Ninja Sushi's Guide to Graphic Processing Units. (GPU)

Not open for further replies.


Jul 6, 2012
Ninja Sushi's Guide to Graphic Processing Units. What does it all mean?

Alrighty. Here is a quick guide to buying a GPU (Graphic Processing Unit/Graphics Processor Unit) I will break down what each part means and then sum up the 4 best GPU IMO on the market right now... and this some.

Reference Cards vs. Aftermarket Cards a.k.a. Retail Cards
Alright. So here is possible the biggest difference in all GFX cards. I will be using ATI Radeon as an example as it is much easier to spot a reference card vs. aftermarket. Let's take the newest cards out and use them. 7950 vs. 7970. All 7950 models are reference. Meaning they're stock from the factory as ATI intended with NO additions. Aftermarket additions could be any of the following and even all of the following. Better PCB, better heatsink and fans, pre-overclocked, etc, etc, etc. (PCB - Printed Circuit Board) *The back of the board... lol) Reference cards are good, but I always urge people to wait for the non-reference version as the perks given are much, much better.

Now you will notice that most cards out right now bounce between 3GB and 4GB of ram. This RAM is possible the most important thing about the GPU you want to buy. GFX card RAM - Random Access Memory - like CPU ram is extremely important when computing. The graphics card caches all volatile memory on the graphics card itself. All unit movement, shaders, shadows, colors, pixels, resolution size, EVERYTHING is stored on the GPU RAM. Without adequate ram you will notice a major decline in the performance of any games or GFX intense program you are using.

Core clock
Long story short: Core clock speed for a GPU is the same basic idea as for a CPU. It's how fast the GPU processes all task at hand. It is the speed at which a GFX card can run a 3D program and is derived from it's architecture. (More on this later.) So basically: The faster the CPU you have - I.E. i5-2500k vs. i3-2100 - the faster your computer can process a task. Same idea with a GPU core clock speed. 1000 MHz is better than a 900 MHz processor yet a 1084 MHz (EVGA 680) is faster than a 1000 MHz (XFX 7970) GPU. Core clock is important, but it is not the end all deciding factor like most people assume to believe.

Memory Clock.
Memory Clock speed is how fast the GFX card processes the RAM. More memory clock and more bus speed = more memory bandwidth. You may have 4 GBs of RAM, that is you can store 4096 MBs of data onto the onboard memory of the GPU, yet a faster memory clock means how fast the GPU can access this memory and process it. All GPU memory today is ran at DDR. Double Data Rate. This is saying that the GPU can read from and write to the RAM at twice the speed. 500 MHz of DDR is effectively 1000 MHz of memory clock.

Memory Interface
This one is pretty easy and most people will understand it. When describing memory interface I like to think of it as a major highway. The more lanes you have, the more traffic can flow through. The more memory interface you have, more data can flow through and be processed. 128 bit < 384 bit. etc. etc. etc.

Power consumption and wattage required - Here is a more indepth look at actually wattage vs. required/recommend wattage.
Power consumption and wattage required are two different beast. When a company post benchmarks of their cards, they will often show a required wattage amount - say 500W required. This is a safety buffer and many factors play into this: Mobo, cpu, ram, heatsink, etc. play into this 500W buffer zone. In all actuality most GPU pull from the wall about 200 watts. (Not counting OCing, full load test, etc etc etc. Now, if you buy to 7970s and they both require 500W, this DOES NOT MEAN!!!! that you will need an 1000W power supply. An 800W will be plenty of power to power all your components as well as two 7970 in X-fire mode.

Pretty easy stuff right here. It's what PCI-e slot the card requires. All new cards run at PCI Express 3.0x16. Don't worry about this unless you have a very old computer in which you want to upgrade it. (Wise man once said! "Tis better to buy new, than upgrade old.")

Stream Processors vs. CUDA Cores. (Let the flame war begin.. jk jk)
So here is the low-low on Stream Processors vs. CUDA Cores. SP is ATI whilst CC is NVidia. Plain and simple. Don't even try to compare the two as you will be comparing apples to oranges. Basically speaking, SP and CC are the architecture of the GFX card itself. This is saying how the card processes, runs, retrieves data, etc. Hence why some ATI cards run games better than some NVidia cards where as some NVidia cards run games better than some ATI cards.

ATI Radeon vs. Nvidia GeForce
All preference. Some people are fan boys while others are benchmark fan boys. I myself love reading benchmarks. I don't give a damn who makes what! I just want the best item I can get for my money in my PC. Plain and simple! I think both companies excel at different points and both flounder at different points.

Card Size and Dimensions!!!
THIS IS EXTREMELY IMPORTANT STUFF!!! Seriously.. If you buy card that is too big for your case, you will need to do one of two things. Go back and get a smaller card or get a bigger case! Pretty simple stuff right there. Now this probably will never happen to anyone, but it's always a good idea to check before you spend your money. No one likes paying a restocking fee and a shipping fee because something doesn't fit.

Things that do matter!
Everything above and most importantly.... DISPLAY PORTS!!! Make sure your card and monitor both have matching display ports. Otherwise you will be buying either a new card, new monitor, or a converter, all which cost way too much money.

Things not worry about as the probably don't apply to your situation.
GDDR5, Power Pin Connectors, Stream Processors and Cuda Cores, RAMDAC, HDCP Ready, Item and Model #. If you don't know what any of this is, it doesn't matter or apply to you. Do not worry about it.

Personal Opinion Time
Here a few cards that I feel are the best in their field right now. (This is not including the pending GHz addition of the 79xx series.)

XFX 7970: [...] 6814150586
MSI 7970: [...] 6814127670
MSI 680: [...] 6814127693
EVGA 680: [...] 6814130794

Here is the easy version: [...] 127-693-TS

If you have the money, go big.

Oh hey. Just upgrade to a 690. lol jk jk [...] 0GTX%20690

Keep in mind the big thing you want to look for is the RAM. Because if your computing at 2560x1600 (in my case 2560x1440) and you run out of GPU RAM, then that game will in turn go to the file location on the hard-drive and start caching HDD space as GPU-VRAM (Little known fact). Hence why putting a demanding game onto your SSD is a smart idea in case you do run at of RAM. It's cheaper than buying another GFX card and still operates pretty darn fast depending on the RAID, and if you have it as SATA 6.0. But above all the 680 does have 1 GB of GPU VRAM over the 7970 which is a large factor. (Also! If you run at of VRAM, it is better to lower the settings or even your resolution as the game with run MUCH smoother than caching on an SSD)

Now does this 1GB and some minimal clock speeds out weigh the price tags of well under 500 bucks to over 600? No. A 4GB, 1536 cuda core, clocked at over 1GB memory etc, doesn't really warrant paying 100-150 bucks more. With that money you could put away to X-fire another card.

NOW! That being said. IF!!! and only IF!!! You have the money, and lots of it. NVidia is the way to go right now. The 680 is better than the 7970 regardless, etc, etc, etc, because it does at 1 more GPU-RAM than a 7970 and some of the EVGA cards clock the highest, etc, etc, etc.

So if money = pocket change, then just go out and buy 2 690s and profit. But yeah, trolling (PATROLLING... not trolling as in forum trolling.) around newegg, ncix, and other websites I find the XFX 7970 Black Edition Double Dissipation is THE - best - bang - for - your - buck.... at $460 bucks. Or if you are an Nvidia fan than the EVGA Superclocked 670 is also a very tempting offer. 4GB of GPU RAM and some pretty nice clocks, all sitting around $484.

As always, I recommend someone correct me if I am wrong and add in anything I might have forgotten.


great guide but the 680 does not have 4gb of ram, at least the stock version does not have 4gb of ram, only aftermarket versions do. And i hope i'm not asking too much here but i feel like this guide is aimed at higher budgets rather than lower budgets. You generally need tips on buying cards when you're on a budget because the more money you spend generally means the better the card. So can you make one a bit more specifically on what's more important when it comes to buying cards rather than what means what.


Jul 6, 2012

Great point! Indeed this is a higher GFX card guide. But the GTX 670 and 680 both come at 4GB if you get the EVGA version. I personally will only buy EVGA and XFX due to their AMAZING customer service, OC protection, and life time warranties. I also only recommend those two companies for those two reasons. lol XFX and EVGA have been known in the past to upgrade your card free if it ever dies. XFX however has been known to upgrade X-Fire cards for free too.

Example: You have a 6850 and it dies. The replace it with a 7950, but wait! You had 2 6850. Tell them and they'll send you another 7850 in exchange for the old 6850.

Budget cards this applies to the same mind you. The guide applies to all GPU old and new; the only difference is that the specs of the older cards won't be the same. A really good budget card I can recommend is a Fermi 460 Ti or a GTX 580. For ATI I'd say 7850 and 6850 are both REALLY good cards.
Lots of things wrong here.

ATI does not exist.

there are plenty of non-stop reference 7950s

Memory is not important for the vast majority of people as there is no difference between 1 and 2 GB at 1080p and memory is sized as appropriate for the gpu

windows shares system memory with the GPU. The GPU absolutely DOES NOT default too using virtual memory and an SSD will absolutely not improve FPS or GPU performance.

core clock is irrelevant and cannot be compared at all between nvidia and AMD. You can only compare between cards with the same architecture, and even then it's not a great indicator.

Stream processors and cuda cores are the processors NOT the architecture.many different architectures use these.

Interface is not what the card REQUIRES. Pcie 3 cards do not require pcie 3 slots and won't run at pcie 3 without the right board and CPU. Any pcie card will run in any pcie card slot. Even the high end cards won't bottleneck in a 2.0x16 slot

Pretty sure power pins are important a it's a problem if you can't plug in your card.

You can only compare cards by looking at benchmarks.looking at core clocks and processors and memory interface does not help at all



Given your stress on VRAM, I think it's arguable that either the 460ti or a 580 are good buys right now. especially the 580 given it's price. we're headed into the 2GB of vram era...

as for the 680 with double vram (4GB), I'm still not convinced that we'll get there, but given your resolution etc I can understand. looking forward to the sapphire vapor-x with 6GB (now that's overkill LOL)


Jul 6, 2012

All right.. let's set some things straight.

First off this was a high end GPU guide with some affordable budget cards thrown in.

Second off ofc there are plenty of non-reference 7950, etc. I don't recommend 50s, etc models, because for a little more money you can get an after market gpu that OC's better, maybe is pre-oc'd, has non-reference pcb, custom heatsink, etc. etc. etc.

Memory is important for most users regardless. I know most monitors run around 1920x1080p and most games are converting to the standard 16:9 instead of 16:10 because of xbox, ps3, and console makes of the like. But as the game designers ramp up their graphics with more titles like Skyrim, Fallout 4 (coming out), Crysis 3 (The end all for GPU benchmarks), etc. games will be more demanding and having higher video random access memory will allow the GPU to process smoother. Much like CPU ram as I'm sure you know.

Also VRam is not virtual ram as in a CPU, paging files, etc. VRAM on a GPU is video ram, thought you should know.

Stream Processors and CUDA Cores are the architecture of the device itself. It is how it processes the information from 1s and 0s to data on a screen with colors and pictures.
GPU Architecture.

Read this by Nvidia
"Fermi: NVidia's new Cuda core Architecture"

I never said these cards require a PCI-e 3.0 port.

Power pins don't matter now days because most people buying a 7970 or a 680 won't be using PSUs from 2001.

Processing speed is processing speed.
Bus speed is bus speed.

I can compare these just as I can compare the speed of a i5-2500k to that of an 965 by AMD.


Apr 5, 2011
First off you guys there was never a 460 Ti, it was just 460, or 560 Ti.

Next, I see some flaws in your guide. Cuda cores/SPs ARE important if you want to compare just ATI or just Nvidia. Also, higher core clock does not always mean better performance. For example, an HD 6950 has a core clock of 800Mhz, but an HD 6870 has one of 900Mhz. Yet the 6950 is still faster, because you have to take into account the number of cores as well.

Performance within one architecture of cards, in an arbitrary unit can be viewed as:
P = K * Core count * Core clock.
This does not directly translate into real life performance however!


there's a world of difference between 50's and 70's. we're not in the world where a 50 is just a crippled version of a 70 anymore. the price and performance gap between the 7850 and the 7870 is HUGE. that and the best 7870, the MSI hawk, is currently selling for the same price as the MSI 7950 with a twinfrozer III cooler for $330 on newegg. now that's a sweet deal, MUCH better than any 7970 out there unless you're on a large budget and absolutely want the best (a pair of those 7950s will smoke a 4GB 680 in performance and cost almost the same).

I agree with you on the memory bit, but only to an extent. we've barely left the 1GB vram era in the high-end. in the mid to low end vram is still pretty much irrelevant. Nvidia likes to be a bit stingy on VRAM, and I agree that the stock memory on the 680 does leave a little to be desired, but is VRAM is what you're concerned over, really just get a 7970 or a pair of 7950s, the extra price for a 4GB 680 is NOT worth it. you're almost better off just selling your 680 when the 780 comes out. the money you're paying for the 4GB to "future proof" will cover the differences (if you take performance difference into account). you buy extra VRAM if you need it NOW (such as for triple monitor or higher). if not, just get stock design, future proofing is a dead concept. especially if you're buying high-end. you'll still want to get a new card within 2-3 years, otherwise you're almost always better off just buying mid-range for now and upgrade later. a mid-range card in 2-3 years will be hitting the 680's performance.

Unkol is not talking about vram being virtual ram. he's saying windows will share system ram if Vram runs out. you'll still take a big hit on performance though.

I think you guys are defining the term architecture differently... unkol is talking about the literal, physical architecture, you're talking about the abstract processing "cores". CUDA and Stream processors do not directly relate with physical architecture at some set ratio. but the physical architecture does define how many Stream/CUDA cores you'll get out of it

Power-pins DO matter. a 7970 or 680 can run off of a 600 or 650W, and a NEW PSU from a bad company won't have the correct cables (not to mention it may burn the card). it's always good to sort these things out with an OP when recommending a card. you can't just write "EVGA 680-4GB FTW!!!"

processing speed is NOT THE SAME as clocks... I think you're getting things confused.


Jul 6, 2012
Alright sure.. Ignore everything I say then. I have a xfx 7970 and I'm perfectly happy with it, might get a second at 460 bucks. Regardless, architecture, power pins, etc. etc. etc. Y'all are getting to detailed IMO. This was simple basic idea for people and some how we got into some serious expert arguing and not the basic low down of a gpu. Besides it's posted in the wrong section regardless. The End. I'm going to sleep.

I meant to say 460 Fermi btw.

The end.
Not open for further replies.