entire OS which would eventually crash anyway, if programs are being opened, closed and opened again, as in an ordinary desktop environment, because the memory management was so poor, the RAM would eventually fill up, resulting in information being swapped to disk.

That is what I remember as well. Don't forget that Windows 3.x was what made Ctrl+Alt+Del become infamous. Windows 3.x releases were extremely unreliable if they were at the mercy of user intervention - one misstep and you could easily cause the infamous GPF (General Protection Fault). If the keyboard buffer became full or the mouse stopped responding... the recovery was impossible. Hard drives were also much less reliable and the filesystem did not provide for much recovery - a cluster or sector failure or any disk-intensive application were real threats for the endurance of any system. Almost all reliable stations I saw in production environments that ran Windows had one or two "features": either the keyboard/mouse were removed/locked or any games/network connectivity were removed (including the famous Trumpet Winsock when internet started to become widespread). Windows NT 3.x was better but very resource hungry.

DOS-based systems, on the other hand, were quite reliable as I recall.

[james_s]

It's largely just a matter of complexity. DOS was very simple, the OS really didn't do much, there wasn't a bunch of stuff running in the background, it was just the command interpreter. Most of the software back then was fairly simple as well although I did have plenty of DOS programs crash. Modern computers and operating systems are orders of magnitude more complex so there are many, MANY more ways something can go wrong. Even so, ever since Win95 most crashes have been down to bad drivers or hardware faults. Win95 did a pretty good job of isolating programs from each other so it was difficult for a program to take down the whole system.

I've had some PCs that were more stable than others but the ones that were less stable I strongly suspect was due to hardware faults or unlucky combinations of components and drivers.

[rstofer]

Of course DOS was simple, it was based on CP/M and that was based on something else (I don't know what).  These OSs did very little.  There was a need for expanded memory and that's where the problem started.  No real standards, pretty much the wild west.  In the earliest days of CP/M, 64K bytes was a huge machine.

But all these 'complaints' are really just looking back at growing pains.  Since I started with mainframes in the early '70s, we've come a very long way.  At one point Control Data owned the scientific market with a machine that did 1 MFLOPs!  Maybe the biggest machine could get to 2 MFLOPs.  My tower probably does close to 500 MFLOPs!  I have no idea what the GPU can do.

We've come a long way in a fairly short period of time.  Imagine what the next 50 years will bring!

[oPossum]

My tower probably does close to 500 MFLOPs!

You are still using a Pentium III ?

My ~8 year old computer will do 50+ GFLOPS (CPUs only).

[Syntax_Error]

How? Even if clocked at 5GHz, that would be 10 floating point operations per cycle.

[rstofer]

Yup, once again, I slipped a couple of digits!  The I7-7700k is good for 132 GFLOPS.  No, I have no idea how that works but the pipeline is 14-19 stages and multiple execution units probably helps.

https://techgage.com/article/intel-kaby-lake-core-i7-7700k-performance-z270-chipset-overview/4/
https://en.wikichip.org/wiki/intel/microarchitectures/kaby_lake

I don't see the site where I got that lowball number!  I like the bigger number!

In any event, it's a pretty quick processor even with Win 10.  32GB of 4 GHz RAM and a pretty high end SSD help with that.

[NiHaoMike]

Modern GPUs are in the TFLOPS range - the latest 2080Ti about 11.75 TFLOPS for single precision (or well over 100 TFLOPS for Tensor operations - haven't looked into the details) and even my now considered old 970 does 3.5 TFLOPS. With a fine tuned 1080p-to-4K upscaler running on it, I see utilization averaging about 18% - that's over 600 GFLOPS just to display some pretty pictures on the screen!

[Berni]

Yup, once again, I slipped a couple of digits!  The I7-7700k is good for 132 GFLOPS.  No, I have no idea how that works but the pipeline is 14-19 stages and multiple execution units probably helps.

https://techgage.com/article/intel-kaby-lake-core-i7-7700k-performance-z270-chipset-overview/4/
https://en.wikichip.org/wiki/intel/microarchitectures/kaby_lake

I don't see the site where I got that lowball number!  I like the bigger number!

In any event, it's a pretty quick processor even with Win 10.  32GB of 4 GHz RAM and a pretty high end SSD help with that.

Its done using SIMD instructions.

If you pack the data in memory just right you can use one instruction to multiply or add 4 or 8 floating point numbers in one shot. Additionally there is often a separate execution unit for add and multiply to allow both to happen simultaneously and is often combined into one MAC (Multiply Aculumate) instruction so they x2 the FLOPS for that instruction because its doing whats considered two operations. This can allow a single modern Intel core to do somewhere between 4 to 32 float operations per instruction cycle (Depending on what instructions it supports and if its single or double precision). Usually each core has its own FPU so you can then simply multiply this number by the number of cores you have. So 8 core CPU can do a total of 256 float operations per cycle and that sounds pretty insane!

The support for these special parallel math instructions is called SSE, SSE2, AVX, AVX2 and AVX512.

As for GPUs they have even higher numbers because they are optimized for float number crunching performance. Same SIMD instruction idea applies, but the cores in a GPU are very different. Each core is so dumb that its more like a state machine who's only job is to shovel data trough the FPU as fast as possible. This makes it terrible at running any sort of complex branching program like the main CPU is built to do, but it makes the core take up very few transistors. As a result the guys at Nvidia and AMD can really mash that copy paste button and put >1000 of these cores inside a single GPU. Because of this GPUs are insanely fast at any task that can be broken down in to a lot of very repetitive float math and this happens to be graphics, video and AI stuff.