Request: How to interpret GPU specs, and why is one board "better" then the other?

[jghaverty]

I am coming from old school GPU's. Yes. I remember my 4mb geforce being the shiznit running counter strike 1.3 at 40fps. It used to be entirely memory based... the more it had, the better it was, that was almost a universal constant.

Now that I am getting back into the PC mod world, Im honestly flabbergasted at the level of changes between different cards with the same memory.

Specifically asking
1) which specs are most important now?
2) Can you compare a 7770 to a GTX 780 Ti, and why the 780 is (clearly) the winner, and why it's better. (im not dumb, I understand higher memory, faster clock speed, "MORE IS MORE" is better, but why is it SO MUCH better?)

I appreciate any information given... thanks.

 

[huilun02]

Now the GPU is more important than memory. A bigger factory with a better layout plan to have a greater number of workers, and more efficient workers.

You cannot quantify how well a GPU/card will run just by looking at the specs. You have to look at benchmarks nowadays, and see how well each GPU compares to each other at doing the same thing.

 

[jghaverty]

Now the GPU is more important than memory. A bigger factory with a better layout plan to have a greater number of workers, and more efficient workers.

You cannot quantify how well a GPU/card will run just by looking at the specs. You have to look at benchmarks nowadays, and see how well each GPU compares to each other at doing the same thing.

The GPU has memory...

and Im asking why a similar spec GPU benchmarks higher/lower then another similar specc'd GPU. More going past the general "which is better" questions around here.

 

[huilun02]

The graphics card has memory.

The GPU is a component on the card PCB, separate from the memory modules.

By 'similar specc'd' I assume similar hardware? Then the difference in performance would be due to one GPU running at a higher clock speed.

 

[jghaverty]

The graphics card has memory.

The GPU is a component on the card PCB, separate from the memory modules.

By 'similar specc'd' I assume similar hardware? Then the difference in performance would be due to one GPU running at a higher clock speed.

*sigh... GPU's have their own memory...

Let me be direct...

1000MHz Engine Clock
1GB GDDR5 Memory
1500MHz Memory Clock (6.0 Gbps GDDR5)
96GB/s memory bandwidth (maximum)
1.79 TFLOPS Single Precision compute power
GCN Architecture
14 Compute Units (896 Stream Processors)
56 Texture Units
64 Z/Stencil ROP Units
16 Color ROP Units
Dual Geometry Engines
128-bit GDDR5 memory interface
PCI Express 3.0 x16 bus interface


The bold things, I do not understand. I would like to. Every page I bring up is simply repeating that more is more, which doesn't matter to me. I cant figure out what htose things matter, and at what scale they matter.

Yes, clearly there ram modules on the mobo, and the gpu has its own ram. I want to know about these other things that apparently matter more now.

 

[smyler]

Look at memory bandwidth, amount of cuda cores (nvidia), or stream processors (amd), power consumptions, and vram (or memory). All of these are factors as the amount of informations in a frame rendered is dependent on the information being presented in each pixel. Resolution is partly a determining factor in this, however, the amount of AA, AF, Physx, lighting settings, etc... All determine what the application or game will use from the card. Also, the memory interface, ie. the bus width is important as well. You won't find most newer cards using less than 256-bit, and a lot of higher end cards use 384-bit or so, though it varies from card to card. However even the bud can be high, and a 256-bit card will outperform a 384-bit, because of the quality of the processors. A lot of people like nvidia, and a lot amd, however it's the same as amd vs intel cpu's, quality over quantity. The intel cpu is per core almost always 20-30% stronger on a single core than a comparable amd cpu. Thats why you see 4670k matching well vs 8350's. Same with gpu's. An amd card with 2000+ stream processors, can either match or be beaten by say a 770 with 1536 cuda cores. I hope this helps.

 

[smyler]

Oh, and you wonder about memory bandwidth as far as gb's a second. The one you stated is very under performing. GPU's now work in teraflops and computer thousands of times more information than the cards you were used to. Try and find something with say, atleast 160+ gbps memory transfer. Look up the specs of a 660ti which was the midgrade high performer a little while ago, and compare it to a 770 per say, you will see what I mean. Also, the higher clocked the memory is, the faster it will process the information, and the bigger the bus interface combined with the amount of information it sends through it per second ie. 160gbps for example, is like it says, the speed at which the information travels from the card to the mobo. Also important, is the MB's ability to transfer that information and compute it through the cpu. Most cpu's will handle it now a days, however i wouldnt go lower than a 3570k or so for intel as the amount of information it handles from the gpu low than that cpu is not so great. AMD, I'd say for the price to performance, get a fx-8120 and above or you're wasting money.

 

[cst1992]

Other things are there.
The Geforce 210 and the 650, for example, have 1GB each. But one is DDR3, other is DDR5, so 2x fast in that regard.
Other thing is that the 210 is 40 nm, and the 650 is 28 nm. So it is possible to fit more transistors in 1 mm2.
Third is the clock speed of each. The 210 is 589/600 core/mem speed whereas the 650 is 1058/5000 (in MHz).
Fourth is the cores, shaders, and ROPs. That is 16/8/4 for the 210 whereas 384/32/16 for the 650(although only 192 cores are considered as the 650 is the kepler architecture).
Fifth is the floating point performance. The performance for the 210 is 35.2 GFlops whereas the 650 is 812.5 GFlops.

 

[renz496]

The graphics card has memory.

The GPU is a component on the card PCB, separate from the memory modules.

By 'similar specc'd' I assume similar hardware? Then the difference in performance would be due to one GPU running at a higher clock speed.

*sigh... GPU's have their own memory...

Let me be direct...

1000MHz Engine Clock
1GB GDDR5 Memory
1500MHz Memory Clock (6.0 Gbps GDDR5)
96GB/s memory bandwidth (maximum)
1.79 TFLOPS Single Precision compute power
GCN Architecture
14 Compute Units (896 Stream Processors)
56 Texture Units
64 Z/Stencil ROP Units
16 Color ROP Units
Dual Geometry Engines
128-bit GDDR5 memory interface
PCI Express 3.0 x16 bus interface


The bold things, I do not understand. I would like to. Every page I bring up is simply repeating that more is more, which doesn't matter to me. I cant figure out what htose things matter, and at what scale they matter.

Yes, clearly there ram modules on the mobo, and the gpu has its own ram. I want to know about these other things that apparently matter more now.

if your concern is gaming performance only then there is no need to know about them in detail. what really matters were real world performance. that's why there are review to test how the device will perform in real world application.

 

[smyler]

Other things are there.
The Geforce 210 and the 650, for example, have 1GB each. But one is DDR3, other is DDR5, so 2x fast in that regard.
Other thing is that the 210 is 40 nm, and the 650 is 28 nm. So it is possible to fit more transistors in 1 mm2.
Third is the clock speed of each. The 210 is 589/600 core/mem speed whereas the 650 is 1058/5000 (in MHz).
Fourth is the cores, shaders, and ROPs. That is 16/8/4 for the 210 whereas 384/32/16 for the 650(although only 192 cores are considered as the 650 is the kepler architecture).
Fifth is the floating point performance. The performance for the 210 is 35.2 GFlops whereas the 650 is 812.5 GFlops.

The graphics card has memory.

The GPU is a component on the card PCB, separate from the memory modules.

By 'similar specc'd' I assume similar hardware? Then the difference in performance would be due to one GPU running at a higher clock speed.

*sigh... GPU's have their own memory...

Let me be direct...

1000MHz Engine Clock
1GB GDDR5 Memory
1500MHz Memory Clock (6.0 Gbps GDDR5)
96GB/s memory bandwidth (maximum)
1.79 TFLOPS Single Precision compute power
GCN Architecture
14 Compute Units (896 Stream Processors)
56 Texture Units
64 Z/Stencil ROP Units
16 Color ROP Units
Dual Geometry Engines
128-bit GDDR5 memory interface
PCI Express 3.0 x16 bus interface


The bold things, I do not understand. I would like to. Every page I bring up is simply repeating that more is more, which doesn't matter to me. I cant figure out what htose things matter, and at what scale they matter.

Yes, clearly there ram modules on the mobo, and the gpu has its own ram. I want to know about these other things that apparently matter more now.

if your concern is gaming performance only then there is no need to know about them in detail. what really matters were real world performance. that's why there are review to test how the device will perform in real world application.

This is true, however you have to take reviews, and benchmarks with a grain of salt and average your findings. The problem is, that not only do tests happen mostly in single player, and even if in MP, they are dependent on the components used in the testers computer. Also, the temp's that are provided are important as components almost always slow down at higher temps. Overclocks, and even monitors play a role in buffering frames from a gpu to the monitor. The response time in comparison to how fast a gpu gets a frame ready. Even the ability of a cpu to compute the information received from the gpu. Reviews are a good reference point, but as you stated, real world performance is more important, and that varies from person to person and how there system is setup hardware and software wise.

 

[Omegaclawe]

While they won't really matter, for academic purposes:

Engine Clock is the frequency the GPU itself runs at, similar to CPU clock rate.
Memory Clock is the speed the VRAM runs at. The faster this is, the more memory can be moved at once. This is actually often a limiting factor of a GPU at high resolutions. GDDR5 can run in dual or quad modes, which is why it also lists the 6.0Gbps, since this is running quad.
Memory bandwidth is the effective memory clock multiplied by the memory bus width. Higher bus width or higher memory clock makes this number bigger. At high resolutions, this number alone can differentiate some cards, but you can't rely on it.
Compute units is just a fancy AMD term for a block of GPU "Stream" Processors. There are 64 stream processors per compute unit, and each stream processor can do a "thing" per clock (to grossly oversimplify). Extra Compute units are almost like "Cores" for a CPU, except they can do the same thing in bulk (in this case, 64 times per clock) with each stream processor.
Texture units are "cores" of a processor set up specifically to handle texture-based math as efficiently as possible. More of these means more texture processing.
ROPs, or Raster Operators, tend to do per-pixel operations, whether this is post-processing or determining what color each pixel should be from the mess of 3D objects with transparency, lighting, bloom, etc. that the rest of the GPU generates.
Not sure on Dual Geometry Engines past AMD marketing buzzwords. You see it a lot but never and explanation of what it means. Possibly it has the capability to render two unrelated scenes at once or something.

As to the "scale" at which they matter, it's safe to say at a very minor one (except Memory Bandwidth). They tend to be designed with as much balance as possible to prevent any transistors from idling. In the original GPU's, You would have something like a compute unit attached to exactly one texture unit with exactly one ROP on the end, but that left the GPU sitting idle much of the time. Additionally, Texture Units and ROPs are more or less useless for direct compute functions and wot. The reason why they don't matter is because there is a massive difference in the efficiency of these parts between GPU's, and while these parts are the reason why some games favor some GPU's over others, nothing in these numbers can actually tell you which GPU is better.

Though if there's a significant difference in Memory Bandwidth, that's a pretty good tip that one of those chips is considerably faster. You'll still have to look at benchmarks to know how much, though, so honestly, you may as well start there are save yourself time, and treat those other numbers like the marketing buzz they are these days.