5 ms·

The OP's advice, if applied in CPU industry, would be disastrous. Modern desktop/server CPUs are incredible complex... in order to drive maximum performance. Pi

by whack 8y ago

The OP's advice, if applied in CPU industry, would be disastrous. Modern desktop/server CPUs are incredible complex... in order to drive maximum performance. Pipelining, OOO execution, branch prediction and speculative execution: these are all features that introduce tremendous amount of architectural and design complexity. In many cases, they also harm correctness, because they can lead to functional and security bugs. And yet, if you try to compete with Intel with a CPU missing the above optimizations, you will get absolutely creamed in the marketplace. No one, not even those touting the importance of simplicity and correctness, will buy what you're selling. Today's free market is too complex for these overly simple rules. Choosing between simplicity, correctness and performance, is a complex tradeoff that needs to be made on a case-by-case basis. Trying to find shortcuts to avoid these analyses may feel liberating... but you're ultimately only shooting yourself in the foot.

Sir_Substance 8y ago
The OPs advice would probably be dissapointing if applied to making a curry too. It's fortunate he regularly used the word software throughout the blog, really.
Sir_Cmpwn 8y ago
The attitude of the CPU industry in this regard led to some recent well-publicized, very bad, and nigh-unfixable security vulnerabilities, as you might have heard.
- whack 8y ago
  And yet, I don't see you or anyone else committing to buy ultra-simple non-pipelined non-OOO desktop/server CPUs. If you insist on only hiring chauffeurs who drive at 100mph, you can hardly complain when they get into a few accidents.
  - Sir_Cmpwn 8y ago
    The main problem is that most software leans on those misfeatures as a crutch to excuse heaping layers of abstractions. Unfortunately this problem comes from several places, so it's not as easily fixed. That being said, consider me lined up to buy one of these CPUs.
  - a1369209993 8y ago
    > And yet, I don't see you or anyone else committing to buy ultra-simple non-pipelined non-OOO desktop/server CPUs. If you can find a CPU that has the same number of non-cache[1] transistors as a Intel/AMD chip, but spends them on a larger number of simple (and preferably independent/non-hyperthreaded) cores, rather than squandering them on speculative execution and ten thousand obscure model specific registers, I would absolutely buy several of them. 1: and similar amounts of cache, of course.
    - Const-me 8y ago
      Intel makes them and you can buy them today, with up to 72 Atom CPU cores, e.g. (1) https://ark.intel.com/products/95830/Intel-Xeon-Phi-Processor-7290-16GB-1_50-GHz-72-core https://ark.intel.com/products/95830/Intel-Xeon-Phi-Processo... Very niche products. For massively parallel number crunching, GPUs are much better in both performance/watt and performance/dollar. That Xeon Phi 7290 delivers up to 3.45TFlops, costs $3200, and consumes 245W. Compare with GeForce 1080Ti 10.6 TFlops, $700, same 250W. For general purpose software they don’t work particularly well either. Most IO interfaces is serial, SATA, PCI-X, they have very few wires going to CPU. If you’re IO bound and you don’t have enough single-thread performance you’ll struggle to saturate the bandwidth, doable but very hard. Also for general-purpose software latency matters. Namely, input to screen latency for desktops and mobiles, or request to response latency for servers. Get Windows or Android tablet with Intel Atom Z8300 (available for $80-100), and see how it performs, it has 4 very similar cores (minus AVX-512), and frequencies are very similar, too.
      - a1369209993 8y ago
        https://www.intel.com/content/www/us/en/processors/xeon/xeon-technical-resources.html https://www.intel.com/content/www/us/en/processors/xeon/xeon... shows at least six volumes of datasheets, and I still haven't found a instruction set refence. I have found https://www.intel.com/content/www/us/en/processors/xeon/xeon-phi-processor-x200-product-family-vol-2-datasheet.html https://www.intel.com/content/www/us/en/processors/xeon/xeon... (helpfully labeled "Datasheet, volume 2", rather than anything related to it's contents) which describes a subset of the aformentioned ten thousand random control registers. So no, Intel does not make [simple cores], it makes heaping piles of shit complete with malware ("Intel® Management Engine") buried at D22:F0 on a interal PCI bus. It isn't simple, it's designed to be incorrect (and even the parts that are supposed to be correct aren't), and I'm not surprised it fails on fast as well.
zurn 8y ago
A counter-anecdote: The features you listed started shipping (from Intel & MIPS) in microprocessors in 1996, 22 years ago. Intel's out-of-order Pentium Pro was beaten by the in-order DEC 21164 the same year. Also, there's the case of Intel losing to in-order ARMs in mobile. First with XScale, and later on with the in-order Atoms. (https://appleinsider.com/articles/15/01/19/how-intel-lost-the-mobile-chip-business-to-apples-ax-arm-application-processors https://appleinsider.com/articles/15/01/19/how-intel-lost-th...)
- whack 8y ago
  Sure, specific optimizations in specific markets may not be worth the cost they incur. Or they may not be valuable enough to overcome other weaknesses in the project. And yet, if someone tried to sell a server CPU today that was not pipelined, not OOO, and didn't have branch prediction, it would absolutely tank in the marketplace. I never said that performance optimizations should always be implemented. Just that performance optimizations should sometimes take precedence over simplicity.
  - ben509 8y ago
    > And yet, if someone tried to sell a server CPU today that was not pipelined, not OOO, and didn't have branch prediction, it would absolutely tank in the marketplace. You could sell it as a niche product for high security applications, since OOO execution is a nasty side-channel.
    - whack 8y ago
      That's an interesting idea. A "so simple it can't have bugs" design would never win over the mainstream-market, but it might be able to find a niche among extremely security-conscious users. This might be a great project for the open-source community to take on.
jstimpfle 8y ago
In the free software world, this is not a huge problem. You can make new CPUs and recompile programs with far fewer effort. All that complexity really comes from monetization.
fwip 8y ago
They did use OP's strategy in the CPU industry. The first CPUs were simple as heck. They spent 20 years making them reliable(and faster without compromising simplicity, mostly just node-shrinks), and they've only really been complicating the architecture in the last 30 years.
- davidgay 8y ago
  iAPX432

by whack 8y ago

Sir_Substance 8y ago

The OPs advice would probably be dissapointing if applied to making a curry too. It's fortunate he regularly used the word software throughout the blog, really.

Sir_Cmpwn 8y ago

The attitude of the CPU industry in this regard led to some recent well-publicized, very bad, and nigh-unfixable security vulnerabilities, as you might have heard.

whack 8y ago
And yet, I don't see you or anyone else committing to buy ultra-simple non-pipelined non-OOO desktop/server CPUs. If you insist on only hiring chauffeurs who drive at 100mph, you can hardly complain when they get into a few accidents.
- Sir_Cmpwn 8y ago
  The main problem is that most software leans on those misfeatures as a crutch to excuse heaping layers of abstractions. Unfortunately this problem comes from several places, so it's not as easily fixed. That being said, consider me lined up to buy one of these CPUs.
- a1369209993 8y ago
  > And yet, I don't see you or anyone else committing to buy ultra-simple non-pipelined non-OOO desktop/server CPUs. If you can find a CPU that has the same number of non-cache[1] transistors as a Intel/AMD chip, but spends them on a larger number of simple (and preferably independent/non-hyperthreaded) cores, rather than squandering them on speculative execution and ten thousand obscure model specific registers, I would absolutely buy several of them. 1: and similar amounts of cache, of course.
  - Const-me 8y ago
    Intel makes them and you can buy them today, with up to 72 Atom CPU cores, e.g. (1) https://ark.intel.com/products/95830/Intel-Xeon-Phi-Processor-7290-16GB-1_50-GHz-72-core https://ark.intel.com/products/95830/Intel-Xeon-Phi-Processo... Very niche products. For massively parallel number crunching, GPUs are much better in both performance/watt and performance/dollar. That Xeon Phi 7290 delivers up to 3.45TFlops, costs $3200, and consumes 245W. Compare with GeForce 1080Ti 10.6 TFlops, $700, same 250W. For general purpose software they don’t work particularly well either. Most IO interfaces is serial, SATA, PCI-X, they have very few wires going to CPU. If you’re IO bound and you don’t have enough single-thread performance you’ll struggle to saturate the bandwidth, doable but very hard. Also for general-purpose software latency matters. Namely, input to screen latency for desktops and mobiles, or request to response latency for servers. Get Windows or Android tablet with Intel Atom Z8300 (available for $80-100), and see how it performs, it has 4 very similar cores (minus AVX-512), and frequencies are very similar, too.
    - a1369209993 8y ago
      https://www.intel.com/content/www/us/en/processors/xeon/xeon-technical-resources.html https://www.intel.com/content/www/us/en/processors/xeon/xeon... shows at least six volumes of datasheets, and I still haven't found a instruction set refence. I have found https://www.intel.com/content/www/us/en/processors/xeon/xeon-phi-processor-x200-product-family-vol-2-datasheet.html https://www.intel.com/content/www/us/en/processors/xeon/xeon... (helpfully labeled "Datasheet, volume 2", rather than anything related to it's contents) which describes a subset of the aformentioned ten thousand random control registers. So no, Intel does not make [simple cores], it makes heaping piles of shit complete with malware ("Intel® Management Engine") buried at D22:F0 on a interal PCI bus. It isn't simple, it's designed to be incorrect (and even the parts that are supposed to be correct aren't), and I'm not surprised it fails on fast as well.

zurn 8y ago

A counter-anecdote: The features you listed started shipping (from Intel & MIPS) in microprocessors in 1996, 22 years ago. Intel's out-of-order Pentium Pro was beaten by the in-order DEC 21164 the same year. Also, there's the case of Intel losing to in-order ARMs in mobile. First with XScale, and later on with the in-order Atoms. (https://appleinsider.com/articles/15/01/19/how-intel-lost-the-mobile-chip-business-to-apples-ax-arm-application-processors https://appleinsider.com/articles/15/01/19/how-intel-lost-th...)

whack 8y ago
Sure, specific optimizations in specific markets may not be worth the cost they incur. Or they may not be valuable enough to overcome other weaknesses in the project. And yet, if someone tried to sell a server CPU today that was not pipelined, not OOO, and didn't have branch prediction, it would absolutely tank in the marketplace. I never said that performance optimizations should always be implemented. Just that performance optimizations should sometimes take precedence over simplicity.
- ben509 8y ago
  > And yet, if someone tried to sell a server CPU today that was not pipelined, not OOO, and didn't have branch prediction, it would absolutely tank in the marketplace. You could sell it as a niche product for high security applications, since OOO execution is a nasty side-channel.
  - whack 8y ago
    That's an interesting idea. A "so simple it can't have bugs" design would never win over the mainstream-market, but it might be able to find a niche among extremely security-conscious users. This might be a great project for the open-source community to take on.

jstimpfle 8y ago

In the free software world, this is not a huge problem. You can make new CPUs and recompile programs with far fewer effort. All that complexity really comes from monetization.

fwip 8y ago

They did use OP's strategy in the CPU industry. The first CPUs were simple as heck. They spent 20 years making them reliable(and faster without compromising simplicity, mostly just node-shrinks), and they've only really been complicating the architecture in the last 30 years.

davidgay 8y ago
iAPX432