Sign in with
Sign up | Sign in
Your question

nForce, Will we be able to up the FSB?

Last response: in CPUs
Share
August 16, 2001 8:23:33 AM

I was hoping that the nForce would be able to support the proper AGP/PCI dividers to enable an FSB higher than 133mhz DDR. Assuming the SDRAM can be set asynchronous to the FSB on the nForce(is this true?), if you had up to 1/3 AGP & 1/6 PCI dividers then this would allow the FSB to be pumped up to 166 or 200MHz DDR without overclocking the AGP & PCI interfaces.

The reason why I am asking is because the Athlon would definitely benefit from a faster FSB and if the nForce supports 2 channels of PC2100 it would easily be able to feed an Athlon with an FSB up to 200MHz DDR with memory bandwidth to spare. At 200MHz DDR you equal the 100MHz quad pumped FSB of the P4, leveling out any performance advantage the P4 has in certain applications.

This doesn't have to be a publicized feature of the chipset, but I we would have no other choice but to use the nForce we wanted to get the most out of an Athlon platform.

More about : nforce fsb

August 16, 2001 1:52:43 PM

unfortunately for all that theorising there is no DDR memory that can support a 200Mhz bus (except on video cards). I think hitting 150Mhz (and still maintaining CL2 timing) is pushing the envelope.

Sure there is talk of (166/333?) DDR soon, but that is still 33.3Mhz shy of the 200. No idea if they can support that at CL2 though....

-* This Space For Rent *-
email for application details
Anonymous
a b à CPUs
August 16, 2001 2:24:01 PM

CL2 isn't per se necessary. On SDRAM it only equated to an 8-10% performance delta. Often you could get better than a 10% increase running CL3 so it was a good move.
Related resources
August 16, 2001 2:32:40 PM

my point is that CL2 150Mhz is faster than 166Mhz CL3 (for SDRAM). I am asuming the same applies to DDR.

-* This Space For Rent *-
email for application details
August 16, 2001 4:12:08 PM

Yes, Peteb is right. You need to get quite a high MHz boost to make up for the higher latency. For SDRAM, at least. For RDRAM, a higher speed is preferable, since latency is completely different.



<font color=blue>Quarter pounder inside</font color=blue>
<font color=red>Change the Sig of the Week!!!</font color=red>
August 16, 2001 4:47:35 PM

higher FSBs!

I guess it has a upper limit. that is why the concept of <font color=red>FSB x multiplier = CPU speed</font color=red> came to being.

imagine a 100 MHz frequency, its period is 10 nS. light travelling at 299792.458 km/sec will travel approximately 2.99 meters in this duration.

electrical current has to travel through physical media (conductor, like copper) where it has to face capacitance as well as inductance. they have such a effect that it appears to be travelling a longer distance, hence in the same duration there is a speed drop. this effect is used in power line simulators in electrical labs where, a capacitor and inductor circuit simulates a conductor of say, 50 km length, but the gadget could be just about a foot long!

so in practice, the electrical signal (the square wave) might travel about half the distance light travels. thus, it an travel a distance of about one and a half meter in 10 nS.

if the FSB is increased, to say 166 MHz or 200 MHz, this period decreases in proportion. these distances are now already closing in to the total track length on a average sized motherboard PCB!!!

plus, frequencies this high are already in the FM range and the EMI (Electro Magnetic Interference) generated by the board might be too excessive, which could in turn, affect other signals on the board.

that is to say, upping the physical FSB beyond a certain limit is practically impossible. i guess 200 (with effort) or 266 (with much effort) is the practical upper limit and it will take a while to achieve these. it might need to develop new PCB material with lower dielectric constant, techniques to print finer tracks and techniques to lower the voltage and current requirements of hte semiconductor chip so that tracks could be made smaller and thinner. integration of more and more components on single chips (a 10x10 mm die does not suffer with such physical consrtains, at least as yet so it can run at whatever speed you could... going for 0.13 micron from 0.18 micron could be a analogy to develop finer PCBs) and thus reduce the physical traces on the PCB making it still smaller.

so I guess the FSB speed will hereafter will really be how the P4 specified, 400 MHz, Quad pumped 100 MHz FSB! there will be a number of channels to aggregate the bandwidth of individual channels. as far as the physical speed of FSB is concerned, we are nearing the end.

girish

<font color=blue>die-hard fans don't have heat-sinks!</font color=blue>
August 17, 2001 5:48:51 AM

Hmm, I was thinking about that.

Now - I would pay for a combined cpu/memory package.

Take today, there is a need and a predominant market for upgradeable cpu and memory.

But - lets sa you know you need 512Mb memory, and that will last the life of your cpu.

If you package the memory and cpu on a board (aka video card) will that allow 200, 250, 275Mhz DDR 'FSBs' due to the distance reductions?

GeForce cards have a GPU and memory running effective speeds of 500+Mhz - why not create a cpu package. Back to the slot processors again, but instead of having the L2 cache off die on the board, we now put the RAM there. Double side the DDR chips a-la GeForce2, good heatsink and off you go. Okay, RAM still cannot read from PCI any faster, but RAM to CPU and RAM to cache is massivley fast.

Seems too simple?

Oh yes - the CPU could still be in a socket, allowing re-use of the memory for a compatible cpu.

-* This Space For Rent *-
email for application details<P ID="edit"><FONT SIZE=-1><EM>Edited by PeTeB on 08/17/01 03:52 PM.</EM></FONT></P>
August 17, 2001 6:35:22 AM

I guess you didn't read my post closely enough. I said "Assuming the SDRAM can be set asynchronous to the FSB on the nForce". There seems to be no reason the think that the memory and FSB have to be sychronous on the upcoming nForce chipset (read up on it, you'll see). At 133 FSB, the nForce (with its dual channel DDR SDRAM) will already provide twice as much bandwidth as the CPU can use. I know you can use the leftover bandwidth for other devices, but you would still have a lot of leftover memory bandwidth. I don't think any of us would even really care about reserving bandwidth for the onboard video as we'll all probably put in our own vid cards.

Basically, what I am saying is that you could leave the SDRAM at 133 DDR x2 (dual channels) and up the FSB to 166 or 200 Mhz. I know a lot of tweakers like to use weird FSB's, but I don't like to run my PCI/AGP out of spec.

OK for comparison:
Standard Dual channel RDRAM P4 platform:
P4 800Mhz RDRAM x2 = 3200 MB/s Bandwidth
P4 100MHz QDR FSB = 3200 MB/s Bandwidth

Here's Athlon on an nForce in standard trim:
133MHz DDR SDRAM x2 = 4200 MB/s Bandwidth
133MHz DDR FSB (std)= 2100 MB/s Bandwidth
Leaves 2100 MB/s available for other components but you are still stuck with the same data path to the RAM

Here's what I hope will be possible with nForce:
133MHz DDR SDRAM x2 = 4200 MB/s Bandwidth
200MHz DDR FSB (OC) = 3200 MB/s Bandwidth
This leaves you with 1000MB/s available to the system components which is still very generous and now it has the same FSB bandwidth as the P4 (and we all know that the FSB of the P4 is responsible for a large part of its performance).

The EV6 bus is spec'd at up to 200MHz DDR (or 400MHz in Intel's terminology) so this would be very possible to do. Remember how much of a performance jump happened when going from the KT133 (100/133 FSB/RAM) to the KT133a (133/133)? That was accomplished by only increasing the FSB by 33% and this was only on standard SDRAM! Imagine what you could do with another 33% FSB increase, 2x the memory bandwidth and extra bandwidth to spare! Oohh... Makes me shiver just to think about it...
August 17, 2001 6:38:29 AM

The EV6 bus is spec'd at up to 200MHz DDR (or 400MHz in Intel's terminology) so this would be very possible to do. Also, check out my other post before this one...
August 17, 2001 9:17:22 AM

the sis735 supports 1/5 which allows for 166mhz cpu fsb with 33 pci and 66 agp 1/6 would allow for 200mhz but the sis doesnt support that, i presume theyre going to try to outdo the sis735 with the nforce so i presume it will have 1/5 at the very least...

if in doubt blame microsoft...
Anonymous
a b à CPUs
August 17, 2001 10:53:08 AM

Quote:
Remember how much of a performance jump happened when going from the KT133 (100/133 FSB/RAM) to the KT133a (133/133)? That was accomplished by only increasing the FSB by 33%"


Not quite right. Aces did a great article on the issue. From what I remember, the performance increase did not really come from increasing the fsb speed, but from running fsb and memory IN SYNCH. Read this:
<A HREF="http://www.aceshardware.com/Spades/read.php?article_id=..." target="_new">http://www.aceshardware.com/Spades/read.php?article_id=...;/A>

and this:
<A HREF="http://www.aceshardware.com/Spades/read.php?article_id=..." target="_new">http://www.aceshardware.com/Spades/read.php?article_id=...;/A>

Here is a quote:
Quote:
The biggest benefit of the VIAKT133A compared to the VIAKT133 is the ability to run the FSB in sync with the memory. Look at the VIA KT133A's performance when we disable interleaving and run at 200/133: almost identical to the VIAKT133! The 2% difference that we measured is probably due to the fact we used ABIT and MSI boards. Our measurements show that the running the FSB and memory in sync provides a 10% boost of performance in this memory intensive game. Our loyal readers might remember that we explained a few months ago why running with a synchronous FSB and DRAM reduces access latency significantly.


So basically, 1/5 and 1/6 dividers would help only if there is memory that can keep up with the FSB.

---- Owner of the only Dell computer with an AMD chip
August 17, 2001 2:45:36 PM

Anybody remembers the Cyrix MediaGX processor???

well, it had a display controller <i>AND</i> a audio controller on the CPU itself. The CPU was their flagship M-II core. The special chipset they made for the MediaGX was probabely the first one to integrate the Northbridge and the Southbridge. But what was the problem? The onchip display shared the system memory, which ran at just 66 MHz back then! 66 MHz is too slow to be shared between the CPU and display (and the sound card as well!), so the result wa obvious - too poor performance!!

While MediaGX was a hint as to what Cyrix had in future, I vaguely remember they had also published a doc with their next generation processor, something like MediaGX plus - I dont remember it well. It integrated the memory controller, a LAN adapter, I/O ports (almost all the Northbridge) and connected to the southbridge on a PCI bus (I guess the CPU itself would be a PCI device). I might wrong on the details. But what happened to Cyrix we all know, went over from National Semiconductors and then to VIA. the project seems to be cancelled as of now.

anyway, such an integration could be done. But I dont think it memory could be integrated onto the CPU itself. the L2 cache has to be there on the CPU, as much as possible and as fast as possible. and due to the amount of memory on the CPU, and bad memory will render the CPU useless (wanting external memory again). maybe arrangements could be made so that you could select any bank from the 512 MB to get some lower amount of memory. but i doubt it.

multiple channels will be the way of life for next generation memory subsystems, and any lack in MHz would be easily compensated by its width and paths.

girish

<font color=blue>die-hard fans don't have heat-sinks!</font color=blue>
August 17, 2001 3:13:01 PM

its high time the PCI and AGP worked on just 33 and 66 MHz! these days all cards can work with a bit overclocked bus. running PCI at 40 and AGP at 80 will contribute to the total performance, while the FSB can go well higher to 160 MHz.

girish

<font color=blue>die-hard fans don't have heat-sinks!</font color=blue>
August 17, 2001 3:52:11 PM

i was wondering at what speed isa worked... i am seriously agains isa slots being removed because there are really few pci hardware modems and barely any pci pcmcia adaptors available... oh well, wireless network cards are so much better like than... old pcis just cant hack it...

if in doubt blame microsoft...
August 17, 2001 4:07:20 PM

ISA slots are 8mhz and some fancier ones are 16 mhz

Change what sig? And keep your C00KIE to YOURSELF!
August 17, 2001 4:17:46 PM

ISA bus is 16 bits wide and works on 12 MHz, giving a peak throughput of 12 MB/sec or less (it takes two cycles for data transfer). Earlier it used to be 8 bits wide and operated at 8 MHz. In the earlier PCs, it had to match the processor speed which was as low as 4.77 MHz.

why do you need any ISA bus? Newer PCI cards will be more efficient and cheaper in the long run.

<b>Not</b> having any ISA slots is a requiremnt of PC99 specification.

and where do you stay? i have rarely seen any ISA modem, we all have PCI modems here in India!

girish

<font color=blue>die-hard fans don't have heat-sinks!</font color=blue>
August 17, 2001 5:05:14 PM

try to find a good PCI based PCMCIA card adaptor on PCI and you'll have a long search.

If you want a wireless PC (desktop/tower) you need ISA...

-* This Space For Rent *-
email for application details
August 17, 2001 5:11:43 PM

correct me if I am wrong, but the dual channel of the nForce is not dual channel for CPU, it allows twin single channel for video and cpu if two DDR modules are inserted. I do not think nForce allows dual channel cpu memory access - but I may be mistaken. I seem to recall this was the case.

I did read you correctly (kind of) but to my current understanding nForce is not dual channel cpu memory access. Dual channel is a way to provide dedicated memory to the video...



-* This Space For Rent *-
email for application details
August 17, 2001 5:12:37 PM

girish, I wasn't talking about integration on die, just on a cpu board. place the RAM close enough to the cpu and you can run 250mhz ddr.....

-* This Space For Rent *-
email for application details
August 17, 2001 5:31:47 PM

no, you werent! :smile:
there was a separate thread on integrated memories somewhere around this week.

integrated memory (if at all you want to eliminate the DIMM sockets) will be fixed, and the motherboard manufacturers will have to produce different versions with different memory sizes. not everybody will buy 512 MB and not everybody will settle for 64 MB.

we had discussed about implementing it in low cost systems like 810 based P-II/Celeron systems for home users who are unlikely going for any upgrade. it would be best suited for large OEMs like Compaq who make people pay through their nose for even a simple memory upgrade.

girish

<font color=blue>die-hard fans don't have heat-sinks!</font color=blue>
August 17, 2001 5:40:58 PM

yes you are right, nForce twin bank architecture enables the onchip display (the GeForce2MX on the IGP) access memory simultaneously with the CPU, and reduces the memory bus contention a bit, but does not completely eliminate it. will the IGP take 128 MB out of 256 MB installed as two 128 MB sticks? the CPU <i>WILL</i> share some bandwidth with the graphics processor, the one from the other DIMM, maybe the ratio of the total bandwidth used by the CPU/graphics will be better than that with other single channel/bank SMA architechtures.

so it becomes important what code/data is stored in the independent DIMM and what is stored in the shared DIMM!

girish

<font color=blue>die-hard fans don't have heat-sinks!</font color=blue>
August 17, 2001 6:13:41 PM

but - as a spin - what if you used a sloket type device, slot type cpu connectors to mobo, but the slocket incorporated RAM?

You buy a slocket type thing to the amount of RAM you need, the cpu could still be on a socket, but because it is so dedicated and near the RAM can run 100% faster than if on a mainboard slot....

-* This Space For Rent *-
email for application details
August 17, 2001 6:16:13 PM

but, if you disable onboard video, does the cpu have dual channel or 2 x single channel DDR RAM slots.

I though it was single channel architecture for the nForce still (in regards to cpu).... Dual meaning capability one for video and one for cpu (if more than one installed) rather than the fully shared feature of older boards...

-* This Space For Rent *-
email for application details
August 17, 2001 6:56:50 PM

yes that could be. basically its a mixture of two, but no twin channels. the bandwidth of the two channels isnt aggregated by the nForce in the sense the P4 does.

when you disable the onboard video, you will be using an external video card and that would have its own memory onboard. the CPU will now get all the available bandwidth of this TwinBank. but I seriously doubt that could be. the memory is <b>NOT</b> defined as dual channel, just two banks. the nForce has two separate 64 bit memory controllers that can work simultaneously to give 4.2 GB/sec. but the CPU itself cannot access memory in two channels simultaneously. it has to acces the 64 bit bus at 133 MHz DDR!

its called the <b>Crossbar</b> architecture where two or more devices can simultaneously a single resource, which has been divided into a number of parts/banks/planes. so we could say that each device (the CPU and IGP or the MCP) has a bandwidth of 2.1 GB/sec each, and it could perform at 4.2 GB/sec only if they simultaneouly access the memory.

but what the nForce can do is to allow interleaving of blocks of memory so that a memory transaction can begin even before the earlier one has finished. this reduces the latency in the memory performance and does contribute to overall performance of the system.

so you could say it is still the same, 2.1 GB/sec of PC2100 DDR memory!!

girish

<font color=blue>die-hard fans don't have heat-sinks!</font color=blue>
August 17, 2001 8:39:42 PM

Raystonn was saying that it was one bank for CPU and one for GPU, but he later admitted he was wrong. You can have both banks for the CPU. I don't know if you have to have the GPU disabled for that or not.

But as a side note, the Athlon can't even accept that much memory bandwidth, so it's kind of pointless.



<font color=blue>Quarter pounder inside</font color=blue>
<font color=red>Change the Sig of the Week!!!</font color=red>
August 18, 2001 12:19:26 AM

sure you can use both bamks for the cpu - but not a dual channel I think - you can still run the board off 1 DIMM therefore the Video and cpu can also share memory.

-* This Space For Rent *-
email for application details
August 18, 2001 4:23:27 AM

it would be a good idea to provide a single SODIMM (or even DIMM) slot for the GPU memory.

nVidia says it can work with mixed combinations like 128MB + 64MB in different slots, so dedicating the other slot for video would give dedicated bandwidth for the CPU, and dedicated bandwidth for the GPU. and the Graphics subsystem could be customised by the user by the amount of memory he wants on his graphics adapter!

anyway, the Athlon CPU wont be accessing at any greater than 2.1 GB/sec so that okay with its dedicated 64 bit DDR 133 MHz memory channel.

girish

<font color=blue>die-hard fans don't have heat-sinks!</font color=blue>
August 18, 2001 8:14:04 AM

isa modems tend to be hardware modems and not win modems which waste cpu resources... not good... and dont work in linux as nobody wants to port the software required to run them, waste of time cuz theyre total bolox... external or isa is the way to go... it is also near impossible to find a pci pcmcia adaptor while isas are easy to come about...

if in doubt blame microsoft...
August 18, 2001 2:08:38 PM

Quote:
sure you can use both bamks for the cpu - but not a dual channel I think - you can still run the board off 1 DIMM therefore the Video and cpu can also share memory.

Both banks of memory are accesable to the cpu via the memory controller regardless of if you are using the integrated video or not ( assuming you are using 2 dimms). by making two banks of memory available to the memory controller allows for more bandwith to be shared with the integrated video. Both the I850 and nforce offer the same 64 bit bus from the cpu to the memory controller. The only differnce being that the p4 is 100 mhz quad pumped verses the t-birds 133 double pumped. This is where you get the difference in bandwith between the cpu and memory controller. The p4 does not have a dual channel ( 128 bit ) cpu to memory controller bus. It does have two banks of memory each supplying 1.6 gigabytes of bandwith ( once again assuming you have two Rdram's installed) to the memory controller. So in effect the nforce is even more sophisticated ( at least on papper) then the I850 as with the nforce you get the crossbar memory controller, something not seen on the I850. In order to achieve parity with the p4( in bandwith)the t-bird would need to run its fsb at 200 mhz. the I850 chip0set was designed to help offset the latency issues of RDRAM.

Video editing?? Ha, I don't even own a camera!
August 19, 2001 7:15:54 AM

but intel's proposed quad pumped 133 MHz (533 MHz) bus offers 4.2 GB/sec, almost the same offered by Dual channel DDR!

and theres still a question about P4 really using it all!

girish

<font color=red>No system is fool-proof. Fools are Ingenious!</font color=red>
August 19, 2001 1:11:57 PM

you can get 150mhz cas 2 ddr already, and i dont think itll be too hard to reach 166 (pc2700) at cas 2.5 which i think will perform better than 150 cas 2...

if in doubt blame microsoft...
August 19, 2001 7:57:11 PM

That is my point!!
The Twin Bank memory architecture will allow for twice as much bandwidth the cpu can use if the fsb is set to 133. but if we allow the cpu FSB to be increased to 166 or 200 than we can increase the usable amount of memory bandwidth.

The reason I would like more divider options is because I have played that FSB game and I am tired of my hard drive getting hosed because the PCI clock is too high. Xtra performance is great to me but not at the expense of stability.
August 20, 2001 1:37:08 AM

Corsair has official PC150 DDR @ Cas2.

i would assume that it might be able to reach 166 at cas2.5


My hamster really wants to give you a big kiss, you sexy hunk you!
August 20, 2001 1:40:06 AM

i read somewhere around on toms (no idea if its true) that one of the P4 motherboards will have a completely independent PCI/AGP clock...
thats to say the PCI & AGP speed will be determined by one clock, and the FSB by another completely different one, essentially removing the divider and capable of running asynchronously.

anyone else heard of this?

My hamster really wants to give you a big kiss, you sexy hunk you!
August 20, 2001 5:05:56 PM

that would be a very good option, but unfortunately it would not let us to squeeze out more performance from the PCI and AGP overclocking.

I'd settle for more dividers rather than separate clocks.

or you could opt to fix the PCI and AGP clocks or set a multiplier.

girish


<font color=red>No system is fool-proof. Fools are Ingenious!</font color=red>
August 20, 2001 5:10:44 PM

Why not just have the option to fix the agp/pci clock at whatever mhz you want?

~Matisaro~
"Friends don't let friends buy Pentiums"
~Tbird1.3@1.55~
August 20, 2001 5:19:22 PM

that would be better!

and so PCI and AGP overclocking science is born!

you can also have separate settings for the AGP and PCI busses!!!!!

girish

<font color=red>No system is fool-proof. Fools are Ingenious!</font color=red>
August 20, 2001 5:21:31 PM

Remember folks, I am the founding father of pci/agp overclocking science, WOO HOOO!
now I can pull a rambus and sue every overclocker for using my ideas....I will be the king of the world.....or suffer a humiliating defeat in a virginia courtroom....maybe I will pass on the lawsuit thingy.

~Matisaro~
"Friends don't let friends buy Pentiums"
~Tbird1.3@1.55~
Related resources
!