AGP, PCI-E, PCI? Differences?

[eidairaman1]

well the 3850 is the Fastest Card available, and it does perform the same as the PCI E variant.

 

[kiriakost]

What i expect to see before i die, in 30 years or so ,
is one software to measure PCI-e bandwidth.

The results will be good for laughs , considering the fact, that in order to use all the bandwidth, the textures of the game , must have the size of an elephant .


Dear PCI-e fanatics .... this message of my, called as " hard landing " ..

 

[kiriakost]

AGP was one slot per board, with an independant bus purely for video cards.

PCI and PCI-E are designed for anything you want, Video cards, RAID cards, network cards, sound cards.

AGP was also limited power-wise, i'm sure you wouldnt like the amount of connectors a modern card like an 8800GTX or 4870x2 would need, if it was still on AGP.

Oh and dont forget, without PCI-E theres no SLI or crossfire.

1) Wow who needs dual 6800GT , when you can have one HD3850

2) Why you need 4GB or more bandwidth for RAID , when your hard drives the most that can do is 80 - 120 megabytes /S both .

3) I am done here ...

 

[eidairaman1]

They all say Theoretical Bandwidth, now certain boards will be quicker than others due to quality of components used, its all about moving electrons around along with DC waves.

 

[kiriakost]

They all say Theoretical Bandwidth, now certain boards will be quicker than others due to quality of components used, its all about moving electrons around along with DC waves.

Thats true , exclude Nvidia s chip sets , they do more harm than good.

 

[eidairaman1]

eh IDK, they are pretty good, and i was talking about motherboards not Graphics cards in general, I am using an NV Chipset with an AMD vid card.

 

[Mussels]

kiriakost: you obviously convinced yourself you're right, but are you aware this has been tested?

"who would go two 6600GT when they can have one 3850"
No idea. the 3850 is fairly weak in comparison to my 1 year old (and then some) video cards. you can buy faster for very little, in PCI-E

modern video cards most certainly were pushing the bandwidth required from PCI-E 1.0/1.1, particularly due to motherboards doing SLI and crossfire in 8x/8x configurations.

you can spout as many theories as you want, but until you actually have some proof of what you're saying, i doubt anyones going to beleive you. (And what you're saying seems to be "I'm still on AGP so AGP is awesome, hahaha im so glad i didnt spend money on PCI-E"

 

[eidairaman1]

Mussles, its worth maxing out the AGP slot, but that is after building a new machine, For me Im going to move up and keep this machine as is because at the time the 1950 Pro was the fastest AGP card, no one was expecting AMD at the time to release the 3850 as an AGP part.

 

[Mussels]

Mussles, its worth maxing out the AGP slot, but that is after building a new machine, For me Im going to move up and keep this machine as is because at the time the 1950 Pro was the fastest AGP card, no one was expecting AMD at the time to release the 3850 as an AGP part.

dont get me wrong, its not like i think AGP is useless, or incredibly slow. AGP has hit its limits with the 3850, and i just get annoyed at people who spout things with no proof of backup, when they contradict common thinking. If you want to change peoples minds, you need more than a theory.

all video cards in DX9.0c (and under) require the video cards ram to be synced with the system ram (it duplicates), therefore a slot needs to be as fast as the ram on the video card, to 'max out' its performance (odds are the GPU or something would hinder it before here, but you get the point), slap a GDDR5 card into an AGP slot and watch it choke, just like modern cards (HD2400, Nv 8400) do in PCI

 

[r9]

Putting it simply every card that is made for AGP is as fast as PCIE counterpart. I`m talking for AGP X8.

 

[eidairaman1]

well i doubt AMD will be releasing anything faster than the 1950 or 3850, but it is sweet to see an AGP card take down a PCI Express card. News of Driver reduction (quarterly Driver releases instead of monthly for X1K and lower) I doubt anything will be faster, but since the OP is making the P4 his main machine, I dont see why not max out the Slot. For me I learned something on this machine, Project Torque Lags real bad with the CPU at 1.8 GHZ (133x14) I overclocked the System to 2.2GHz ~ 400 MHZ boost (200x11) and i had no more problems, it just proves that the CPU was the bottleneck for my setup). Now i probably wouldnt upgrade to a 3850 Unless if i could get this machine at 2.4 GHz or Higher Stabily (CPU is cool, Mobo isnt).

 

[kiriakost]

kiriakost: you obviously convinced yourself you're right, but are you aware this has been tested?

"who would go two 6600GT when they can have one 3850"
No idea. the 3850 is fairly weak in comparison to my 1 year old (and then some) video cards. you can buy faster for very little, in PCI-E

modern video cards most certainly were pushing the bandwidth required from PCI-E 1.0/1.1, particularly due to motherboards doing SLI and crossfire in 8x/8x configurations.

you can spout as many theories as you want, but until you actually have some proof of what you're saying, i doubt anyones going to beleive you. (And what you're saying seems to be "I'm still on AGP so AGP is awesome, hahaha im so glad i didnt spend money on PCI-E"

Step down from the broom, its not a horse , i can teach you more ,
if you are willing to learn .

 

[Error 404]

Putting it simply every card that is made for AGP is as fast as PCIE counterpart. I`m talking for AGP X8.

Not so; for example, the 7800 GT AGP is crippled by nVidia. It performs nowhere near as well as a PCI-E 7800 GTX.
Also, all nVidia cards since the 6x00 series have had to use an AGP - PCI-E bridge chip, because they were all native PCI-E cards. Theoretically, this would impact on performance due to the slight delay in converting AGP to PCI-E signal.

Also, kiriakost; The textures are not "as big as an elephant", they are comparatively small; there are just THOUSANDS of them, and they're not being loaded off of your hard drive, they're being transferred from your RAM to the PCI-E slot, so no hard drive bottlenecks from that except for when you are loading a level or something.

AGP is slow, but its not terribly slow. It is slow enough, however, to bottleneck many modern cards. Do you think a GTX 295 would perform as well as it could in an AGP slot compared to a PCI-E x16 2.0 slot? I think not!

 

[eidairaman1]

there was no 7800GTX AGP, there was 7800GS, Also on another Point, any card designed for the x16 slot (4850/4870, GF GTX 260/280 and higher) if you drop it in a PCI E 4x slot, the performance drop is considerable because some think PCI E 4x is AGP 8X. PCI E 8x and higher is what is need to not see a bottleneck for the fastest cards out.

 

[Error 404]

there was no 7800GTX AGP, there was 7800GS

My mistake, it must'v been the GT.

 

[eidairaman1]

There was only 1 AGP 7 Series GT card for AGP and that was the XFX Only 7950GT. Everything else has been GS.

 

[Mussels]

the 7900GS in AGP could be unlocked to a GT through rivatuner, its probably why you're getting confused.

the PCI-E 7900GS couldnt be unlocked.

 

[kiriakost]

Also, kiriakost............... Do you think a GTX 295 would perform as well as it could in an AGP slot compared to a PCI-E x16 2.0 slot? I think not!

One card connected in transport channel , as the AGP ot PCIe, it does not perform ... it moves data.
If the channel is empty , as i clearly proved above about AGP bandwidth ...
there is no issue at all , no matter how the channel are named .

Its like having a very small car , in a highway with 8 free lines ... no problem .

The point is that we still have small cars compared with the size of the highway ...
and no VGA , even the latest VGAs, are still like small cars, in comparison with the size of the highway .

This is the message , that i like to become , commonly understood !!!

 

[KainXS]

the 7950GT AGP was one of the worst cards I have Ever, EVER owned, should be avoided at all costs, they die like crazy

 

[eidairaman1]

Gen 1 X1950 Pro AGP from Sapphire were having Overheating Problems, It was blamed on QA of the cards, like heatsinks not fully attached even not contacting components, second Gen also as the 1950GT were perfectly fine (want to replace the fan on mine because its a bitch to clean)

 

[Ketxxx]

I done a whitepaper for AGP and PCI-E, hers a copy and paste job from both;

Originally designed for the Pentium 2 platform, the AGP bus has undergone a multitude of evolutionary steps beginning with just AGP 1x (266MB\s) and ended with 8x (2.1GB\s) using the PCI specification as an operational baseline, the AGP specification adds 20 additional signals not included in the PCI bus.

The aim of the AGP bus is to provide a much smoother framerate and 3D rendered image in a much higher detail level than previously seen before on the PCI or ISA bus standards of old. The AGP bus is able to transfer high amounts of data due to it being able to transfer on both the rising and falling edges of the 66MHz bus clock frequency. AGP allows for direct data transfer between the graphics card and the CPU and \ or system memory. Although AGP is an extension of the PCI interface, it is completely (physically, logically and electrically) separate from the PCI bus. Therefore, activity of PCI peripherals won't affect the AGP card's performance.




AGP Technical Sheet:

Maximum transfer rates (32bit): 266MB\s (1x) 533MB\s (2x) 1066MB\s (4x) and 2100MB\s (8x)
Operating Frequency: 66MHz
Pipelined Requests
Address \ Data de-multiplexed
Single target, Single master
Memory Read \ Write only, no other I/O operations
High \ Low priority queues
Sideband Addressing
Fastwrites

Explanation of DIME

Direct Memory Execute (DIME) is probably the most important feature of the accelerated graphics port. AGP graphic chips have the capability to access main memory directly for the complex operation of texture mapping. AGP provides the graphics card with two methods of directly accessing texture maps in system memory: pipelining and sideband addressing. In pipelining, AGP makes multiple requests for data during a bus or memory access. Sideband addressing offers the highest level of AGP performance. In addition to allowing multiple outstanding transactions and non-coherent access to main memory, Sideband Addressing introduces a separate address \ command bus, the Sideband Address Port (SBA). Because the SBA and data buses are not multiplexed, the graphic controller can use the SBA to make data requests without interrupting the data bus.

During the lifetime of the AGP bus the port underwent a multitude of evolutionary steps forward as briefly mentioned earlier, here is a rundown of the revision history.

AGP 1.0 (1996) specification defined 1x and 2x speeds with the 3.3v keyed connector.

AGP 2.0 followed shortly after (1998), the specification defined 1x, 2x and 4x speeds with the 3.3v, or 1.5v keyed connector or a 'Universal' connector which supported both card types.

AGP Pro specification defined 1x, 2x and 4x speeds with the 3.3v, or 1.5v keyed connector or a 'Universal' connector which supported both card types.

The AGP 3.0 specification followed 4 years later (2002) and was the final evolutionary step in AGPs life, the specification defined 1x, 2x, 4x and 8x speeds with the 1.5v keyed connector, or a 1.5v AGP Universal / Pro connector. Each upgrade was a supper-set of the 1x mode, (4x would also support the 1x speed etc.) The base clock rate is 66MHz, but to achieve 2x, 4x, and 8x speeds the clock is doubled each time.


AGP (1x): 66MHz clock, 8 bytes/clock, Bandwidth: 266MB/s [3.3V or 1.5V signal swing]

AGP 2x: 133MHz clock, 8 bytes/clock, Bandwidth: 533MB/s [3.3V or 1.5V signal swing]

AGP 4x: 266MHz clock, 16 bytes/clock, Bandwidth: 1066MB/s [1.5V signal swing]

AGP 8x: 533MHz clock, 32 bytes/clock, Bandwidth: 2.1GB/s [0.8V signal swing], still uses 1.5 volt mainboard power.

The AGP data bus may be 8, 16, 24, 32, or 64 bits. Due to timing requirements the maximum bus length is 9". The trace impedance is specified as 65 ohms +/- 15 ohms (no termination resistor is specified). For the 8x speed the bus requires a parallel termination or 50 ohms. Some lines may require a Pull-Up Resistor to insure the lines come out of reset in the proper state.

----------------------------------------------------------------------------------------------

Like Hypertransport technology, PCI-E is a 2-way serial connection that carries data in packets, similar to the way it is transferred over Ethernet connections.

The PCI-E bus is an assembly of serial, point-to-point wired, individually clocked “lanes”, each consisting of two pairs of data lines carrying data upstream and downstream. Each individual “lane” is capable of 250MB\s, giving a PCI-E x16 slot a theoretical maximum of 4GB\s bandwidth. Real performance benefits come in when more than one lane is added to a given point-to-point route. Lanes can be stacked together to increase the amount of bandwidth available to specific areas of the I/O system, such as the video card slot.

Using 4 routes instead of 32 for a basic connection significantly cuts the costs of producing motherboards, since each lane is exclusively used for communication between two points, there is no sharing of the available bandwidth either.


Introduction To SLi (Scalable Link Interface)

SLi technology allows two PCI-E x16 slots to split 16 lanes into two sets of 8 and run two identical video cards in tandem. This often leads to dramatic boosts in 3D performance. Latest chipsets such as the nVidia nForce4x16 are capable of 2x16 lanes for graphics instead of 2x8 lanes, though performance increases are minimal and arguably well within the margin of error.

The current standard PCI Express implementation allows for up to 40 lanes in total. 32 of these are used for the PCI-E x16 graphics slots and eight more are distributed between any combination of PCI-E x1, x2 or x4 slots.

PCI Express connectors are similar in appearance and connection method to 32-bit PCI slots. PCI Express 1X slots are about the size of current modem riser slots (about 1" long), while the X16 interface (164-pins) for graphics is very similar in appearance to the standard AGP port. The flexibility to adapt to PCI Express devices of different bandwidths is built into the X4 and X8 slots.

The Technology Behind The Technology

The physical communication method is Low Voltage Differential Signaling (LVDS). Differential Signaling (DS) uses a wire pair to represent logic levels as opposed to the single ended (SE) methods used by legacy hardware. The reason for using DS is that of signal integrity at very high baud rates (level changes per second).

The reasons for this are mainly capacitance between tracks (which opposes a rapid change in voltage) and slew rate (essentially how quickly the driving amplifiers can change voltage levels).

With DS this has much less impact. It doesn’t matter if the voltage levels are reduced in amplitude (within reason), as long as there is a difference for the difference amplifier to work with.

With DS the intended signal (bottom) is converted into the DS (middle) and transmitted along the wire pair into the difference amplifier at the other end (top). The difference amplifier then recovers the original signal by comparing the two voltages.

DS also offers a immunity to interference, which SE transmission does not. Suppose in transmission electromagnetic interference (EMI) caused a large voltage spike in the DS. Because the wire pairs always run in parallel very close to each other the spike will be present in both wires. Again (within reason) the signal can be recovered.

This more advanced signaling is needed due to the very high transmission frequencies necessary in replacing a large parallel interface with a narrow serial interface, for example AGP to PCI-E.

To reduce inter symbol interference at the high signaling rates a method called 8B/10B is used to encode the serialized data on each lane. This represents 8-bits as 10-bits to provide a net DC ‘0v’ on the communication line.

As 10-bits are used to transmit 8-bits (1-byte) of data the 2.5GBaud provides a raw throughput of 250MBps per lane, however not all of this can be used to transmit useful data. Away from the physical side PCI-E is a packet based interface, much more like USB than it is to traditional “wide bus” PCI and AGP.

This means that control information in the form of headers must be transmitted at the transaction layer in addition to frame, sequencing and error checking in the 2 layers preceding. These are not set values so the throughput will depend on the commands being issued or the tasks performed.

True data efficiency probably lies in the region of 80% or so which would give 200MB/s throughput per lane. Beyond the 3 layers listed PCI-E behaves like traditional PCI with regards to Windows \
enumeration etc.

Topology Of PCI-E

PCI-E cards are connected in such a way that each card has a point-to-point link with a switch as opposed to the bus arbitration required by PCI devices.

In this case the only advantage of PCI-E over AGP is that of data throughput (3200MBps vs. 2100MBps) as AGP is a point to point link anyway. If a graphics card is being used for less conventional uses (maybe video capture?) the equal downstream throughput should provide a large advantage over AGP. For PCI devices, especially those requiring frequent bus accesses, the gains in latency will provide a great advantage.


PCI-E Clock Frequency

PCI-E has a base frequency of 100MHz, this is fed into a clock multiplier which increases the frequency by an effective 25x using one (or possibly 2) PLL. For example if the base clock is increased to 120MHz the differential signaling rate is now 3GB\s. The line coding used (8B/10B) should allow for clock recovery.

Due to the limiting factors such as line capacitance and slew rate, the only way to increase the maximum PCI-E frequency is to increase the voltage output of the driver IC’s, even then the de-serialiser at the other end may be/still be the limiting factor.

 

[Mussels]

One card connected in transport channel , as the AGP ot PCIe, it does not perform ... it moves data.
If the channel is empty , as i clearly proved above about AGP bandwidth ...
there is no issue at all , no matter how the channel are named .

Its like having a very small car , in a highway with 8 free lines ... no problem .

The point is that we still have small cars compared with the size of the highway ...
and no VGA , even the latest VGAs, are still like small cars, in comparison with the size of the highway .

This is the message , that i like to become , commonly understood !!!

you didnt prove anything, about 'AGP bandwidth' and modern cards being 'small cars'
you made some comment about hard drives not providing enough data to video cards, when in fact thats sytem ram that links to them.

Please CLEARLY state what it is you're trying to prove, in a coherent single post.

 

[TechnicalFreak]

A friend of mine, his son has this HP system (core 2 duo), with 2 gig of ram and an PCI-E card.
The last system I had that I sold to another friend (P4 @ 3.0ghz) , he bought an HD2600Pro (agp)
and it gets more FPS in CS:S than the C2D system.. The C2D system if I remember it right has an PCI-E 4350 card. I installed the cards on both the systems. I thought the C2D was suppose to be faster, I installed the cards on both the systems. Now I feel that I did something wrong somewhere..

 

[DrPepper]

One card connected in transport channel , as the AGP ot PCIe, it does not perform ... it moves data.
If the channel is empty , as i clearly proved above about AGP bandwidth ...
there is no issue at all , no matter how the channel are named .

Its like having a very small car , in a highway with 8 free lines ... no problem .

The point is that we still have small cars compared with the size of the highway ...
and no VGA , even the latest VGAs, are still like small cars, in comparison with the size of the highway .

So what your saying is that, the same card in agp against that in PCI-e would have the same performance ? If so then I disagree. I think if there was a GTX295 in an agp slot it would be crippled. I think I may be wrong here but check the memory bandwidth of this card and I think that is greater than what AGP can transfer. I'm probably wrong but I will at least learn something from this.

Oh and nice to see you got a new pc u2k. I expect you to max out cryis this time around.

 

[KainXS]

A friend of mine, his son has this HP system (core 2 duo), with 2 gig of ram and an PCI-E card.
The last system I had that I sold to another friend (P4 @ 3.0ghz) , he bought an HD2600Pro (agp)
and it gets more FPS in CS:S than the C2D system.. The C2D system if I remember it right has an PCI-E 4350 card. I installed the cards on both the systems. I thought the C2D was suppose to be faster, I installed the cards on both the systems. Now I feel that I did something wrong somewhere..

because a 4350 is slower than a 2600pro agp, and cs:s wasn't limited by the p4 he had