Ask HN: Why are modern Intel CPUs not getting a lot more cores?

Because common desktop applications still don't do much with multiple cores, and thus a few fast cores are better (and techniques like TurboBoost are useful, where cores sleep to free the energy budget to allow a single core to clock even higher). And for many things, CPUs have been fast enough for quite a while now.

So the common CPUs have only a few cores, people who really want more get to pay the premium for the gamer models or Xeons, since they don't really have alternatives, and Intel knows that ;)

For servers, you can get CPUs with 4-16 small(er), low-power cores instead of a few large ones. (Atom C2xxx, Xeon-D)

Memory bandwidth is another issue. What people forget in the recent deep learning pissing match between Intel and nvidia is that memory bandwidth limits mean the difference between optimized CPU, GPU and FPGA systems is not that much.

A balance needs to be striken between the consumers of data (OOO execution, multiple issues, HT threads and multiple cores), and the providers of data, the cache subsystem (and memory etc).

Bigger caches that would be required by more cores in order to keep the "usefulness" of the design in balance, would lead to higher power consumption, which a) is a primary criterium of evaluation of the product such as laptops, even before the computing power; b) not justified given typical "office+home" workload, as already mentioned in the answers.

I agree with the "less core but faster processors are what is most useful on desktops".

But I ask, has anyone one or any language made the number of processors/hyperthreads transparent? Ideally I would just create a thread and some load balancing system would distribute it to the correct processor. Am I just behind the times?

1) Nobody writes multi-threaded code

For user applications multi-threading is overkill. It's hard to write, and there are very little benefits the average user will see. Good concurrency isn't even that important on servers. You can trash your cache and interrupt tables but you'll still handle 10million clients per node so nobody cares. Great concurrency is really only practiced in HFT and HPC.

2) Modern hardware/OS's suck at multithreading:

IO interrupts are normally handled on 1 core. While you process data on any core. This causes some non-trivial caching issues. Furthermore putting 2 10GBE nic's on the same core can prevent you from actually using all 20GBE of bandwidth you have because the processor literally isn't fast enough. Then you need to set affinity for epoll thread, etc. etc.

3) Amdahl's Law: https://en.wikipedia.org/wiki/Amdahl%27s_law

Every time you double the number of execution units, even if your calculation gets twice as fast, the improvement factor constantly decreases. Take for example a 60 second calculation that can be fully parallelized

       1 units: 60sec
       2 units: 30sec
       4 units: 15sec
       8 units: 7.5s
       16 unit: 3.75s 
       32 unit: 1.87s

The transitions between 8-16-32 really show you the effect coming into play. Yes the calculation is being done faster. But the overall change from 8 to 32 units saves you ~5.63seconds, while using 4x the resources.

     Also it's yet to be seen what chipset we are getting,
     atm you shouldn't buy AMD not because the CPU's are
     utter garbage, but the chipset is utter trash.

     You are also forgetting the psychological impact that a
     few seconds can make. After 3 seconds of waiting without
     any progress indicator you've generally lost the
     attention of the user and they will do something else in
     the meantime.

All the answers about more cores not being that useful for desktops are true, but there is also the fact that Intel doesn't want to cannibalize its high margin Xeon business by selling high core count cheap desktop CPUs.

Finally for desktops, and especially for mobile, the die real estate is probably put to better use by incorporating a larger integrated GPU.

You can put Xeon in desktop motherboard like with X99 chipset. I myself recently got 22 (44 threads) core one, but for day to day tasks it doesn't make a difference and it is quite expensive. If there was a larger demand then perhaps we could see more of these.

The move over the last few years has been towards mobile, and decreasing power consumption.

So over the last ten years or so, Intel has responded to this realisation by, inter alia, using the extra transistor budget afforded by Moore's Law to move more and more of what was once on the chipset on die.

This is, practically, a much better use of silicon than simply adding more cores.

Thoughts: - Diminishing returns in performance per core because some tasks are truly serial, not parallel. - It costs money - putting stuff on silicon costs money. The only justification for spending more money is that you will make it in revenue. - The PC market imploded. Lenovo has killed the laptop and PC market, and soon will be entering the server market. Expect the end of the server market in the next 5 years or so. There is a lot less money to go around. - Moores 1st law (there are several) is hitting a wall, that is what drove the multi-core approach. They couldn't double the clock speed, so they effectively added to the clock-count. - There is on-die architecture including routers and memory busses. They have capacity limits and tend to be the performance limiters for max temperature. They might be the limiting technology. - Low power is important, as is low heat. More cores (aka more ALU's means more heat and higher power use. - The more components, the easier for something to go wrong. - Gartner says the next big money is in IoT, not massively parallel on the PC. (It seems google beat Intel to a machine-learning ASIC, and deep learning seems to be a decent market right now - but because Gartner didn't say it the Intel leadership can't hear about it).

Caches.

Xeons and the like dedicate much of their transistor budgets towards larger caches.

If you take that budget for caches and put it towards more execution units, you get the equivalent of Xeon Phi which is a lot of Pentiums built on today's fabrication processes.

I am the proud owner of a Xeon 12/24 3.0Ghz. It's the best one!

Some people like hot-rod cars.

I don't use a GPU like normal people do. Graphics are great for parallel execution. (Obviously)

If everybody owns a Xeon 12/24 then you can make real-time software that requires it. There's not a chance any software will be written until most people have it.

They are, but in the GPU.

GPUs handle the tasks that can be easily parallelized better than CPUs. Many supercomputers are now made from GPUs.

I have a theory that we will gradually favour business and societal forms that can be easily parallelized, i.e. on GPU.

BTW intel's growth focus is on ASICs (application-specific integrated circuits) - effectively, code pushed down to silicon. https://news.ycombinator.com/item?id=11287511 And will integrate them in future Xeons http://www.pcworld.com/article/2921832/intel-looking-to-boos...

They actually are doing more cores... For now, they integrated the Phi chip into a real cpu. Give it another 2-3 years then this tech will come to consumers (very likely I think).

http://www.anandtech.com/show/10553/asrock-rack-launches-2u4...

And I completely disagree with that consumers don't need many cores... That would be paramount to saying that consumers don't need GPUs. One of Intel's core arguments for bringing Phi to the masses is realtime raytracing (something for which normal GPUs suck bad). Also NVidia is working on a Phi-kind-of-thing as their next generation GPU. So we are getting there, but it was a long journey from 2006.

AMD tried and failed. Multiple cores is what the FX is all about. What's the one type of application that drives processor sales on the desktop? Right, games. What's one of many types of applications that do profit more from fewer stronger cores than from more slower ones? Right again, games.

That said, over the last 5 years processors did get beefier. They just did not got more cores. The i5-2500K (from 2011) is slower than the i5-6600K, both have 4 cores though. But if we go back 10 years, we are looking at the Core 2 Duo, those had only two cores.

And games are also changing. Not too long ago you could still play games (like Fallout 3) with a single core. New games (like Fallout 4) rely on the processor having multiple cores, some don't even start with a dual core (Hyperthreading is needed then and saves the i3). And games supporting DX12 (and probably Vulcan) work better on the FX processors than those using DX11. More cores become more useful now, and processors will have to add more of them also in the consumer desktop area soon.

This is my current understanding:

Mostly down to poor support for multiple cores in software, a lot of which have some kind of legacy heritage, and some tasks not translating well to running in parallel. Take video editing, for example - generally a single task the requires straight-line speed on one core, while the rest sit idle. Games are the most intensive for multi-core CPUs, but gamers usually buy top-end components, up to and including Xeons.

If anything, we tend to have a glut of CPU performance for day to day tasks. Even low-grade CPUs like Celerons are capable daily-use machines; some tech people (a friend included) see no reason to go beyond an i3. The multi-core paradigm shift around 2005-6 also massively increased CPU efficiency over anything that came before, reducing the need for existing applications to be reworked to exploit the extra cores. Intensive Javascript on sites can really drag your browser down, but even Chrome only runs one process per tab, and few people work with more than one tab at once. Multi-cores therefore tend to take care of background tasks, with only one core at a time doing all the heavy lifting. For one of the best examples of multi-core not being the ideal solution, one need only look at the Playstation 3. The 7-core (I think?) CPU is extremely powerful, but very, very difficult to program for, creating real headaches for developers trying to exploit its potential.

Multi-core might have drastically improved power efficiency, but it hasn't completely solved the heat generation problem. With the Prescott P4, Intel realised the core was generating so much heat it presented a serious problem to cool it. Cramming 10 of those on a single die would probably cause a China Syndrome. Adding cores allowed manufacturers to clock each core lower and still crunch a similar amount of numbers while reducing the heat output, but this sacrifices straight-line speed for single tasks. Ergo, AMD and Intel have to compromise for their consumer-grade products - enough straight-line performance that it doesn't feel slow for the current task, and enough cores to keep the throughput high. Looking at the current Xeon E7 v4 series, for example, increasing the core count above 4 results in a decrease in clock speed, likely entirely down to keeping the heat under control.

Add into this, laptop sales overtook desktops sometime around 2010, which require lower power consumption above all else. You could put a 16-core CPU in a laptop, but to keep power consumption realistic each core would probably end up clocked slower than a Pentium II. Further to this, a second shift occurred a year or two later where tablet and phone sales outstripped laptops. Mobile devices based on ARM CPUs are very power-efficient for all but the most demanding tasks. I know several people who've ditched laptops altogether and use Android or iPads exclusively. Essentially, current consumers have no need for the higher number-crunching capability of a desktop, which in turn means the traditional market for Intel and AMD has shrunk noticeably. ARM CPUs are starting to come standard with 8 cores, and it's difficult to find a current device with less than 4, all helped by the ARM architecture's power efficiency. Because they run an OS designed specifically for low power, not for general purpose, the cores can be clocked lower.

A lot of factors, but as noted, Intel and AMD are still improving their CPUs, just not by cramming more cores into the same die.

If the desktop has four cores now, then it has double the cores of a few years ago when two was standard and that was double the number of cores 'anyone' had ten years ago: desktop Conroe chips started shipping in 2007. Which means the number of cores has quadrupled in less than ten years.