CAS latency hasn't been measured directly in nanoseconds for some time now. It is now measured in clock cycles.

Yah, so to compare two different sticks of RAM you have to multiply the time/cycle by the number of cycles. Which gives you (wait for it....) time!

Which the parent did, to point out that all these "new" memory technologies haven't been decreasing the RAM latency much at all. RAM latency is still a _VERY_ important part of overall execution performance. Particularly for single threaded operations reading RAM in unpredictable manners. Cache misses are overwhelmingly the single largest optimization variable for modern applications.

Yes, and for a desktop machine probably 90% of what I do is limited by the single thread performance .Hence why I haven't upgraded in a while myself.

So, I do welcome faster machines, what I don't welcome is the fact that the vast majority of machines being sold today are actually _SLOWER_ than what was available a few years ago.

This happened at work, we replaced a couple of older machines that cost a fortune with a couple newer far less expensive one and the performance was actually worse.

There are LOTS of applications outside of gaming where more speed is appreciated.

But a lot of those applications are also runnable in networked clusters. I stopped compiling code on my desktop probably 15 years ago and haven't looked back. Buying a single machine with 32-cores and a super fast RAID shared between a dozen or so developers both improves individual compile times and saves a bunch of money over buying a bunch of faster desktop machines for everyone. Edit the code locally, save to a network share, compile remotely.

Same thing for VMs, ray tracing, transcoding, scientific computing, etc,etc.

There are still a few "workstation" level applications but its questionable whether the i7 line is more appropriate in those circumstances than just buying multisocket xeon configurations (which provide even more cores and memory bandwidth).

All that said, don't get me wrong, I really like my single threaded performance which is where I think people have been sort of missing the boat for the desktop. AKA I would pick a dual core machine over a 16 core one if the cores were even 2x as fast at single threaded operations.

Its not even about optimizing the code, its making choices that from the beginning cannot result in faster code. People like to focus on the overhead of JITs, GC's, and hidden object copies, etc, in many "modern" languages, but frankly while they have an effect, the mindset they bring is a worse problem.

Modern machines can lose a lot of performance with poor memory placement/allocation in a NUMA configuration, doing cache line ping ponging, and on an on. Things that are simply not controllable if your language cannot even guarantee a consistent location for the data in question.

Lets not even talk about the horrors of HTML/javascript/CSS/AJAX/etc.

Now, all that said, a huge percentage of applications are going to be "fast enough" if they were written in bash, running in an emulated x86, in javascript, in firefox on a $50 tablet. Simply because even the slowest thing available today has 100x the performance of the machines 15 years ago which somehow managed to be useful without storing all their data in the "cloud" for the NSA to peruse.

The problem is that trying to fab a processor without a foundry seems to be a big disadvantage. For IBMs mainframe business its probably not a critical problem as they aren't as performance intensive.

But for something like POWER which directly competes with x86 I suspect that they will have an even harder time selling their processors if they follow the AMD (or sparc, mips, etc) route. The ARM vendors seem to do fine without foundries, but the best performing ones seem to regularly come from companies that actually have their own in house fabs.

If the workload naturally fits into more nodes of smaller size, it frequently makes sense to opt for the higher node count. There is of course different break points depending on judgement calls, but most places seem to think of two sockets as about the sweet spot.

That describes the problem I work on, the throughput scales pretty nice as the machine size grows, but the costs of the larger machines grow much faster than their performance. So, it is far more cost effective to ship a few 2 socket machine with higher clocked processors than try to cram it all into one or two large machines.

But! While the throughput of the larger machines scales, their latency does not. In fact for the latency sensitive portions of our application we are far better off with smaller machines with faster ram, faster clocked CPU's , and closer IO busses. There are points where its actually impossible to buy better latency than we get for just a couple grand in our mid-range machine.

The pricing I saw a couple months ago didn't even approach what we are paying for our machines. Sure the machines in question _may_ have been ~30% faster but they cost literally 4x as much.

For customers buying larger Intel platform machines (4 sockets or more) the power8's are possibly competitive, but compared with the mid-range dual socket machines its wasn't even close.

Maybe IBM has adjusted the pricing since then, they keep telling me its going to be better than x86, but I have yet to see that for our use cases. Plus, I suspect that Intel will adjust their pricing in a few months if POWER is actually competitive. They have a habit of doing that. Just taking back the 4 socket "tax" they added a few years ago when AMD stopped being competitive will probably blow a hole in IBM's model.

If you go to IBM conferences you will find a fair amount of talk on this very topic by 3rd party vendors. There are probably a dozen vendors that want to provide AS400/iSeries cloud instances, but IBM won't let them because it violates the terms of the IBM i license which is tied to a hardware instance.

Plus, the whole software ecosystem piggybacks on the same idea, (often based on machine capabilities). This means that even if you can rent an iSeries for an hour its likely your software vendor won't license you their application.

So, while it is entirely possible, IBM seems to be dragging their feet on the license issues, and the vendors seems to be in a chicken/egg situation.

Not sure what the GP actually intended, but I'm convinced the fact that the kernel and a few thousand drivers all simultaneously have to be bug free for any given "release" is a serious problem. Should your hardware experience a driver problem you get to roll the dice again and hope the next version fixes your problem without breaking something else. Good luck, especially if you have a couple dozen different hardware configurations to contend with, especially if any of them are not x86.

Its futile, the drivers and the kernel should be separate and there should be a stable API, if not a full blown ABI for them. Linux has been evolving for ~20 years now its probably time to start trying to maintain some kind of actual kernel mode API. That way the _USER_ can pick and choose the kernel and any given set of drivers independently from one another. If kernel X happens to be "good" but you need a driver newer than that kernel you shouldn't have to upgrade to the latest buggy kernel just to get a driver for a more recent piece of hardware.

Android avoids this problem because the OEM spends time assuring that the driver set for their device is working/stable before shipping the device. Then rarely are they ever upgraded for anything other than bug fixes.

Hmmm, compare the options available via:

http://www.eightforums.com/tut...

with the options available in windows7 (and previous) in the "window color and appearance" dialog.

There are probably 4x the number of options in win7, including font selection/size changes that are simply not available on win8.

If this device is anything like the dell venue pro with 32GB, it works out to something like 17GB usable when you turn the device on, but by the time windows update runs its going to be less than 10GB free.

Lots of discussion about this on the internet, for example:

http://en.community.dell.com/s...

My only complaints are that Chrome actually performs quite poorly on sites with heavy AJAX (specifically Yahoo Mail), and that Flash is better off left not installed (darn). But Firefox appears to be much better optimized for low-end hardware, so I just use Firefox with no Flash.

On low end hardware I've been installing qupzilla. Its a webkit based browser minus much of the junk. It actually runs pretty good on windows2000 era machines, something that cannot be said for chrome/firefox at this point.

2GB per process - closer to 1.75 in practice.
3.25GB total usable.

First the 2GB limit is not fixed, it can be increased to 3GB. See http://msdn.microsoft.com/en-u...

The latter is a license restriction, not an OS one. There are 32-bit windows machines that can use 64GB of RAM and by default PAE is actually enabled by default on all versions of windows since XPSP2 (or was it SP3?) when DEP became the default as well.

http://msdn.microsoft.com/en-u...

So, if MS cared they could release a 32-bit windows professional that supports more RAM, they just choose not to. The real question is about drivers failing to use the HAL provided API's for DMA and such. While there were a few driver issues in the early 2k days most of them actually tended to work.

Classic shell gives you a start menu, and some of the customizations back, but your still up a creek if you want to say, customize the coloring of your window controls at a fine granularity, or a large number of other less noticeable UI changes.

I am, because after a single install for classic shell, you get a Windows that functions almost identically to a Windows 7 box

I use classic shell on the win 8/8.1, and frankly while its pretty tweak-able, it can't actually replicate the way I ran windows 7 nor the way I run windows XP. A big part of the problem in 8 is the removal of the ability to control with fine granularity pieces of the window decoration.

There are a lot of other issues, but win8 is just another half ass interface from microsoft layered on top of the half dozen other application UI paradigms they have tried to thrust on people over the past two decades. It doesn't take long before you notice that there are actually more than just metro/modern applications and applications with ribbons in win8. In fact last I counted there were 6 different application paradigms in the applications shipped with the OS going all the way back to windows 3.0. That is even before you install any 3rd party software.