384-bit Memory Interface for HD5000-series?

[Zero82z]

well then you should stop talking also.

I can at least show you how it has worked up until now where all you have done is pull a theory out of your lower region. nothing supports what you have theorized at all where as I can look at the history of every card made in the last 5-7 years and see a pattern or even formula if you will.

Again, the fact that something hasn't been done before makes no difference to the possibility of it being done in the future. So unless you can give me some actual technical reasons why what I have suggested is impossible, there is nothing to support that claim. I admittedly have nothing to support my claim either, but that doesn't mean that it isn't possible (nor am I claiming that it is definitely possible or impossible).

 

[necrophile]

Do you guys have me on ignore or something.

 

[Zero82z]

Do you guys have me on ignore or something.

No, I've just elected to not respond to your posts since they're completely off-topic and not relevant to the discussion in this thread.

 

[necrophile]

So I give an explanation as to why bus widths and all numbers in a computer system are why they are and you ignore them because they refute your position?

 

[Zero82z]

So I give an explanation as to why bus widths and all numbers in a computer system are why they are and you ignore them because they refute your position?

No, I ignore them because they're wrong. There's no real reason why absolutely everything has to be a multiple of two. Take triple-channel RAM as an example. Just saying that things involving computers are usually powers of two doesn't refute anything I've said. Most of the time, that's simply because of convention, not necessity.

 

[necrophile]

Lol, you have got to be kidding.

Just because something has tri in it does not mean it has anything to do with 3 except in name.

dual channel is two 64 bit channels seen as one. 1 channel plus another not times.

Tri channel will be 3 bus widths(not too sure on the exact number for one) however, it is not a multiple of 3, it is 1+1+1. The bus width per channel is still a binary number.

Seriusly now, this is a tech site.

Everyone here does know that when it comes down to it, computers work by processing 1's and 0's right?

it is just transistors switching states that is all a computer is when you get down to it.

You do realise this yeah?

 

[Zero82z]

Tri channel will be 3 bus widths(not too sure on the exact number for one) however, it is not a multiple of 3, it is 1+1+1. The bus width per channel is still a binary number.

So 1+1+1 is not equal to 3? Wow, I must be really bad at math then . In triple-channel, the total effective bus width is 192 bits, which is, you guessed it, a multiple of 3.

Everyone here does know that when it comes down to it, computers work by processing 1's and 0's right?

it is just transistors switching states that is all a computer is when you get down to it.

You do realise this yeah?

I think you don't really know what you're talking about.

 

[bigdogchris]

If priced right this card will be an epic success for ATI.

Since when do GPU manufacturers EVER ship GPU's at a killer price that makes it a must buy? Hell, Nvidia makes the best cards yet refuse to sell them at a competitive price point. Nvidia could own the market if they really wanted.

 

[necrophile]

Okay, so multiples was the wrong word, try powers and get back to me.

 

[Zero82z]

Okay, so multiples was the wrong word, try powers and get back to me.

192 isn't a power of two, and neither is three. By the way, every power of two is a multiple of two, so if there's no requirement for everything to be a multiple of two, then there cannot be one for powers of two either. So again, you're wrong.

 

[necrophile]

You are trolling now.

Do you not understand, computer work on the principle of 1 and 0, on and off. Everthing has to be a power of two.

A byte is 8bits of data, a bit being represented by either 1 or 0.

If you want a different number then you add them up.

1010=10.

See.

192 is 128 plus 64, 2 to the power of 6 plues 2 to the power of 7 or 2 tot he power of 6 plus 2 of the same.

That is why the address range of 32bit computers is limited to 4gb, 4GB= 2 to the power of 32.

 

[ihira]

Well, the fact that a 256bit + 768MB card doesn't exist to this date is probably a good indication that it doesn't work that way no?
Unless they change the way video memory controllers work (which I dont know in depth) I doubt this formula is going to change.

That being said the 384bit seems a bit 'weird' because we're stuck with 768MB or 1536MB. The earlier is lower than todays enthusiast standard sand the latter seems a bit overkill.

 

[Zero82z]

You are trolling now.

Do you not understand, computer work on the principle of 1 and 0, on and off. Everthing has to be a power of two.

A byte is 8bits of data, a bit being represented by either 1 or 0.

If you want a different number then you add them up.

1010=10.

See.

192 is 128 plus 64, 2 to the power of 6 plues 2 to the power of 7 or 2 tot he power of 6 plus 2 of the same.

That is why the address range of 32bit computers is limited to 4gb, 4GB= 2 to the power of 32.

192 may be the sum of two numbers that are powers of two, but then all even numbers are as well. And the number 192 itself is not a power of two. So you are wrong. There's no two ways about it. I am quite familiar with how a base-2 number system works, and I know why the 32-bit memory address limit is the way it is. However, it's foolish to think that everything inside a computer is determined by powers of two. If that were the case, then wouldn't it be impossible for a computer to process anything involving odd numbers? I know that's a silly example, but it's the most straightforward and obvious one I can think of.

Well, the fact that a 256bit + 768MB card doesn't exist to this date is probably a good indication that it doesn't work that way no?

Correct, the memory size has to be a multiple of the bus width. However, 1152 is a multiple of 384. The issue is that some people don't believe that it's possible for a video card to address a memory size that's equal to an odd multiple of its bus width, but so far nobody has come up with any actual evidence to support that notion.

 

[necrophile]

The basic operation of a computer is 1 and 0's. the hardware has to follow this as everthing is sent as that, software that is programmed can be used to represent other things, hell no one actually programs in 1 and 0's or so I believe unless you need maximum power out of hardware and have a hell of alot of time although I ain't a programmer.

Prove me wrong here, stating I am wrong through flawed logic and trying to use your opinion to change the fundamental way semiconfuctors work is not acceptable.

 

[Zero82z]

The basic operation of a computer is 1 and 0's. the hardware has to follow this as everthing is sent as that, software that is programmed can be used to represent other things, hell no one actually programs in 1 and 0's or so I believe unless you need maximum power out of hardware and have a hell of alot of time although I ain't a programmer.

Prove me wrong here, stating I am wrong through flawed logic and trying to use your opinion to change the fundamental way semiconfuctors work is not acceptable.

Fuck it, I'm talking apples and you're talking oranges. I'm not going to bother arguing with you any more since what you're saying has very little to do with what I'm talking about.

 

[necrophile]

No, the only reason they are a multiple of each other is because they are both binary numbers!!!!!!!

There is no direct correlation between bit width and memory amount, sorry if I have said something to indicate that or seemed to be in agreement with other posters.

Memory bus and the speed of the memory is what is actually of importance, these two determine the bandwidth of the memory system.

If you require the gpu to be fed with data at a certain rate then you need a certain bandwidth, ie you need the data to be pushed odwn a certain number of paths at a certain speed.

Memory size is nothing to do necessarily with the bus width, it is to do with the amount need to store locally all the information needed for the gpu to process a frame quickly. Again, large amounts of memory which does not have the bandwidth will create a bottleneck so that is why you either have to increase the speed of the ram or the bandwidth when using high amounf of memory other wise it is a gimmick like all those 1GB 9500gt's with 128 bit bus's.

I ain't lying here, just think about this logically.

 

[necrophile]

also remember what I said about dual channel, it has some worth here.

Instead of think about the amount of memory as one lump think of it as how you make numbers using binary.

It is always, always powers of two so if you have a 384 biut bush and 768 of ram it is best to think of it as 256 + 128 and 512MB + 256MB for each bus.

 

[Krenum]

Since when do GPU manufacturers EVER ship GPU's at a killer price that makes it a must buy? Hell, Nvidia makes the best cards yet refuse to sell them at a competitive price point. Nvidia could own the market if they really wanted.

Since the AMD/ATI 4870. Nvidia could own the market , but they don't.

 

[necrophile]

Please remember that there are costs involved and how much choice nvidia and there partners have on what to sell them whilst still making a profit is unknown. Indeed it has been speculated that nvidia especially may have been having to see their top ends at a loss at least when the cards were new out.

Just rumour of course unless someone has substantiated it. That is what annoyed so many people either nvidia is now selling the cards at a loss or if they are making a profit then the price that the 200 series was orginally at was a kick to the teeth for the public.

 

[necrophile]

Ninjaedit's don't work with e-mail notifications.

 

[Sojuuk]

When you find a video card with an amount of memory totaling 3x bus width.... I will eat it.

I don't know why it hasn't been done but it hasn't. Leads me to believe that it can't or shouldn't be done.

 

[SonDa5]

Since the AMD/ATI 4870. Nvidia could own the market , but they don't.

Good response.

I'll add a some comments to build on your post.

I think ATI may surprise the world by making the pricing much more affordable than what NVIDIA has done in the past for the following reasons:

1. Economicly speaking times are tough but people still want the latest and greatest technology to play games.

2. Building good will by offering great pricing on the latest technology is bound to win a whole bunch of new ATI/AMD fans which will also help out ATI/AMD as a whole.

3. Value is going to be the win win for ATI/AMD.

 

[Valset]

Good response.

I'll add a some comments to build on your post.

I think ATI may surprise the world by making the pricing much more affordable than what NVIDIA has done in the past for the following reasons:

1. Economicly speaking times are tough but people still want the latest and greatest technology to play games.

2. Building good will by offering great pricing on the latest technology is bound to win a whole bunch of new ATI/AMD fans which will also help out ATI/AMD as a whole.

3. Value is going to be the win win for ATI/AMD.

I hope you are right because nvidia is looking to be a bit behind and I think I am going to have to bite on one of these if its half as good as its suppose to be.

 

[necrophile]

I don't think ATI have too much choice in regards to pricing. If they have high prices then they will look as bad as nvidia and I don't think they will make the same mistake.

 

[e-geek]

I would think 1152 would be possible, couldnt 896 + 256 = 1152

I mean the gtx260 was 896mb then you couldnt you just add 256mb?

Any way 3 128bit channels so 1+1+1(384bit), 384mb per channel = 1152mb (384 is just 128 + 128 + 128 or 256 + 128 etc) So tri channel ram on x58 is 64 + 64 + 64 = 192 bit, so just like the dual channel example - you could put 128 + 256 in each of the 3 channels to equal 384mb per channel (in tri channel) and if the 5870 is 128 + 128 + 128 it should work.

There, it makes sense to me anyway. I totally agree with zero it should be possible, im not saying it is possbile but theoretically it should be.

 

[necrophile]

Yes you are correct there.

The real question is why would you.

Zero got that number by not realising things work in binary. you do not use powers of 3 in computers.

When it all come down to it, ram chips are made of powers of two.

All information is done in essentially bits, then bytes, then multiple's of bytes. Whilst the total number of ram may not be a power of two, they are made up of numbers that are powers of two.

How can you expect a computer to address ram sizes that are not powers of two?

Will data just get cut off or what?

Lets not get him started on it again, the whole debacle of the discussion was because of a lack of knowledge about how computers work, christ I am newb but I know how the basic of computers use a transistor either being in an on of off state.

 

[necrophile]

Does anyone know the base capacity of the ram chips used in gfx cards. There is only so mcuh room for placing them and so How much you use must be taken into account no?

 

[BinarySynapse]

Memory manufacturers make 8, 16, 32, 64, 128, etc (power of two) megabit chips. Those chips have 8,16,32, 64, etc (again powers of two) bit outputs. There is no physical or technical limitation at this point, it just easier and cheaper to make them this way.

Using Nvidia as an example, a 384-bit bus using 12 chips would mean that each chip outputs 32-bits each. So using memory chip size of 512Mbit, you'd get a maximum of 768MB, Going to the next size up (1024Mbit) would yield 1536.

The total size of the memory depends strictly on the chips used, the bus width of the system, and how many banks the memory controller can address. All chips in the same bank have to be the same size, so if the memory controller can address at least two banks of RAM there is no reason you can't use six 512Mbit and six 1024Mbit chips to yield 1152MB.


So it's not impossible at all to have 1152MB on a 384-bit bus, but it really depends on how the memory controller is designed. If it can only address a single bank of ram, then you'll be stuck with [number of chips required to fulfill bus width * chip size]MB of RAM.

 

[Zero82z]

Zero got that number by not realising things work in binary. you do not use powers of 3 in computers.

You have no fucking idea what you're talking about, nor about what I said. You clearly didn't read any of my posts, or you just aren't educated enough to understand them. But don't say that I don't understand how binary works, because I'm quite certain that I have a better understanding of it than you do. The fact that transistors are based on an on or off state at a fundamental level does not mean that a computer is unable to deal with numbers that are not powers or multiples of two. And if you can't understand that, then there's no hope for you.

 

[necrophile]

You do not understand damn all.

Computers do not understand anything, they are machines. They can only do what they are programmed to do. Any advanced calculations have to either be programmed for or designed for.

Hardware does not understand anything at a basic level, a signal is sent which is either on or off. That is it.

We are talking about hardware at the most basic levels and they have to be arranged for data to be stored and transmitted. That is bits of data.

You don't seem to understand that because something in a computer can be divided by three it does not mean they operate by three's or any other number.

The base number a computer works by is two.

THis whole discussion came about because you thought there had to be a relationship between bus width and ram amounts. Instead of just saying to cannondale, oh my bad you started arguing about BS and even when I pointed out the reason for why computers use such numbers you just say, you're wrong without any reasoning behind your comment.

Again, if a computer communicates at the basic level in binary, then what other number is the base?

 

[Zero82z]

We are talking about hardware at the most basic levels

That's where you're wrong. We're not talking about individual transistors here, we're talking about complex integrated circuits which are designed in such a way that they don't have the same limitations as individual transistors do.

You don't seem to understand that because something in a computer can be divided by three it does not mean they operate by three's or any other number.

I never said that they operate by threes. What I did say is that they are not limited to dealing with things that are only powers of two.

Anyway, I'm wasting my time here. I'm not going to bother any longer.

 

[necrophile]

You do that and keep believing what you want and I'll do the same.

I at least have tried to just tried to justify my position, wrong or right I have tried, you have just hid behind ignorance and wallowed in it.

 

[Krenum]

Good response.

I'll add a some comments to build on your post.

I think ATI may surprise the world by making the pricing much more affordable than what NVIDIA has done in the past for the following reasons:

1. Economicly speaking times are tough but people still want the latest and greatest technology to play games.

2. Building good will by offering great pricing on the latest technology is bound to win a whole bunch of new ATI/AMD fans which will also help out ATI/AMD as a whole.

3. Value is going to be the win win for ATI/AMD.

Indeed. AMD/ATI get it. They have excellent prices and hardware to back it up. That is the most basic in business philosophy. Make a quality product that is affordable and it will sell itself.

I wish that Nvidia would lower there prices and compete with AMD/ATI a little more. Competition is good for the technology community, without competition things become stagnant. Thats bad.

 

[necrophile]

This will be intested due to the complete reverse of fortunes.

With the 4 series nvidia were caught with their pants down pricing wise and so had to rashly respond. This time it is looking like they will be second which will allow them time to figure out pricing.

Whether it will be long enough to adjust clocks or other specs to deliberately carve out a price point I am not sure but this delay could be used to their advantage.

 

[Valset]

um guys on the subject of memory and bus width I think your both kinda right here, there is nothing preventing them from putting an odd number of chips or even of using different sets of density chips but it is highly inconvenient to do so and usually incurs a performance hit for doing so.

 

[LAiN_OfTheWired]

I think its just ridiculous they get so hot-headed over something that's pure speculation, especially considering neither of them is an engineer of any kind.

While the 384-bit memory bus is intriguing, no doubt ATI has been keeping tabs on the would-be monster GT300. I wouldn't be surprised if the initial batch of cards were either 256 or 384, but I believe the HD5000-series will see a jump to 512 before ATI goes on to whatever will be HD6000-series (or whatever the successor series will be named). They've gotten progressively more comfortable with the enthusiast market again. First the X2 series cards, now the 4890 cards, why not match Nvidia at the top-end with 512-bit interface on GDDR5 and show them frighteningly high-clocked+binned R800?

 

[Biskquik]

Intel uses odd size L2 caches on some of their CPUs. E7000 series and Q9000 series actually use a factor (not base) of 1.5MB for their cache size. This is on a completely shared system where both cores will access the 3MB so you cannot make the argument that it's 1MB plus 512KB for each core in the example of a E7000 chip.

Hell, Intel puts out 80GB ssds and those use flash so they fall under the realm of a binary "convention".

I honestly do not see why AMD cannot release a card with 1152 MB memory configuration. It actually seems somewhat possible either by utilizing some 3 channel by 128-bit architecture where each channel accesses 3 128MB memory chips. To me, it seems somewhat ideal to do that as cooling for 9 memory chips totaling 1152MB seems easier to control over 12 memory chips in a 1536MB config. AMD could possibly use 256MB chips (I'm not quite sure about the max density of GDDR5 off the top of my head) and use it in a 6 chip config but I'm sure that would hike prices.

 

[necrophile]

It is the fact that he was so dismissive of my reasoning when in truth he had no reason to think otherwise. I would rather admit we were both wrong than even to suggest for a second he was even partly right. Yes I am that bloody minded.

Also, ATi did have the GDDR4 512bit bus on the 2900xt didn't they and they decided at least then it was unnecessary. Maybe they will bring it back, who knows.

 

[Sojuuk]

Intel uses odd size L2 caches on some of their CPUs. E7000 series and Q9000 series actually use a factor (not base) of 1.5MB for their cache size. This is on a completely shared system where both cores will access the 3MB so you cannot make the argument that it's 1MB plus 512KB for each core in the example of a E7000 chip.

Hell, Intel puts out 80GB ssds and those use flash so they fall under the realm of a binary "convention".

I honestly do not see why AMD cannot release a card with 1152 MB memory configuration. It actually seems somewhat possible either by utilizing some 3 channel by 128-bit architecture where each channel accesses 3 128MB memory chips. To me, it seems somewhat ideal to do that as cooling for 9 memory chips totaling 1152MB seems easier to control over 12 memory chips in a 1536MB config. AMD could possibly use 256MB chips (I'm not quite sure about the max density of GDDR5 off the top of my head) and use it in a 6 chip config but I'm sure that would hike prices.

cache is fundamentally different from RAM.

as I understand it ssd's use what is effectively multiple built in raids so that 80 could be any number of different size combinations and SS storage is not the same as RAM.

3 x 128bit channels seems plausible. But you still couldn't use 3mb x3x128bit to achieve 1152.

 

[Biskquik]

Actually cache isn't all that fundamentally different than RAM aside from cache being SRAM and the DIMMs we stick in our motherboards being DRAM, though in some embedded systems (i.e Nintendo Wii), SRAM can be used as off chip memory. It's still primary storage, it either gets a hit or a miss when attempted to being accessed. Now, I'm not factoring things like cache-coherency or associativity but those aren't attributes that should be used to dictate its size, it's still RAM.

Now for ssds, yes, the chips are RAIDed in a fashion of most likely 10*8GB. The problem is that the number of chips is still not a "nice" number, which seems to be the prevalent argument in this thread. If the controller on the SSD can support a non base two number of flash chips, what is really the limitation of developers making a MMU that accesses a non base 2 number of chips.