Hardware review sites are very good at distracting us and making us think that absolute performance is all that matters (hence your comment about "blowing competition away"). And, yes, Intel has the fastest processors. But what matters to most people is performance per dollar. And in fact AMD often beats Intel at this game, especially in the consumer market (sub-$300 range).

Example: I recently built a Linux video transcoding machine to reencode H.264 BluRay movies to smaller H.264 files [1]. A $110 AMD FX-8300 is able to reencode an average BluRay movie around 47-50 fps, while you would need to spend more than twice that amount of money to match this performance with Intel (the performance of the FX-8300 at this tasks falls between a $200 Intel i5-4590 and a $250 Xeon E3-1231 v3).

[1] Using: avconv -i input.mkv -threads auto -s hd720 -c:v libx264 -c:a libmp3lame -sn -b:v 1400k -b:a 128k -ac 2 output.mkv

Both measures have their place. In clusters e.g. rendering farms, then perf/dollar or even perf/watt are paramount. But in a single desktop for a performance-hungry engineer then absolute performance is the number that counts.

Only if he doesn't care at all about money.

> Only if he doesn't care at all about money.

Power is not always the only cost. I work in an industry where you pay per core for software license fees far in excess of the cost of power. There, perf/core is what matters.

Well, even if money does matter, the tradeoff is often more complex than just the one-time hardware cost. If you are a software engineer whose time is expensive, and a slower workstation processor means you wait an extra 5 seconds for a compilation 100 times a day, then the time costs add up quickly.

I would imagine the same is true for e.g. server response latency when that latency correlates to user experience and thus growth/popularity of your service. People might not notice a few ms, but 100ms vs 300ms is definitely visible.

Compiling can be more expensive than 5s/compile, but you should still probably lean toward things like ccache/ramdisks/distcc instead of the latest processor.

I think the real value is new things like TSX/HLE.

Yup, all those things are useful, but they don't solve everything. I've worked on projects where a single change to a header file would lead to a 20-minute rebuild adventure and where individual .cc files would take 10 minutes to compile. Caching and distcc build farms help somewhat but there's still, fundamentally, the critical path of a single file's (re)compilation, and only a faster CPU will help with that.

(And if you say "thin client, distcc, big build farm", then you just need the latest CPUs in the build farm instead.)

Of course this is as much an argument for a real module system in C/C++ as anything else...

(And, I completely agree that TSX and other architectural stuff is totally awesome, too. Perhaps that'll drive adoption more than incremental perf, but I'm just arguing perf is still often really important.)

When workplace is buying - ofcourse not. I have never ever seen a workstation for developers with AMD CPU here in Baltic countries.

This is a real issue. At my company, we always buy the $300-$400 CPUs but they are barely getting faster with each generation. The last time there was a big jump was with Nehalem.

The performance jumps just got more specialized. For instance, introduction of AVX instructions made a huge performance bump for some areas, especially scientific, albeit you needed crazy cooling to run a program using AVX extensively for more than a few seconds. I can remember having around an 80% performance jump with one code I used compiled with vs. without AVX. x86 hardware virtualization extensions were another enormous performance boost. Hardware transactional memory was supposed to be a big (order of magnitude) boost to some things on Haswell but they ended up having to disable it.

Through all this time though, I haven't noticed a massive overall performance boost in a long time. If all I wanted to do was browse the web and read email I'd probably still be using my Core 2 Duo MBP from forever ago. The speed boosts have been piecemeal, one generation improves one thing, the next something else. This is different from Ye Olden Days, I can remember upgrading from a 386 to a Pentium II to a Pentium 4 and each upgrade was so dramatic that it completely redefined what I could do with my computer. The last real performance jump I can remember was the Core Duo upgrade that brought real multicore.

I do kinda miss my beloved old SPARCstation though. Makes me wonder what computing looks like in the alternative dimension where SPARC or one of the other architectures had taken over the PC market like Intel ultimately did.

There have been lots of other performance improvements, but I think they're less visible than other stuff. The CRC instruction that got added in SSE4.2 seems to get fairly wide use in hashes and stuff like btrfs. I think rdrand, aes, and clwb are interesting ones as well. I guess power consumption is kind of another visible gain.

I think the biggest visible change recently though, at least as a software engineer, is the continuously lowering cost, increasing availability of, memory. I can literally store an entire project in memory and not have to worry about any network or disk bottleneck.

Are many people often CPU capped anyway? How would you detect a performance increase unless you're over-utilising the machine.

For example, my work machine has an i7, 8GB of RAM, and an SSD. I can cap IO and RAM, but I am yet to see the i7 max out, heck even over half the cores. I guess what I am saying is: Is CPU really a bottleneck most people encounter day to day anymore?

Confused. Faster CPU means lower latency for all operations. Doesn't have to be maxed-out.

> Faster CPU means lower latency for all operations.

Not really. Per core CPUs have become slower and slower. It become about throughput and not raw per op speed a long time ago. Using that measure you'll want to install a Core Duo in your machine.

Depends entirely what 'it' is. Rendering in parallel - sure. But still most apps are single-threaded.

On the bright side, they're typically getting far more power efficient, and cooling them becomes easier.

For developer boxes, the price of power and cooling do not matter compared to developer productivity, and developer productivity can be in some cases hugely impacted by the speed of their PC (during builds, simulations or renders, etc.)

Intel is still incentivized to provide much higher performance CPUs to the market because they compete with themselves - they want to encourage upgrades.

What you're seeing here is beginning of the end of Moore's Law. Dennard scaling is dead and that will limit performance going forward.

The implications are huge for the computer industry. As Bob Colwell liked to say - imagine if Moore's law ended in the year 2000. We'd never have smartphones.

not sure what you mean by "desktops" - their chips are the fastest CPUs you can get for notebooks, workstations and for many (most?) servers.

> not sure what you mean by "desktops"

I would assume he means "non-mobile". I.e. things not ARM.

I meant that I have no clue and intereset in notebook situation and workstation is a desktop PC. Same for servers - I have no idea what sysadmins use, nor I care. Monitoring shows CPU percentage, it reaches 80% - we tell them to upgrade. :)