Here are some slides I created to help my sons and their friend select CPU and GPU/Card for their Gaming Computer. I spend a lot of time talking about different CPUs and some of my intuition on the efficiency of different CPUs

1. How to Build a
Gaming Computer
Don Gillies
June 4, 2016

2. 1. Algorithm
● Select a CPU
● Select a Video Card
● Select a motherboard and RAM
● Select a storage device (disk or flash memory)
● Select a power supply and case
● Any extras? (DVD, Speaker, card reader, etc?)
● Put it all together
● Install Windows-10
● Select Mouse, Keyboard, LCD

3. 1. CPU Selection
● CPU = Central Processing Unit. The “brain” of the computer.
● Main properties of a CPU
○ Instructions per cycle (KHz, Mhz, Ghz) ~ speed (length of footstep stride)
○ Maximum clock speed (Ghz) ~ number of footsteps per second
○ Millions of instructions per sec per Watt of Power (Mips/Watt) ~ energy efficiency
○ Power of the instruction Set (floating point / scientific arithmetic is difficult to make fast).
There is not much difference between 32-bit and 64-bit arithmetic (32 is generally enough)
● Who knows something about CPUs?
○ AMD CPU?
○ Intel ATOM CPU?
○ Celeron (2955U, 3205U) CPU?
○ I3, i5, i7 CPU?
○ Xeon CPU?

4. Intel CPUs + Computer (Intel of Santa Clara, Oregon)
CPU
(i7 4-core)
CPU-Core
Thread
Thread
L1 Cache
CPU-Core
Thread
Thread
L1 Cache
CPU-Core
Thread
Thread
L1 Cache
CPU-Core
Thread
Thread
L1 Cache
L3
Cache
L1 Cache : 16-32KB. CPU Cores:
Good: 2 cores = 4 threads (~ i3+ laptops)
Better 4 cores (desktops only)
Best: 4 cores, 8 threads (~ i7 desk/laptops)
L3 Cache/core:
Bad: 1.5MB
Better: 2MB
Best: 2.5MB
RAM
RAM (shared all cores)
Good: 4MB (chromebook)
Better: 8 MB
Best: 16 MB
Graphics Processing Unit
(GPU) - massively parallel
EU/
RE
EU/
RE…………...
Texture
RAM
GPU RAM (shared all cores)
Good: 1 GB
Better: 2 GB
Best: 4 GB
Wow: 8 GB
Display
128-512 REs (AMD)
Up to 2048 (Nvidia)
Very high speed
(fast, wide)
GPU - integ
Integrated Intel GPU
Okay: 10-12 Exec Units
Good: 20-24 Execution Units
Better: 40-48 Execution Units
Best: 72 + 128KB Cache RAM
external
L2
Cache
L2 Cache/core:
Bad: 128KB
Better: 192KB
Best: 256KB

5. AMD CPUs (Advanced Micro Devices Austin TX)
AMD Bulldozer and PileDriver (A10-5750 etc) CPUs
● AMD chips didn’t actually have 2 cores. They have 2 copies of some arithmetic units to reduce
stalls (i.e. 2 adders #6, 2 fetching units #2), but can issue only 1 instruction at once.
● Intel chips truly can issue and run 2 parallel instructions at once, hence, 2 cores.
● Each instruction executes in many (8-15) pipelined “steps”.
● AMD Includes affordable 3D ATI GPUs and makes a great cheap 3D laptop.
● AMD Zen (early 2017) CPUs will be true 2-core (40% more instructions/clock) cpu, lower power too!
● AMD is in a (looks like successful) turnaround situation - no new CPU design in 5 years!
1 3 4 8 DONESTART
AMD - calls this charade
2 cores
Intel - 2 cores
Run 2 instructions at once
1 2 3 4 5 6 7 8 DONE
START
1 2 3 4 5 6 7 8 DONE
STALL
5
6
7
6
2
2

6. ARM CPUs (Acorn Risc Machines - UK)
● Used in Cellphones. Sold as a “digital file” for printing at a “VLSI foundry”.
● VERY Low-Power: Qualcomm Snapdragons and Apple A5 - A9.
● About 2x less energy per instruction vs. Pentium
● About 2x less instructions-per-clock (2 Ghz ARM = 1Ghz Pentium)
● Doesn’t run Windows
○ (Windows-RT was supposed to be for ARM but nobody bought it & it’s discontinued)
● Does run ChromeOS and Linux. Strawberry Pi = ARM.
● Does run Steam OS (a flavor of Linux)
○ Most of the Valve Games can be played on ARM, and very few others.
○ Runs Half-Life 1 and 2, Counterstrike (all flavors), Team Fortress, Portal, Left 4-Dead.
● ARM GPU is called MALI (integrated) and it’s about as slow as Intel HD.
● Qualcomm offers Adreno GPU 300-540 (integrated) which is a bit better.

7. ATOM CPUs (Intel)
● An Intel power-efficient CPU to compete with ARM
● Got a very bad name an now is called “Celeron nXXXX or Celeron XXXXn”
● The “Netbook” (2007) used these ultra-cheap low-power CPUs
○ Windows is very slow to begin with; intolerable on Netbooks!
● Horrible for games, run away! Run away!
● The very-best ATOM chip has the very worst pentium GPU (10-12 EUs).
● Recently (2014) Intel Started Lying by renaming the ATOM cpus “Celeron”
○ Celeron n2840, n3050, n3060, n3150, Bay-Trail, Cherry-Trail, XYZ-Trail are actually Atom
○ You can tell an Atom because cpubenchmark.com is 300-530 per core, not 600+
● Normally have 4 gerbils (i.e. 4 cores, no multi-threading).
● Less than 10 watts of power, suitable for tablets and Chromebooks.
● People buy them because the device needs no fan and lasts for 11 hours

8. What matters in a Gaming Computer
CPU - runs the AI algorithms:
1. Bot movements (20%)
2. Controls the GPU (60%)
3. Controls the AI (20%?)
4. Reads input - mouse, keyboard, takes maybe 0.1% of CPU
5. Most Games don’t need really powerful CPUs. 2 Ghz Pentium is okay.
GPU - shades triangles
1. Any polygon can be decomposed into triangles
2. Shades (rotates, clips, fills) between 100M and 1B triangles per second
3. Paints the polygons from a “texture” memory - see earlier diagram.

9. What Matters in a Gaming Computer
Display
1. Want a dedicated computer display
2. TVs are bad because color processing can add 20-50 milliseconds lagtime
3. Some LCD display are inherently slow (20 ms) and smudge or leave “trails”.
lagtime = delay until seen, slow LCD = delay until picture disappears.
4. Want a fast LCD display (< 8 milliseconds response time)
Keyboard, Mouse
1. Personal Preference
2. Many people prefer true keyswitch (Cherry) keyboards - clicky clicky
3. Many people prefer a wired mouse (might save 1 millisecond)

10. How to pick a CPU
Want a high clock speed (2.4+ Ghz laptop, 3.2+ Ghz desktop)
● Pentium CPUs are almost 100% faster than AMD at same clock speed
● Pentium CPUs are 100-150% faster than ARM, Atom @ same clock speed
● Does it make a difference i3 @ 3.0 Ghz vs. i5 @ 3.0 Ghz vs. i7 @ 3 Ghz?
○ Absolutely not <<===
○ Intel Prints only about 3 types of CPUs per generation:
■ i7 cpu
■ Celeron cpu (cost-reduced, designed in Israel)
■ Atom cpu
○ Intel tests each CPU and using lasers “disables” the broken parts
■ Laser-removal of 1-2 cores - an i7 becomes an i5
■ Doesn’t overclock at all, an i7 becomes an i3 (no turbo-boost)
■ Part of the RAM is broken, a 2.5MB L3-cache becomes a 1.5MB L3-cache.
■ Intel sets the CPU : Bus speed ratio (“clock multiplier”) to restrict overclocking

11. How to Pick a CPU
Secondary considerations after you have a 2.4 Ghz - 3.2 Ghz CPU :
Want at least 2 cores x 2 threads = i3, i5, i7
Extra Threads boosts performance by ~30%, costs Intel almost nothing.
Want turbo boost (= temporary overclocking, i5, i7)
Optional : 4 cores (i7-3630QM, Lenovo Y580 computer), why Optional?
● CPU controls the bots & AI & GPU
● Does not require a ton of CPU power.
● Most games are not written for 4 cores, they are written to exploit 2 cores.
Decide if you want to overclock (hotrod) your computer:
● Adds about $100 - $200 to CPU cost (+$30 cooler,+ $50 motherboard, +$30 cpu)
● Need to buy a cooler and overclock-able CPU
● Intel i5-4590K, i7-4770K, i7-4790K, i5-6600k, i7-6700k are some overclocking CPUs

12. What is overclocking
Popularized by www.tomshardware.com from Germany
Buy a CPU at speed X, run it at speed Y >> X. Need to do three things
1. Buy a high peformance (water?) cooler to keep it cooler (circuits run faster when cool)
2. Increase CPU voltage
3. Increase CPU clock speed
Many motherboards allow you to set the latter two parameters in the BIOS
Starting in ~2004 Intel tried to prevent overclocking by crippling the CPU at manufacturing time
1. Memory bus (MAIN cpu bottleneck) runs at 100 Mhz
2. CPU runs 24 faster than memory = 2.4 Ghz CPU
3. CPU runs 30x faster than memory = 3.0 Ghz CPU
4. These multiplier numbers (24, 30) are set by laser @ manufacturing time.

13. Why does ‘Locking the clock multiplier’ work?
Really fast (> 100 Mhz) memory is extremely difficult to make
● Dynamic RAM (DRAM) is optimized to be VERY CHEAP, not VERY FAST.
● Almost nobody cools-down their memories, the 2-4 DIMM module slots are too large
● All memory runs at 100 Mhz (10 nanosecond fetch time) for a decade+
● So if you overclock a CPU 2.4 Ghz to 3.6Ghz
○ The 100 Mhz memory is now being run at 150 Mhz
○ At 150 Mhz the memory will fail to produce results fast enough
○ The data will be corrupt and the CPU will crash
● DDR, DDR2, DDR3, DDR4 memories are not getting faster
● The memories are getting wider (more bits per transfer) and cheaper; not faster.

14. How to pick a GPU
A built-in GPU (Intel HD, Mali, Adreno) is usually not very good.
● Apple Macbook Pro 2015 had the 2nd best built-in GPU (HD 6100) available
● It won’t play most games at FullHD graphics, even at low settings
● Our HD 5500 (Toshiba Chromebook 2) is 75% - 90% as fast.
An external GPU is a thing of wonder
● Typically 2x-20x faster than a built-in GPU
● Supercomputer: Does 1, 5, 10 Teraflops of floating point arithmetic / second.
● Takes a lot of power, 75w - 500w compared to a laptop:
○ Apple Macbook Pro 2015 CPU + Integrated GPU uses 35w
○ Our Lenovo Y580 Gaming computer CPU + Intel CPU ~45w + GPU(GT660m) 50w

15. How to find a GPU
Nvidia makes the best GPUs
● They are descended from people who invented the GPU (Silicon Graphics)
● They pioneered new geometry innovations (DirectX 11 and onwards):
○ Waterfalls
○ Mist
○ No Radiosity => ⇐ Radiosity
You don’t need an expensive GPU to have a good gaming experience
● $200 will get you a fairly crappy Intel CPU chip
● $200 will get you an awesome NVidia GPU card - with fan & 4x speedier!

16. How to interpret NVidia GPU Models
Nvidia GT 760 = 7 + 6 =~ 13, roughly a measure of the performance.
NVidia GT 760 ~= NVidia GT 670, roughly, both have a performance level = 13.
7 is the generation of the GPU, higher is better (10 = current generation)
6 is the model level within that generation, normally it’s 4-9.
2, 3, 4, 5 = budget models
6, 7, 8 = intermediate models
9 = most advanced model

17. AMD Models
● We ended up picking an AMD Rx480 GPU card (Diamond Multimedia)
● I like supporting the underdog and AMD always dominates the low end
● The card does 5.8 GFlops of performance 8GB for $272 (with tax & shipping)
● The card is the first card < $300 that can handle Virtual Reality, i.e.:
○ 90 frames per second
○ 2160×1200 (stereo view of 1080 x 1200 per eye)
● The minimum VR card from NVidia is the GTX 970 or GTX 1060
● Our card can do
○ 1080p @ 60 Hz @ high detail, any game
■ 270 fps, League of Legends, 170 fps CS:GO, 90 fps Overwatch, high detail
○ 1440p @ 50 Hz,
○ 4K @ 30 Hz (for many but not all games)

18. End (For Now)
Next time
● Take apart an old gaming computer (Celeron 400Mhz with 3Dfx Voodoo-1)
● Produce a shopping checklist for components
● Pick a graphics card
● Size the power supply

19. How to Estimate CPU Speed
There are many benchmarks
I like the website www.cpubenchmark.com the most. Geekbench is similar.
Single core should be at least 1200-1500.
Multi core should be at least 3000-4000.
1. Acer c720 touch laptop - Intel Celeron 2955U 1.4 Ghz - 2 cores, 786, 1461 rating - okay
2. Toshiba Chromebook 2 2015 - Intel i3-5015U 2.1 Ghz - 2 cores x 2 threads, 1200, 3166 rating - good
3. Macbook Pro 2015 - Intel i7-5557U 3.1-3.4 Ghz - 2 cores x 2 threads + TB - 1900, 4934 rating - better
4. Lenovo Y580 - Intel i7-3630QM 2.4-3.4 Ghz - 4 cores x 2 threads + TB - 1689, 7900 rating - best
Notice the 1-core benchmark is almost *directly* proportional to clock speed (1.4 Ghz - 3.4 Ghz)
Not enough L2 or L3 cache can have a crippling effect on a CPU; all these CPUs have 2 MB L3 cache.

20. How to estimate GPU speed
How to find a good GPU?
● Look at the ratings on www.notebookcheck.com from Germany
● Look up the GPU
● Find it in the sorted list on the right side of most screens.
● Find your GPU class rating (class 1 - best, class 2 - good, class 3 - widespread, 4, 5, 6(very old))
● Class #1 will run any game at the highest settings, except for 4K. Y580 was class #1 originally.
GPUs are *not* rated in instructions per second.
● They are rated in Gigaflops/Sec or Gigatexels/second (FLOPS = FLoating-point OPerations Per Sec)
● Memory bus is ALWAYS the main bottleneck.
● External GPUs have an extremely fast extremely wide ( up to 512-bit ) data bus for textures
● Some AMD laptops run 2x faster with a 2nd bank of RAM, going from 64-bit to 128-bit data bus.
(Lenovo Thinkpad E545).

21. Proposed Computer Budget - $900 total (alan)
Spend $700 on Case, CPU and GPU Graphics card including Windows
● Spend ~25% on the graphics card - $200
● Spend ~50% on the PC itself - $500 including Windows($50)
● Get an overclocking motherboard
○ Fill half the RAM slots with half your eventual RAM (one 8GB DIMM) - expand it later.
■ AMD graphics (R7, R9) can go faster with 2 RAM channels, buy 2 x 4GB for AMD CPUs.
○ Watercooler and overclocking CPU can be added later
● SSD can be added later
Spend $200 on the screen, keyboard, mouse
● Economy Chinese Gaming mouse - $15
○ Redragon M601 CENTROPHORUS-2000 ; or Etekcity Scroll 6E
● Get a good 24” Full-HD monitor (IPS or TN panel) - $140 - $150
○ Asus VS239H-P (IPS), Asus VE248-H (TN)
● Keyboard - HP KU-0316 economy - $15 (beats all keyboards < $70)
● Headset (Turtle Beach refurbished - EBay - $30, Sades SA921 $22, Sades SA928 $36)

22. Moore’s Law
● Moore’s Law
○ Transistor sizes shrinks by sqrt(2) x sqrt(2) = 2x every 2 years
○ VLSI circuit speed will DOUBLE every 2 years : circuits get closer together
○ VLSI Circuit =
■ Transistors = Amplifier components (input, output, base = amplifier control)
■ Wires
■ Mostly. There are also capacitors, resistors, inductors, too.
○ DRAM truly has doubled every 2 years, DRAM designs are simple and regular.
○ CPUs increased 1.7x every 2 years
■ The circuits are irregular and change in every new design.
■ Multiplier has more parallelism
■ CPU needs more registers, more adders, more boolean arithmetic units, etc.

23. Moore’s Law
● History of Moore’s Law (Gordon Moore, Intel Founder, 1965)
○ 1971, Intel 4004, 4-bit CPU, 500 Khz, 125,000 4-bit instructions per second
○ 2015, Intel i7-4790K, 64-bit CPU, 4.0 Ghz, 1,4 billion 64-bit instructions per sec
○ 11200x speedup * 16x wider circuits (4 bits … 64 bits) = 179200x
○ From 1971-2016 this speedup is 1.7x every 2 years!!!!
○ Moore’s Law stopped working in ~2008.
● It is said that if cars followed Moore’s Law
○ Cars today would go 1000 miles per hour
○ They would cost $1

24. Moore’s Law has ended - why?
1. Imagine a city with houses (transistors) on 1 acre (our house = 1/5th of an acre btw)
2. Every two years the acre size is cut in half
3. However, the cars and roads (wires) never change size
4. Eventually you have many tiny tiny shrunken houses
5. Separated by relatively giant distances since the roads are still enormous!
6. The houses stopped getting closer together because the roads were too wide!
7. This really happened. The circuit wires stopped following moore’s law.
○ Faster parallel circuits have more transistors and connections and need more
wires.
○ Wires were made taller vertically, first.
○ Wires were changed from aluminum to copper to reduce resistance, second.
This helped shrink and speed up the wires but they stopped improving >10 years ago.

25. Intel has slowed down a lot recently!
1. Soon transistors will stop shrinking. Many people predict RIOTING in the streets!
● Skylake is 14 microns. At 5 microns geometry, transistors = 1-2 electrons wide
● 14 microns => 10 => 7 => 5. We are only 6 years from the end of the world!
● Actually, Intel recently slowed down (2015) to 3 years/shrink so we are 9 years away!
2. Transistors @ 22 microns (2014) - got much harder to shrink.
They became crazy-looking 3D waffle objects to pack more power in a tiny space
3. Maybe circuits can become entirely 3-D ? No progress yet.
Every time you print a VLSI Layer you lose ~ 0.5% - 2% of the chips on a wafer; chips have 20-30.
125 nm Planar Transistors
2001 2012 2015

26. Techniques to defy The End of Moore’s Law
Add more processor threads
● Add a second “register file” for idled cores (hyperthreading ~ 30%)
● When it stalls running one “thread” CPU swaps files and runs a 2nd thread
● Good for another 30% improvement
Add more Cores(CPUs) (80%, 70%, 50% improvement for first few cores)
● 2, 4, 6, 8, 10, 12, 18, 36-core CPUs
● Google uses 36-core CPUs (works because all run *different* programs)
Make the circuits wider
● DDR, DDR2, DDR3, DDR4 RAM
● Produce more data and use wider (32-bit, 64-bit, 128-bit, 256-bit) data bus

27. Techniques to defy The End of Moore’s Law
Old Way (1 change each year, 2 years per shrink)
● Tick: Improve the circuit designs and enlarge the memory paths
● Tock: Shrink the geometry (22 → 14 nm, for example)
New Way (1 change each year, 2 years/shrink, pioneered in ‘Kaby Lake’)
● Tick: Improve the circuit designs and enlarge the memory paths
● Tock: Shrink the geometry (22 → 14 nm, for example)
● Toe: Improve the onboard graphics (faster GPU, more MPEG hardware)

28. Summary from Last Time
1. Most games cannot make good use of > 2 cores. Orange column indicates winner.
2. Intel hyperthreads are counted as 0.3 cores/each, so 4-cores x 2-threads=5.2
3. Intel Turbo-Boost lasts 15-60 secs before CPU overheats &slows down; ignore it
Geekbench 3.0 multicore Instr/clock/core Cores Clock Speed Geek3m 2 cores L2/L3 Cache Power
Intel Pentium i7-6700 1 ~5.2 Up to 4 Ghz 16365 6294 0.5/2.5MB Good
Apple A9 1.56 2 Up to 1.85 Ghz 4423 4423 1.5/2MB Best
AMD Fx8350 0.54 8 Up to 4 Ghz 13200 3300 0.5/2MB Awful
Qualcomm Snapdragon 820 0.52 4 Up to 2.2 Ghz 3500 1750 2 MB (L2) Best
ARMv8 A57 (Snapdragon 810) 0.66 4 Up to 2.4 Ghz 6476 1619 2 MB Better
Qualcomm Snapdragon 800 0.39 4 Up to 2.3 Ghz 2722 1361 2 MB Best
Intel Atom / Celeron n3160 etc 0.30 4 Up to 2.4 Ghz 3022 756 2 MB Better

29. Conclusion: Things I want you to remember
1. PCPartPicker.com is a good website to select components that work together
2. Cpubenchmark.com is a good website to compare CPUs
3. Notebookcheck.com is a good website to compare GPUs
4. Most games only need 2 physical cores (circa 2016)
5. Hyperthreading increases multi-core performance by about 30%
6. Power = Performance, don’t be tricked by 3 Ghz Y-series (6.5w) or U(15w) cpus;
these spend their life “throttled” at 1 Ghz because the CPU overheats throttles
7. In a gaming computer about $500 is enough for a good CPU. Invest the rest of
the money in better and better GPUs.
8. Overclocking adds at least $100 to the cost of the CPU (i.e. $600)
9. I was surprised at how efficient (instructions/cycle/watt) Apple A-series
processors are

30. Extra Slides

31. Building Guide - Addenda
Who builds an AMD Gaming Computer?
1. You have less than $200 for a CPU and you want 8 real cores
2. You aren’t paying the power bill(s) or the air conditioning bill(s)
3. You think that high clock speed means “fast”
4. You want to overclock to go even higher
5. You want to edit videos, do 3-D ray-traced rendering, compile operating system software ie run
massively parallel apps quickly & cheaply.
Who builds an Intel Gaming Computer
1. You have $160 - $1000 to spend on a CPU
2. You don’t want to overclock or you buy a K-series (overclocker) CPU (i.e. i7-4790k, i7-4770k).
3. You want to brag to all your friends that your computer is the fastest, losing friends is okay ... …
4. When you lose at CS:GO or LOL it will be *all* your fault… Never the fault of the computer!

32. Additionally …
1. CPUs are only getting 2% faster every year, 4% every new generation.
2. Need a 25% speedup to notice a difference (12 years wait!)
3. Last 2 generations (Broadwell iN-5xxx, Skylake iN-6xxx)
a. Intel is saving *power* (broadwell graphics actually got slower per GPU core)
b. Intel is *not* making CPUs much faster-per-clock.
4. Speed is *Proportional* to the power going into the CPU
5. A 15-watt or 28-watt laptop will never beat a heavier 45w or 90w laptop.
6. Where is the power suffix on an Intel CPU? It’s the character
a. i3-4005U, i3-5015U, i7-5775U, i7-3630QM, n3160, m7-7Y60 etc.
b. U - 15 watt (chromebook, macbook air, 2 core) or 28watt part (macbook pro, 2 core)
c. QM - 45 watt (Y580 gaming computer, 4 core) part, more power → more cores
d. n < 10 watts = no need for a fan - ATOM suitable for a tablet or Chromebook - SLOW!
e. Y prefix ~ 4.5 watts - slowest! Pentium (Macbook 12”) but spends most of its life at 1 Ghz!