«it turns out today a top consumer (not HEDT) CPU barely matches this pair of CPUs»

Well, the X5670 was a high-end 2P server CPU sold for $1400 (https://ark.intel.com/products/47920/Intel-Xeon-Processor-X5...) so to fairly measure how performance has improved, you should compare it to a similarly priced product, like the AMD EPYC 7401 (https://www.amd.com/en/products/cpu/amd-epyc-7401). And, oh, look at this: the SPECint_rate "h264ref" benchmark subcomponent, which measures multithreaded video encoding performance, shows a 4.2× improvement (472 to 2000):

• X5670: https://www.spec.org/cpu2006/results/res2010q2/cpu2006-20100...

• EPYC 7401: https://www.spec.org/cpu2006/results/res2017q4/cpu2006-20171...

Edit: oops, fixed major screw up. I had linked to SPECint numbers, instead of SPECint_rate. The real improvement is 4.2× not 45×! Far from what Moore's Law predicts (2^(8/1.5)) but still a very notable improvement. I'm sure your friend would be happy to transcode 4× faster...

2x or 4x doesn't matter , you'd need magnitudes of difference to need to think what an OS can or can't do. Which is what the old ten year was about -- and look, 2^(10/1.5) is a bit above 101. Two magnitudes. That matters.

To compare, let's presume (this is a great exaggeration I know but it serves the purpose) each Chrome tab takes equal CPU time. Say, on the current computer the 100th tab makes it unusuably sluggish. At 2x, the previous computer could open 50, even at 4x we are looking at 25. At 100, we are looking at not being able to run Chrome.

Is it possible to buy a Mac Pro with an AMD EPYC 7401 CPU? Surely the fairer comparison would be against the current top-of-the-range in Apple's lineup.

Except the Mac Pro hasn't been updated in a while, partially due to thermal constraints. That isn't related to long-term progress, but an uneven cycle.

I'm not sure I buy the analysis, for many applications the single threaded performance of a 8700k vs an X5670 is doubled, e.g. ~6000 vs ~2600 in geekbench single threaded. You also can't draw conclusions of single threaded performance from multi-core comparisons due to turbo boost.

It's also worth noting that a X5670 cost ~$1400, so you're comparing ~$2800 worth of CPU to ~$350 worth of CPU in something like an 8700k, and that doesn't cover platform cost differences. Another way to look at this is that 8 years after the release of the X5670, you can get better performance for ~10% of the price, that's not bad. If you did plop down ~$2800 today for a CPU or two, you'll have 28+ cores at your disposal.

> If you did plop down ~$2800 today for a CPU or two, you'll have 28+ cores at your disposal.

In just over a weeks time you will be able to get a 32 core / 64 thread AMD Ryzen Threadripper 2990X [1] for around $1,500-$1,800 which is pretty amazing.

[1] https://wccftech.com/amd-ryzen-threadripper-2990x-2970x-2950...

I have a desktop with an intel 3570k bought in 2012, I wanted to build a new gaming desktop with a ryzen 2700x.

The ryzen is only 20% better in single threaded performance. I could easily overclock my cpu to close further the gap. Just need to upgrade my gpu and I'm good for a few more years. These cpus have aged quiet well.

I recently replaced my main desktop PC after nine years and the new CPU benchmarks at only about twice the performance.

Double the performance isn't too bad obviously, but that's about the same timespan as between the 486 and the Pentium III.

I'm right there with you guys.

I made my desktop in 2014 with a Haswell i5-4670K @ 3.4GHz. Overclocked it to 4.3GHz, and I still don't have a reason to upgrade. I really want a Ryzen 2700X also but staying patient. Newer and better hardware will keep coming out so I'm in no rush. (but my next CPU will definitely be AMD)

My current thing right now is compiling the Linux kernel exactly catered to my Haswell machine, and using the latest gcc with "-march=native -O2" optimizations. It might seem minor, but man if she ain't screaming right now. I should mention I have an AMD RX 480 graphics card also. So I'm up-to-date in that world and enjoying the best of team red and blue right now. New drivers, software, kernels, and optimizations keep coming out that make my current hardware faster. But then came Spectre and Meltdown lol.

Just remember,

http://funroll-loops.teurasporsaat.org/

So... always use `-O0 -march=i586`, because using obvious optimizations for your hardware is a total waste of time and will definitely never ever be beneficial? Seriously, I get that it's possible to go off the deep end over optimizing, but that site is way off the deep end in the opposite direction.

I'm sure it was just poking fun. We can't be kernel ricers and get mad when people make fun of it. CFLAGS matter! lol

One thing I will say, compiling your own has many benefits, as all the distribution kernels use Generic x86-64 and miss out on all the Intel/AMD optimizations.

For example, just by switching from "Generic-x86-64" to "Core 2 or newer" in Processor type and features -> Processor family in kernel config, adds these 5 optimizations:

CONFIG_X86_INTERNODE_CACHE_SHIFT=6

CONFIG_X86_L1_CACHE_SHIFT=6

CONFIG_X86_INTEL_USERCOPY=y

CONFIG_X86_USE_PPRO_CHECKSUM=y

CONFIG_X86_P6_NOP=y

You can say it's pulling hairs, but hey, that's what I want. Not many people have the time, know-how, run Linux, etc. to be able to make these optimizations.

7y son's computer is my old, and nearly as old as him! 2600k overclocked to 4.4ghz, 16gb ram, 512gb 840 Pro, NVIDIA 980 on a 1920*1200 U2412M.

My much newer one (six core 6850/950 NVME/1080ti/32gb) really isn't that much faster, GPU aside.

In a similar boat with my old 4770k and a new RX580. I'm kicking myself for the k model though, as I would like to pass my old GPU through to a Windows VM, but the unlocked k model CPUs don't have VT-d, necessary for IOMMU.

Buy a 4th gen second hand CPU? No need to kick yourself, those are cheap now.

For example: https://www.ebay.com/itm/Intel-i7-4770-3-4GHz-Quad-Core-Syst...

This mistake was fixed in Devil's Canyon:

https://ark.intel.com/products/80807/Intel-Core-i7-4790K-Pro...

Intel® Virtualization Technology for Directed I/O (VT-d) ‡ Yes

Can AMD Ryzen and Threadripper run Linux without crashing? What about compiling large source code bases? There were some scary bugs reported on Linux when Ryzen came out.

And that's not even covering motherboard incompatibilities with Linux (e.g. sleep/shutdown not working, WiFi/Bluetooth broken, etc.)

Have those problems been fixed definitively? Or should one stick to Intel for Linux at the moment?

Longtime Linux user who switched to Ryzen 1500x recently from an old Core i7 (2011 model). As far as the processor goes, I have faced no problems w.r.t stability. Paired with an ASRock motherboard, there are no problems with other functionality such as suspend/resume either. Occasionally the Desktop compositor hits a snag freezing my machine and requiring a forced reboot, but I would fault the NVidia drivers for this.

Also, since I use the Arch distro, installing packages from the AUR requires compiling stuff. So far ffmpeg has been among the packages with large source code base that I compile regularly and I've never had issues. I would like to compile the kernel too, but I'm unsure as to whether the difference will be noticeable for my daily usage. On my old machine, the Zen-kernel did improve responsiveness, but on my current one, I'm sticking to the LTS kernel.

Word is that the 2nd generation of Ryzen a.k.a Raven ridge are more stable and worth it if you can afford them (watch out for motherboard compatibility though).

I can't resist crosslink my post from two days ago https://news.ycombinator.com/item?id=17656645

I for myself couldn't suffer a machine where I occasionally need to reset , destroying flow.

I read your post twice and you didn't mention rebooting. Were you rebooting because of docking?

Most of your gripes come from running Linux on a laptop machine. If you are a person who likes to work on powerful desk workstations, Linux behaves great on those. Granted, some things, such as multifunction printers (which are mostly software-driven) don't have support. You do have to be careful when selecting hardware, unfortunately.

No, I just found your off-the-cuff comment absolutely typical of the desktop Linux situation: it's totally stable except when it freezes so hard you need to find a reset button.

> Can AMD Ryzen and Threadripper run Linux without crashing?

Yes, of course!

> What about compiling large source code bases? There were some scary bugs reported on Linux when Ryzen came out.

Are you referring to this? https://www.extremetech.com/computing/254750-amd-replaces-ry...

The later revisions of Ryzen 1000 CPUs have been fixed, you could RMA your CPU if it was affected (I did this myself, AMD's support was excellent). Ryzen 2000 CPUs aren't affected.

> Have those problems been fixed definitively? Or should one stick to Intel for Linux at the moment?

I've never heard of any motherboard incompatibilities. The WiFi chips in some of those AMD motherboards are actually from Intel btw.

I'm running an AMD Ryzen workstation with Linux for over a year now and it works perfectly.

From what I remember those crashes were hardware defects that caused issues under load on any platform. Ryzen 2 hasn't had this issue and my 1600X worked beautifully out of the box on Ubuntu 18.04 except for the lack of sensors.

So your software has no idea whether the hardware is overheating?

Hello, fellow 3570K user - I've also been using one for 6 years now as my main (and gaming) machine. Had to upgrade my GPU once (HD6870 originally, now an RX 470) but I can even play Quake Champions without any problems, in addition to all the games that are a few years old. I've decided to evaluate getting a new machin in 2019, not earlier.

The same pair of video codecs won't magically transcode orders-of-magnitude faster on newer CPUs†, because video codecs—like crypto primitives, compression algorithms, and a few other categories—are designed, in the choices of the types of maths they do, to take advantage of the existence of certain accelerated instructions (e.g. SSE) that exist at the time of their creation. You can't make them take advantage of newer accelerated instructions (e.g. AVX-512) without it being a different algorithm.

In other words: new CPU architectures don't make old algorithms faster†. Instead, new CPU architectures enable newer, faster algorithms.

† Yes, this used to happen, in the 90s, when "new CPU" meant "appreciably higher clock speed." Nobody is making computers faster by increasing clock speeds any more, because higher clock speeds = higher power-draw and higher TDP. CPUs are now advanced by making them work smarter, not harder: accomplishing the same abstract tasks faster by providing hardware that enables newer algorithms to be used for those tasks, while using the same or less energy than the previous implementation would have used.

There is also a ~33% decrease in power usage. So a i7-8700 which also has a similar score now has 65W TDP vs a 95W TDP.

Since moores law is leveling off, programming skill & quality will become more important to extract more performance out of computers. I forsee less javascript and python and more golang, java & rust and other %100 statically typed languages in the future.

And a return to hand-optimized assembly! I knew this day would come, when my skills would be in demand once again!

Like client-server vs local, progress is a cycle.

Dynamic/static typing and execution performance are largely orthogonal. Java manages to be slow while being statically typed.

If your language is inherently heap bound and/or targets bytecode with ad-hoc JIT trickery, no amount of static typing or hard inference is going to help.

My performance hierarchy from fastest to slowest goes in this line. Numbers are based on how much slower it is from the baseline of C / ASM.

C (static typed, static dispatch, manual memory, x1) -> C++ / Objective-C (static typed, dynamic dispatch, manual / refcount memory, x2) -> Java (static typed, dynamic dispatch, GC memory, x3) -> JavaScript (dynamic type, dynamic dispatch, GC memory, x7)

Java is a actually fairly fast for a language with GC. Far faster than equivalent javascript unless you do things like asm.js, which might as well be just a binary runtime.

I feel you are generous with x3 for Java: empirically, a straightforward rewrite to C easily gains you 30x, or more for heavily cache-local datasets.

And it's not setting any speed records among automatic memory management languages either. Swift is well ahead in performance without breaking a sweat. Some Common Lisp implementations generate faster code than Java without any JIT overhead and a billion investment into development.

Swift is a C++/Rust style language vs a Java language. It just looks really pretty, so you might not realize what it's implicitly doing under the hood.

It doesn't use full automatic GC, so of course it is going to be faster than Java.

Swift also has performance gotchas, like it's strings. It's strings are very easy to use in an non-performant way because they are going for full unicode correctness, even when you don't want it.

But it's also a young language with a whole bunch of low hanging fruit in the perf realm.

My perf scale is also more orientated towards business logic / app type things. If your really thinking about your perf and doing a rewrite in fully static C/C++, then I have no idea how much better you can get :)

Is common lisp fully GC?

Swift is distinctly not Rust although it looks a lot like it on the surface. It uses reference counting, just like Perl, Ruby or Python. RC is the most rudimentary form of garbage collection. Also has its problems, like dealing with circular references.

And yes, CL is fully GC. But my point is that's not the culprit. If your problem calls for allocation and management of heap memory, it has to be done one way or another, manually or via GC. GC overhead in that light can really be negligible, and it's other concerns (like scheduling and real time constraints) that typically make its use problematic.

A language's documentation, ease of use, and community are all far more important than CPU performance. Point me to one java project on Github with >1000 stars and I'll give you a hundred Python and JS projects.

High quality TV shows are more popular than oxygen. Unless oxygen runs out.

Yes, finally I will be able to use pen and paper again!

Higher resolution, higher contrast, tactile feedback, cheaper, lighter, thinner

Does not explode, just burns

Dijkstra preferred pen and paper to computers.

You answer to his/her assertion that popularity will shift is that it hasn't shifted yet?

There are heaps of Android Java (or nowadays Kotlin) projects and libraries with well over 1000 stars. I'd guess way more than Python but quite a bit less than JavaScript.

Swift, Kotlin, Rust, Objective-C, C++ and Java have huge communities and code behind them too :)

Design imperatives change too, though. The old machine probably had 3x the power draw. A 4-5x improvement in instructions per joule is nothing to sneeze at. Frankly there are laptops that compete reasonably well with that old space heater.

What's great is how cheap X5600 series CPUs are. I'm upgrading my off lease Dell PowerEdge T410 from dual E5640s with 8GB 1066 MHz RAM to X5670s and 32GB 1333 MHz RAM for less than $200 CAD. I should be also able to sell my old hardware for a few dollars.

The only annoying issue is having to use a piggyback PSU for my graphics card, an RX480. Dell just had to use a proprietary ATX-style connector or else I would have just replaced the stock PSU all together.

What kinds of activities do you need to do video transcoding for?

Vlogging? I hear it's pretty big with the kids these days.

Plex and other home media servers for example.