One Article Review
| Source |  Errata Security |
|---|---|
| Identifiant | 1493999 |
| Date de publication | 2019-12-13 16:50:17 (vue: 2019-12-17 22:04:18) |
| Titre | This is finally the year of the ARM server |
| Texte | "RISC" was an important architecture from the 1980s when CPUs had fewer than 100,000 transistors. By simplifying the instruction set, they free up transistors for more registers and better pipelining. It meant executing more instructions, but more than making up for this by executing them faster.But once CPUs exceed a million transistors around 1995, they moved to Out-of-Order, superscalar architectures. OoO replaces RISC by decoupling the front-end instruction-set with the back-end execution. A "reduced instruction set" no longer matters, the backend architecture differs little between Intel and competing RISC chips like ARM. Yet people have remained fixated on instruction set. The reason is simply politics. Intel has been the dominant instruction set for the computers we use on servers, desktops, and laptops. Many instinctively resist whoever dominates. In addition, colleges indoctrinate students on the superiority of RISC. Much college computer science instruction is decades out of date.For 10 years, the ignorant press has been championing the cause of ARM's RISC processors in servers. The refrain has always been that RISC has some inherent power efficiency advantage, and that ARM processors with natural power efficiency from the mobile world will be more power efficient for the data center.None of this is true. There are plenty of RISC alternatives to Intel, like SPARC, POWER, and MIPS, and none of them ended up having a power efficiency advantage.Mobile chips aren't actually power efficient. Yes, they consume less power, but because they are slower. ARM's mobile chips have roughly the same computations-per-watt as Intel chips. When you scale them up to the same amount of computations as Intel's server chips, they end up consuming just as much power.People are essentially innumerate. They can't do this math. The only factor they know is that ARM chips consume less power. They can't factor into the equation that they are also doing fewer computations.There have been three attempts by chip makers to produce server chips to complete against Intel. The first attempt was the "flock of chickens" approach. Instead of one beefy OoO core, you make a chip with a bunch of wimpy traditional RISC cores.That's not a bad design for highly-parallel, large-memory workloads. Such workloads spread themselves efficiently across many CPUs, and spend a lot of time halted, waiting for data to be returned from memory.But such chips didn't succeed in the market. The basic reason was that interconnecting all the cores introduced so much complexity and power consumption that it wasn't worth the effort.The second attempt was multi-threaded chips. Intel's chips support two threads per core, so that when one thread halts waiting for memory, the other thread can continue processing what's already stored in cache and in registers. It's a cheap way for processors to increase effective speed while adding few additional transistors to the chip. But it has decreasing marginal returns, which is why Intel only supports two threads. Vendors created chips with as many as 8 threads per core. Again, they were chasing the highly parallel workloads that waited on memory. Only with multithreaded chips, they could avoid all that interconnect nastiness.This still didn't work. The chips were quite good, but it turns out that these workloads are only a small portion of the market.Finally, chip makers decided to compete head-to-head with Intel by creating server chips optimized for the same workloads as Intel, with fast single-threaded performance. A good example was Qualcomm, who created a server chip that CloudFlare promised to use. They announced this to much fanfare, then abandoned it a few months later as nobody adopted it.The reason was simply that when you scaled to Intel-like performance, you have Intel-like liabilities. Your only customers are the innumerate who can't do math, who believe like emperors |
| Envoyé | Oui |
| Condensat | 000 100 1980s 1995 a72 a76 abandoned about accept across actually adding addition additional adopted advantage again against ago all already also alternatives always amazon amd amount announced any apple approach architecture architectures are aren arm around attempt attempts avoid back backend bad barrier based basic became because becoming beefy been believe benefit better between big box boxes brings bunch but buy cache can care cares cause center championing chasing cheap chickens chip chips class clearly clothes cloud cloudflare college colleges companies compete competing compilers complete complexity computations computer computers consoles consume consuming consumption content continue core cores cortex could cpu cpus created creating custom customers data date decade decades decided decoupling decreasing dell design designs desktops didn different differs digital disappearing doesn doing dominant dominates effective efficiencies efficiency efficient efficiently effort efforts eke emperors encrypted end ended enough entire environment equation essentially etc even evolution example exceed executing execution exist fab fabrics fabs factor fanfare fast faster feature fewer finally finest first fixated flock forerunners found free from front game gaming gen generation get gets going good graviton1 graviton2 ground had halted halts has have having head heavily high highly hosts how however ibm ignorant important increase indoctrinate infrastructure inherent innumerate instances instead instinctively instruction instructions intel interconnect interconnecting internet introduced invest iphone isn its just keep know laptops large largest later latest leading legacy less liabilities like little logical long longer lot made mainframe make makers makes making manufacture manufactured many marginal market market: markets math matters meant memory microcontroller microsoft might million mips mobile months more moreover most moved much multi multithreaded nastiness natural new next nobody none not notebook notice now obsolete often old once one only ooo optimized oracle order other out own parallel partnership past people per performance phones pipelining platform plenty politics portion power press pretty previous processing processors produce promised qualcomm quite rackspace raspberry rather reason recently reduced refrain registers remained replaces resist rest returned returns risc risk roughly same samsung scale scaled science second sell server servers set shipping similar simplifying simply since single slower small sold some sometimes sony sparc specs speed spend spread started stored students succeed such sufficient summarized superiority superscalar support supports sure surface tech techies technology ten than that them themselves then there these this: thought thread threaded threads three through thus time today top traditional transistors trouble true tsmc turns two upshot use vendors very volumes waited waiting want wasn watt way western what when whether which who whoever why will wimpy windows without won work workloads world worth x86 year years yet you your |
| Tags | Guideline |
| Stories | |
| Notes | |
| Move | 
|
L'article ne semble pas avoir été repris aprés sa publication.
L'article ne semble pas avoir été repris sur un précédent.