Pi 3 Benchmarks: The Marketing Hype Is True

The spec bullet list for the latest Raspberry Pi begins as you’ve already heard: WiFi and Bluetooth, now standard. While this is impressive itself, it doesn’t tell the whole story. The Pi 3, with an ARM Cortex A53, is up to 50% faster than the Pi 2 from last year. That’s an astonishing improvement in just 12 short months.

In playing with the Pi 3 for a few hours, it’s apparent the Pi 3 is fast. It passes a threshold of usability. The Raspberry Pi isn’t a computer that just sits on a shelf and runs a few cron jobs and blinks LEDs anymore – this is a computer that’s usable as a computer. But how fast is it? By stroke of luck, the official website for the Cortex A53 gives us a direct comparison between this chip and the CPU in the Raspberry Pi 2:

image credit: arm.com
image credit: arm.com

In real devices, the performance improvement from the Pi 2 to the Pi 3 is somewhere between 40 and 60 percent. At least that’s what ARM and the Raspberry Pi foundation are claiming. Is this true? There are tests we can run, and the marketing speak, for once, isn’t too terribly off the mark.

Benchmarks

The standard benchmarking tool for the Raspberry Pi and other Single Board computers is Roy Longbottom’s Raspberry Pi benchmarks. These benchmarks are able to test the computational power of Pis in nearly every respect, from integer performance to floating point performance, to how fast NEON instructions (the same idea behind Pentium MMX) are executed. There are benchmarks for OpenGL, memory speed, and Java.

Many of these benchmarks can be disregarded. There’s no need to test how fast the Pi 2 and Pi 3 can access memory; the RAM chip is the same. Java will die a slow and miserable death about a hundred years from now. The state of OpenGL drivers for the Pi is in flux right now. For these tests, I’ll only concentrate on the integer and floating point performance of the Pi 2 and Pi 3.

Dhrystone

The Dhrystone benchmark is the standard measure of integer performance, producing a number of Millions of Instructions per Second (MIPS). Tests were conducted on the Raspberry Pi 2 running a 1 GHz, and the Raspberry Pi 3 running at 1.2 GHz. The Raspberry Pi 2 clocked in at 1822 MIPS, The Raspberry Pi 3 clocked in at 2451 MIPS, an increase of 25%.

Linpack Single

Linpak Double

The Linpack Double Precision and Linpack Single Precision benchmark is the standard measure of floating point performance, producing a number of Millions of Floating Point Instructions per Second (MFLOPS). Tests were conducted on the Raspberry Pi 2 running a 1 GHz, and the Raspberry Pi 3 running at 1.2 GHz. The Raspberry Pi 2 rated 151 and 157 MFLOPS for single precision and double precision Linpack, respectively. The Raspberry Pi 3 rated 181 and 192 MFLOPS for single precision and double precision Linpack, respectively. The Linpack performance increase for the Pi 3 over the Pi 2 was increased 17% for single precision and 18% for double precision.

Power

The choice of power connector for the original Raspberry Pi was a stroke of brilliance. In 2012, a micro USB connector was found in every good smartphone, and chargers were plentiful. You could walk into any dollar store and find something that would convert mains power to 5VDC into a USB receptacle. The quality of these cut rate power supplies wasn’t really there, but it didn’t matter.

The Raspberry Pi capitalized on the ubiquity of cell phone chargers and used this as the power connector. For the time, it was a great idea. A stock Raspberry Pi wasn’t going to draw more than the 500 mA that were guaranteed to come from these power adapters. For a time, it worked.

PowerCon

Every USB port on the planet should be able to provide 500mA to peripherals, Raspberry Pis, and USB-powered paraphernalia. Going above this is possible, provided you have a good power supply. While the Pi 1, Pi 2, and Pi Zero can be used with any power supply, this is not true for the Pi 3. Compared to the Pi Zero, it’s a power hog. Add in a few USB dongles and the new on-board WiFi adapter, and you’ll need a 2.5A power supply at least.

Networking

Until now, there was exactly one networking option for the Pi. You could either use the on-board Ethernet, or grab a ubiquitous 802.11n WiFi adapter and use wireless networking. Notice I said one networking option.

The USB and Ethernet for the Raspberry Pi all go through one chip, the LAN9514 combination 10/100 Ethernet adapter and four-port USB hub. This chip connects to the CPU through the single USB port on the chipset, limiting bandwidth from the four USB ports and the Ethernet to around 480Mbps. Real world use cases will vary.

The Raspberry Pi 3 includes WiFi and Bluetooth, but it doesn’t go through the USB port. Instead, the best way to connect a Pi 3 to the Internet now happens through the SDIO port on the CPU. This port is accessible through the GPIO pins and has been used for great effect in custom WiFi hats, but does it deliver more bandwidth?

SCROT

The standard way of testing a network interface on a local network is iperf; all you need to do is run an iperf server on one box, and run the test on the other. This is not a test that will give you the absolute throughput of a network interface. It’ll be close if you’re testing two computers connected with an Ethernet cable, but when measuring WiFi, iperf should only be used as a basis of comparison. That’s exactly what we’re trying to measure, anyway: is the Pi 3’s built-in WiFi faster or slower than a USB WiFi adapter?

The results are insignificant. An Edimax 802.11n USB WiFi adapter was able to pull down 22.7 Mbps. The Pi 3’s built-in WiFi managed 21.2 Mbps. That’s more or less the same.

A Real World Test

Dhrystones, MIPS, Linpack, and other benchmarking tools are just that; benchmarks. They don’t measure real-world performance of code you would write. There’s an simple, standard way of figuring out how fast a little bit of Python will run: finding all the prime numbers below one million. There are 78,497, but how fast can the Pi 3 find them all?

Seive

The Raspberry Pi 3 found all the prime numbers below one million in one minute, 7.49 seconds. With the same Python script running on the same SD card, a Raspberry Pi 2 found all the prime numbers below one million in two minutes, five seconds.

The Marketing Hype Was True

The lede from the Raspberry Pi foundation in announcing the Raspberry Pi 3 was WiFi, Bluetooth, and a CPU that’s about twice as fast as last year’s Pi 2. The marketing materials from ARM provided a direct comparison between the Cortex A7 in the Pi 2 and the Cortex A53 in the Pi 3, with the latter being a 40 to 60% speed boost over the former.

The hype, for the most part, is true. In real-world tests, the Pi 3 may be about twice as fast. In the standardized benchmarks, the Pi 3 is about 20-40% faster. Networking speeds remain the same, though, but the Pi 3 still opens up a lot of doors to interesting applications. Personally, I have a USB Nintendo 64 controller on order.

123 thoughts on “Pi 3 Benchmarks: The Marketing Hype Is True

    1. That is the right question…

      Can you.

      and of course how much.

      Here comes the next set of bench marks.

      Oh how I wish I had the cash right now.

      ( But I’m still trying to get a PI 0 First. )

        1. Did you find this useful?
          Help me buy a #PiZero and I make more giveaways:
          They want help to buy what they tell what they us they can’t find. I wonder if they are holding out on us?

      1. The official answer is no, the unofficial (do at your own risk) answer is yesL most people have ramped their Pi3’s to the 1.35 – 1.4Ghz range (plus 500Mhz GPU and SDRAM) with varying success, not a huge improvement but enough to make more clunky emulators run much better; be prepared to deal with overheating though, in my experience the chip will require a beefy (or fan assisted) heat sink, it very easily and quickly reaches the 80+C mark at these speeds.

      1. It’s SDIO which means chances of having monitor mode one day is extremely low. I wish they had chosen a real wireless card such as ath9k or ath10k instead of broadcom.

  1. You need to update all your percentages. The formula is (new value – old value) / old value. It appears you’ve been dividing by the new value, which is causing the numbers to be lower than they should be.

    1. I saw the same error in the author’s calculations. I was just scanning through the comments to see if someone had already posted the error.

      the increased performance percentages should be:
      Dhrystone benchmark: 34%
      Linpack Single Precision benchmark: 20%
      Linpack Double Precision benchmark: 22%
      Python Sieve test: 86%

  2. Could you test if it is possible to use a network cable of 100m. I’ve tried a 80m cable on the Rpi2 and that doesn’t work. If I connect a laptop on the same cable, I get a connection.

    It seems that the Rpi2 doesn’t got enough power to send get a connection over 80m of CAT5e FTP cable.

        1. it is not the cable. If I change the Rpi to a laptop, it works.
          BTW the cable was new.
          Shadly I’ve cut it into 2 pieces 80m and 20m hopen 80m would work, but it is still to long.
          I think that the ethernetchip onboard cannot provide enough power to overcome the resistance of a long ethernet cable.

          Can anybody verrify if his Rpi has the same problem?

          1. LAN9514 was tested at UNH IO Labs and worked fine with 100M cable. Most likely is a cheaper magnetic (cost saving). Swapping it out for a a Bel or Halo Magnetic Jack typically solves the issue.

  3. I can’t tell whether any of these benchmarks take advantage of having twice the cores in the new Pi.
    Getting an overall 50% performance improvement out of a 20% faster-clocked improved chip with twice as many cores seems less than it should be…

    1. Using rhetoric to up-sell hype causes cancer in kittens…

      They proved that for general tasks a 20% clock increase gets a 25% improvement at 64 bit for twice the energy consumption. No mention of the core temperatures or “free -m” under load is a bit annoying.

      My wish for the pi 3 system is still to open the orphan GPU driver (not just the stubs), and minimal 4GB+ ram for 64bit based cores. Note it would become a legendary gaming platform with 16GB ram to buffer resources on the SoC.

  4. A further performance improvement of 15-30% could probably be achieved if the RPi 3 was running a true 64-bit Linux OS. Hopefully the RPI Foundation will start working on a 64-bit version of Raspbian in the future. But I completely understand why they’re being cautious about this (focus more on backwards compatibility).

      1. The cool part of this is that you’re correct, “It’s linux, you can work on that yourself :-)”. If the RPI Foundation does not come out with a 64 bit version in a reasonable amount of time, someone else will. I believe the RPI Foundation is smart enough to recognize that.

    1. performance gains from 32 to 64 bits instructions are about 10% and only in select applications where heavy, large numbers calculations are involve. 15-30% is never going to happen

  5. “The Pi 3, with an ARM Cortex A53, is up to 50% faster than the Pi 2 from last year. That’s an astonishing improvement in just 12 short months.” Really? Astonishing? When they started 4 years ago with what was an old core already, the ARM11, showing big improvements is easy. And now using the A53 that many started using last year, why would it be astonishing? But it sill sell! They are planning 100K a week production and what IS a bit astonishing is that Ebon said the network boot ability will make the board a good choice for factory automation. Interesting target market for educators getting people started with programming. Not very good news for people in that business.

    1. 150% performance for only 200% of the power drawn! (Ok, not fair, as WiFi most likely adds a bit on that power)

      Still the power requirements are a turnoff for me. Most people that I know with Pi2 have experienced some kind of power problem.

          1. True, still for $35 it’s an incredible piece of kit. It runs Linux after all. Just run the PI in pure CLI headless mode without a desktop or use a very lightweight window manager like openbox or i3. 1GB of RAM will be plenty then.

            It’s the desktop environment and web browsers that require large amounts of RAM. most other apps..especially cli apps; do not require that much RAM

    2. I guess when you set the bar extremely low to begin with it is pretty easy to continue making astonishing improvements. Notice no one from the Pi Foundation or pro-Pi news sites like Hackaday dares to post a comparison of the Pi’s performance to other major SBC’s. A person would almost certainly get banned from the Pi forums if they dared to post one themselves.

  6. “The results are insignificant. An Edimax 802.11n USB WiFi adapter was able to pull down 22.7 Mbps. The Pi 3’s built-in WiFi managed 21.2 Mbps. That’s more or less the same.”

    I don’t think this is a good test. I think you’re either WiFi signal strength or router limited.

    The RPi Zero ESP8266 WiFi solution gets as high as 46.4 Mbit/s. USB WiFi on the Pi has been seen to go as high as 81 Mbit/s.

    The SDIO interface’s throughput is clearly theoretically lower than USB (~100-some Mbit/s vs. 480 Mbit/s), but obviously SDIO isn’t shared, USB is, and at some point it really doesn’t matter. But it’d be interesting to know if SDIO actually ends up being a bottleneck on WiFi at any point.

    1. I was just about to say the test is bullshit, the tester’s results just show he lives in an area with a lot of interference and/or his setup sucks. In an ideal setup and environment you should be getting a lot, LOT more than 22Mbps — I live in an apartment – condo and there are a lot of wireless networks in the area causing interference, but I am still getting ~43Mbps using a similar 802.11n USB WiFi-stick.

      The test doesn’t show the performance of the SDIO WiFi – adapter, it just shows the incompetence of the tester.

  7. There’s no way that the WiFi on the Pi 3 is actually limited to ~20 Mbit/s. You’ve got to be limited by WiFi signal strength or your router.

    SDIO WiFi on the Pi Zero through an ESP8266 has been shown to get ~40 Mbit/s reliably, and I’ve seen USB WiFi on a Pi reach ~80 Mbit/s. The SDIO port’s maximum throughput is theoretically lower than USB’s, but both of them are well above what most typical WiFi setups can do anyway, so I don’t think it matters. But that test definitely isn’t telling you anything.

    1. I snagged a 3 yesterday. I’ll run it against my benchmark setup when it gets here.

      Looking at the kernel changes from yesterday, the Pi3 still has the weird SDIO clock divider that keeps it from running at exactly 50MHz, so it looks to be running at the default 41.66MHz. The adapter is using 4-bit mode, so SDIO is unlikely to be the bottleneck anyway (theoretical 166Mbits/s half duplex)

  8. In my testing with the ESP8266 on the Pi, it takes some tweaking to get a good WiFi throughput benchmark. Having the Pi too close to the router produced worse results than a couple of feet away. I had to do a wireless survey of my house to find a room and channel that has the least overlap with my neighbor’s routers.

    1. The Pi and Pi 2 both used by then out of date soc’s, so it is inevitable that large performance gains will be experienced when a more up to date chip is used. What you are seeing is Moores law in effect with old chips, nothing else.

  9. At the risk of spamming, I’d like to plug my solution to the Pi power dilemma. It’s called Pi Power, and it’s a buck converter that will generate a solid 2A @ 5v from anything between 6 and 14 VDC in. It’s a lot easier to get a cheap 10W @ 12v wall wart than it is to get a really *good* 10W @ 5V. Pi Power has a 2.1mm barrel connector for input power too, which is easier to come by, and injects 5v directly to the 5v pins on the GPIO header. The current shipping version has a 2A polyfuse that works with the TVS diode on the Pi for overvoltage protection (as well as protecting Pi Power from shorts).

    It’s available on Tindie.

    1. I just use 4A Mean Well supplies from Mouser. Never let me down yet and being commercial/industrial grade they have about every safety feature in the book. They do have barrel jacks but adapters are easy enough to come by.

      1. I have used some 4A Mean Well PSUs at work and they’re super noisy as far as “quality” PSUs go. They’re just fine for powering a single board computer but a poor choice for powering something with sensitive analog circuits.

  10. It is saf that the authors felt the need to spell out to their readership that for individual bench marks “Larger is better” or “smaller is better”.

    If you don’t have enough knowledge to interpret bench marks, you shouldn’t be looking at them… Spend your time learning the basics.

    1. What utter twaddle. Someone who doesn’t understand benchmarks shouldn’t begin to learn about a benchmarks through a graph on hackaday?

      What is the holy canon source for learning about benchmarks from which we must not deviate oh great planofuji?

      1. Design oversight? Even if the number of users looking to listen to or record FM radio with a Pi is quite low, I’m sure some interesting hacks (such as dirt cheap data transmission with cheap FM transmitters?) could be done with it. They should just bring it to a test point if they haven’t already.

  11. “The quality of these cut rate power supplies wasn’t really there, but it didn’t matter.”

    I beg to differ. It did matter. Many ‘unstable’ Raspberry PI’s became perfectly stable once the power supplies were exchanged for better ones.

    1. Unfortunately it continues to matter to this day. I’ll never understand why people skimp on power supplies. I mean they are not terribly cheap but a quality power supply is a gift that just keeps on giving. In my opinion anything that plugs into an AC wall socket should be as high of quality as you can reasonably afford.

  12. Despite they’ve created another wonderful board they still insist on that fragile micro usb port for powering the board. Why don’t they put a barrel connector or at least a pair of pads for soldering wires for a PSU?

    1. They use Micro USB because it’s cheap, it’s easy to implement, it’s highly compatible with the chipset and to top it all off almost everyone already has one of these wires laying around their house so they don’t need to include one with the unit.

      1. The only relevant thing there is *cheap*. The power port has nothing to do with the chipset; its data lines are not connected and the power lines go via a couple of buck converters to produce 3.3v and 1.8v. And most people have a pile of 12V 1A switchers laying around from old external hard drives etc.

      2. The extra cost of a pair of pads is zero. There’s no way cheaper than that. And the increased power consumption makes the ordinary wall wart useless, that’s clear in the article.

        1. Wall-warts get more beefy all the time, used to be they were 500mA then over time because of USB2/3 and cheaper high-power parts it became a standard to have 2.1 but now we have USB3.1 so I expect the next gen will go 3A.

          1. Micro USB are rated for up to 2.4A, so I won’t expect USB 3.1 ratings without a change of connectors. USB charger voltages cannot be assumed to be any where at 5V at high current at all – see the curve in the specs.
            A footprint for unpopulated barrel connector would be much better for higher current.

          2. Doesn’t matter what the USB specs say to the chinese manufacturers ;) And I simply expect to see 3A adapters, regardless of how well it fits in specs.
            USB2.0/3.0 is also not speced for 2.1A but that is what the standard became because people want fast charging.
            And to quote wikipedia: “In July 2012, the USB Promoters Group announced the finalization of the USB Power Delivery (“PD”) specification, an extension that specifies using certified “PD aware” USB cables with standard USB Type-A and Type-B connectors to deliver increased power (more than 7.5 W) to devices with larger power demand. Devices can request higher currents and supply voltages from compliant hosts – up to 2 A at 5 V (for a power consumption of up to 10 W), and optionally up to 3 A or 5 A at either 12 V (36 W or 60 W) or 20 V (60 W or 100 W).”
            So old USB even has allowances for 5A at 12V.
            It also states that apple’s pad was one of the devices drawing 2.1A when 2.0 was suppose to be the max for USB at that time. So it seems apple engineers decided it was OK. And it seems the standard allows for quite some wattage.

            So the question is do we have material technology to allow a modern micro-USB to push 2.5A over a tiny port.
            Seems the Pi foundation thinks so, since they use that port for a device requiring 2.5A to their own specifications.
            And they got FCC approval too, and I assume a CE certificate and all, so it seems there is no objections from the oversight commissions, unless they completely missed the 2.5A over micro-USB thing.

      1. Sure, for me is that easy, but not for everyone. It’s not a void criticism; on the contrary, I’m proposing two solutions (one of them at zero cost) because the PIs are amazing but I believe that the boards can be improved.

  13. [facepalm] After skimming through some of the article and comments, it’s painfully obvious that most people have forgotten (or perhaps never bothered to learn) that the Raspberry Pi Foundation didn’t design the Pi to make a profit, or to compete with other Single Board Computers (SBCs). Go to https://www.raspberrypi.org/about/ and educate yourselves before spouting more uninformed nonsense. Comparing it’s specifications/performance and then complaining that it’s not as powerful as other SBCs is idiotic. All the other SBCs it’s being compared to are for-profit commercial products. Also, I fail to see how the foundation’s claims regarding performance can be construed in any way as “marketing hype.”

    Keep in mind that the Pi is first and foremost an affordable educational tool. It’s not intended to outperform any conventional desktop or laptop machine.

    One final note: I’ve got an A, a handful of Bs, a B+, and a 2 B. I’ve never had a single power supply issue with any of them. I find that most people who complain of supply issues are simply using ‘Wun Hung-Lo’ switching adapters and USB cords.

    1. I guess someone needs to remind the Pi Foundation of that. Something tells me the Compute Module is not rated for children under the age of 18+ and working for a business enterprise. Nor is the “Raspberry Pi Customization Service.”

      1. In the element14 livestream at launch Eben reported that, although it’s hard to capture actual numbers, they believe that the pi’s ‘in the wild’ are evenly split across hobbyist, educational and industrial, about 1/3 in each application.

    1. *IN GENERAL* the FCC frowns on WiFi antennas that are exchangeable / enhanceable. The manufacturer is supposed to certify Part 15 compliance on the system as a whole. Making it possible to upgrade the antennas invalidates that certification unless the upgraded antenna also passes muster, and if it’s an upgraded antenna, it *probably* won’t, given that it may have increased gain over the stock one (otherwise, why would you bother?).

      RP-SMA was invented because it was – at least at the time – uncommon enough that the FCC allowed the manufacturers to say with a straight face that it wasn’t interchangeable with anything else. Since then, of course, RP-SMA has basically become the “standard” WiFi antenna connector (or at least one of them), so that claim’s harder to make now.

  14. Only one comment (!) has mentioned this, so I thought I’d repeat it, because it’s important for the article:

    All of the percentages you report are incorrect. For example, the improvement in the Dhrystone test is 34% not 25%. You need to divide the higher result with the lower result.

    Also, you should probably be more clear with the fact that the Pi 2 is overclocked by 11%, while the Pi 3 is not overclocked at all. I know you explicitly say what frequency you test at, but those who don’t know that the default frequency for the Pi 2 is 900 MHz will probably assume both are tested at default frequencies.

    It’s interesting (and a little sad) that this whole article deals with the performance difference between the Pi 2 and the Pi 3, yet almost no one seems to be concerned with the accuracy of what’s being reported.

    1. I was a little surprised myself that nobody had commented on the fact that the number were off due to improper calculations, and even more surprised that author *still* hasn’t fixed it.

      They don’t call me Mr. Pi, so I didn’t notice the part about the Pi 2 being overclocked. Good catch.

    1. Could be worse.

      The linked Python prime-finder is kind of terrible. It doesn’t believe 2 is prime, for starters, which is kind of a big oversight. More importantly, it uses Trial Division rather than Sieve of Eratosthenes. Numbers on my high-end computer for the original TD code vs my lightly modified SoE code are 6.4s vs 0.4s (including in both cases about 0.2s of printing time).

      First, the RP3 is in no danger of taking over my desktop, being still 10x slower. Second, who knows what the relative performance would be with SoE: SoE is typically memory-latency-bound, so there would likely be little speedup.

      Other than that this was a fantastic benchmark.

      #!/usr/bin/python2
      # Copyright (c) 2016 Bart Massey
      # Generate prime numbers
      # First million using Sieve of Eratosthenes

      import datetime, sys, math

      # Function to print message
      def print_prime(x):
      print ” Prime : %7i ” %x

      # Largest prime to emit
      upper_limit = 1000000
      # Table of composites for sieve.
      composite = [False]*upper_limit

      # Sieve of Eratosthenes
      def sieve():
      for candidate in range(3, upper_limit, 2):
      if (composite[candidate]):
      continue
      for multiple in range(2 * candidate, upper_limit, candidate):
      composite[multiple] = True

      # Function to search for prime numbers
      # within number range
      def find_primes():
      count = 1
      print_prime(2)
      for candidate in range(3, upper_limit, 2):
      if(not composite[candidate]):
      print_prime(candidate)
      count += 1
      return count

      # Check if the script was called with a
      # parameter. Use that as the upper limit
      # of numbers to search
      if len(sys.argv) == 2:
      upper_limit=int(sys.argv[1])

      # Record the current time
      startTime = datetime.datetime.now()

      # Start the process
      print “”
      print “Starting …”
      print “”
      sieve()
      count = find_primes()
      print “”

      # Measure and print the elapsed time
      elapsed=datetime.datetime.now()-startTime
      print ” Found %d primes in %s” %(count,elapsed)
      print “”

  15. Where are you getting the 192 Mflops (0.1 Gflops) for linpack bench on a stock (1200 MHz) Pi3?

    The stock pi 3 at 1200 MHz passes the linkpack bench at 6.1 Gflops (or 6000 Mflops)

    sudo apt-get install libmpich-dev
    wget http://web.eece.maine.edu/~vweaver/junk/pi3_hpl.tar.gz
    tar -xvzf pi3_hpl.tar.gz
    chmod +x xhpl

    reboot, then run:
    ./xhpl

    pi@raspberrypi:~ $ ./xhpl
    ================================================================================
    HPLinpack 2.1 — High-Performance Linpack benchmark — October 26, 2012
    Written by A. Petitet and R. Clint Whaley, Innovative Computing Laboratory, UTK
    Modified by Piotr Luszczek, Innovative Computing Laboratory, UTK
    Modified by Julien Langou, University of Colorado Denver
    ================================================================================

    An explanation of the input/output parameters follows:
    T/V : Wall time / encoded variant.
    N : The order of the coefficient matrix A.
    NB : The partitioning blocking factor.
    P : The number of process rows.
    Q : The number of process columns.
    Time : Time in seconds to solve the linear system.
    Gflops : Rate of execution for solving the linear system.

    The following parameter values will be used:

    N : 8000
    NB : 256
    PMAP : Row-major process mapping
    P : 1
    Q : 1
    PFACT : Left
    NBMIN : 2
    NDIV : 2
    RFACT : Right
    BCAST : 2ring
    DEPTH : 0
    SWAP : Mix (threshold = 64)
    L1 : transposed form
    U : transposed form
    EQUIL : yes
    ALIGN : 8 double precision words

    ——————————————————————————–

    – The matrix A is randomly generated for each test.
    – The following scaled residual check will be computed:
    ||Ax-b||_oo / ( eps * ( || x ||_oo * || A ||_oo + || b ||_oo ) * N )
    – The relative machine precision (eps) is taken to be 1.110223e-16
    – Computational tests pass if scaled residuals are less than 16.0

    ================================================================================
    T/V N NB P Q Time Gflops
    ——————————————————————————–
    WR02R2L2 8000 256 1 1 55.37 6.166e+00
    HPL_pdgesv() start time Sat Apr 23 15:14:17 2016

    HPL_pdgesv() end time Sat Apr 23 15:15:12 2016

    ——————————————————————————–
    ||Ax-b||_oo/(eps*(||A||_oo*||x||_oo+||b||_oo)*N)= 0.0025941 …… PASSED
    ================================================================================

    Finished 1 tests with the following results:
    1 tests completed and passed residual checks,
    0 tests completed and failed residual checks,
    0 tests skipped because of illegal input values.
    ——————————————————————————–

    End of Tests.

  16. One minute and six seconds to find primes below 1 million? Off-topic for the article, but it’s interesting to note that C++ code to do the same thing takes less than 2 seconds to run; about 500 milliseconds when coded to use multiple threads.

Leave a Reply

Please be kind and respectful to help make the comments section excellent. (Comment Policy)

This site uses Akismet to reduce spam. Learn how your comment data is processed.