What's this MT/s thing, guys look.

[walkerbowe]

Ok, so I'm surfing around the varies computer sites, and I'm looking at the different mobo's. Now suddenly I'm not seeing fsb's written anymore, but this new MT/s??? I go and check other sites and see the same MT/s thing. What this new rating, and how come there not listing the fsb anymore, or how do you convert this MT/s to fsb old school, please someone explain it.

Below is the specs of a mobo I was looking at, look it's got the MT/s thing.

Here's a link, so the site where I saw it, http://shop.ai-tech.com.au/A.i.TECH_327_/Asus_A8N-SLI-DELUXE__Socket_939__nForce4_SLI__2000_MTs__4DDR.ht ml

Supported CPU: Socket 939 AMD Athlon 64FX/64
Chipset: NVIDIA nForce4 SLI
FSB: 2000 MT/s
RAM: 4x DIMM DualChannel DDR 333/400 Max 4GB
Slots: 2x PCI-E X16(SLI mode: x8,x8), 2x PCI-E X1, 3x PCI
Ports: 2x PS/2, 1x LPT, 10x USB2.0(Rear 4), 2x RJ45, 2x SPDIF out, 1x IEEE1394, Audio Ports
IDE: 2x ATA 133 with NV Raid 0/1/0+1
SATA: 4x SATA2 3.0Gb/s with NV Raid 0/1/0+1, 4x SATA with Raid 0/1/0+1/5 by SI3114R
Onboard Audio: Realtek ALC850 8-Ch
Onboard LAN: nForce4 built-in Gbit MAC, MARVELL GbitLAN
Onboard 1394: 2x 1394 ports
Form Factor: ATX

[Cafe_333]

MT/s meaning MegaTransfers per second refers to how many one million data transfers take place.
1MT = 1 million data transfers.

The way all CPUs communicated to the Northbridge and the ram, was through a parallel Front Side Bus. Then when the Athlon64 was introduced, AMD opted to use a new type of high speed bus to communicate with the NorthBridge called the 'HyperTransport' bus - a serial interface which allowed data transfer rates of 3.2GB/s (1600MT/s) in both directions, giving a total bandwidth of 6.4GB/s. Current AMD motherboards with newer revisions of the HyperTransport bus now allow total data transfer rates of 8.0GB/s (2000MT/s).

EDIT: In contrast btw, Intel P4 533fsb CPU's perform data throughputs of 4.2GB/s - and can not communicate in both directions simultaneously like AMD can. Finally, their flagship 800fsb CPU's perform at 6.4GB/s.

Block Diagram of a sample VIA chipset with the HyperTransport bus:
http://www.via.com.tw/en/images/products/chipsets/k8-series/k8t890_blkdiagram.jpg

[goofball]

MT/Sec = mega transfers/second.

technically, there is no FSB in Athlon64 systems, since the memory controller and traffic has been removed from the chipset. now, there is a point to point link (Hypertransport) from CPU to chipset.

[goofball]

In contrast btw, Intel P4 533fsb CPU's can only perform at a max data throughput of 3.97GB/s - and can not communicate in both directions simultaneously like AMD can. Finally, their flagship 800fsb CPU's all top out at 5.96GB/s.



how are you getting those numbers?

533 FSB max throughput = 133.33 x 4 x 8= 4.264GB/sec
800 FSB max throughput = 200 x 4 x 8 = 6.4GB/sec

unless my math and understanding is wrong. I'm not taking bus efficiency into effect, as I am calculating peak throughput.

[Cafe_333]

I remember reading about that on THG when Athlon64 was first introduced.
http://www.tomshardware.com/cpu/20030923/athlon_64-10.html
"By comparison, the Pentium 4 with 533 MHz FSB allows a maximum data throughput of 3.97 GB/s - but not in both directions simultaneously."

I grabbed the numbers from that, but yours does seem more accurate. I suspect mine may be off but i'll google some more.

EDIT: odd, not sure how THG got that figure, but everything else I've read on various sites from google support 533fsb @ 4.2GB/s and 800fsb @ 6.4GB/s. However those may be theoretical and the bus efficiency could be what THG was stating - though that was not made clear. Anyway, I corrected my above post. :confused:

[konfusion666]

all you need to know is that, generally speaking, a Socket 939 motherboard will give you 2000MT/s (because the HyperTransport bus is running at 1GHz).

and a Socket 754 motherboard will give you 1600MT/s (HT Bus runs at 800MHz).

anyways, that just assumes "stock" - if you overclock you can run them at whatever the hell you want. usually you don't need to OC the HTBus though, I don't think modern hardware/applications has come close to saturating it (excepting server stuff)

[walkerbowe]

all you need to know is that, generally speaking, a Socket 939 motherboard will give you 2000MT/s (because the HyperTransport bus is running at 1GHz).

and a Socket 754 motherboard will give you 1600MT/s (HT Bus runs at 800MHz).

anyways, that just assumes "stock" - if you overclock you can run them at whatever the hell you want. usually you don't need to OC the HTBus though, I don't think modern hardware/applications has come close to saturating it (excepting server stuff)

so if it has a 2000 MT/s then it's 1gz bus. I'm assuming whatever the MT/s is, half that and that's what the fsb is?

Is that a fair thought?

I ask all these questions, cause when your chooing DDR memory, you have to know what speed to buy, and since pc3200 runs at 400mhz. You have to decide what speed to buy, if your getting a newer board etc.

[konfusion666]

you don't need to concern yourself with anything.

just get PC3200 ("400mhz") memory unless you are planning on overclocking, then get good brand-name PC3200 (OCZ) or something rated higher (like PC4400... whatever).

most current Athlon64's and most Semprons are on PC3200, so that's all you need...

[Cafe_333]

so if it has a 2000 MT/s then it's 1gz bus. I'm assuming whatever the MT/s is, half that and that's what the fsb is? Is that a fair thought?Sounds right to me. Considering AMD motherboards with 1600MT/s and Intel boards with 800FSB both give 6.4GB/s data throughput, that would be a fair *thought* at most, but keeping in mind HyperTransport is a serial interface - meaning it's point to point direct communication in both directions, while Intel is using a parallel FSB which communicates in one direction at a time. Remember, "technically, there is no FSB in Athlon64 systems" as goofball put it.

[walkerbowe]

you don't need to concern yourself with anything.

just get PC3200 ("400mhz") memory unless you are planning on overclocking, then get good brand-name PC3200 (OCZ) or something rated higher (like PC4400... whatever).

most current Athlon64's and most Semprons are on PC3200, so that's all you need...

so what's the speed of the bus from the memory (pc3200) to the cpu? If I was going to buy DDR OCZ memory or something to that match my system. How would I be able to tell what speed to buy, so that I don't buy faster memory then my MT/s thing whatever is, non-fsb stupid thing, grrrr.???

[walkerbowe]

you don't need to concern yourself with anything.

just get PC3200 ("400mhz") memory unless you are planning on overclocking, then get good brand-name PC3200 (OCZ) or something rated higher (like PC4400... whatever).

most current Athlon64's and most Semprons are on PC3200, so that's all you need...
I thought I wouldn't need to overclock? I would just need to buy faster memory, like say OCZ Platinum PC2-8000 (1ghz) ? Isn't buying memory that fast, actually the same speed as the serial bus MT/s 2000 rating? or is the memory slower or faster then the MT/s rating etc?

[Cafe_333]

so what's the speed of the bus from the memory (pc3200) to the cpu? If I was going to buy DDR OCZ memory or something to that match my system. How would I be able to tell what speed to buy, so that I don't buy faster memory then my MT/s thing whatever is, non-fsb stupid thing, grrrr.???You match the data throughput of the bus speed. In the old days this was easy because with a FSB of 400MHz, you only need DDR400 ram. For Intel systems of 800FSB, you need two bars of DDR400 in dual channel to give 800MHz speeds.

With new AMD systems using the HyperTransport bus, you need to know what the MT/s is. The first Athlon64 boards were 1600MT/s, equivalent to 6.4GB/s. In this case, using two DDR400 modules in dual channel configuration will give you a bandwidth of 6.4GB/s (PC3200x2).

[Cafe_333]

I should point out a change from my previous post. Currently all Athlon64's only perform at 1600MT/s. Two bars of DDR400's in dual channel configuration will actually give you the most optimal performance.

The latest AMD boards now do 2000MT/s, equivalent to 8.0GB/s - which can support X2 processors that also perform at 2000MT/s. The closest ram to match these dual core cpus are two PC4200's (DDR533) in dual channel giving 8.4GB/s bandwidth (PC4200x2). However because of this, I do believe that current Socket939 motherboards are not ideal for dualcore CPUs because it is bottlenecked by the ram. Currently they only support max PC3200 (DDR400).

It is only ideal to buy Athlon64 cpus for the time being. By the time X2's come down in price, hopefully by then there will be AMD boards that can support DDR533 and up.

As an aside note.... if you think X2's are fast now, just wait and see how much more potential these bad boys have when DDR533 AMD boards hit the market.

[konfusion666]

yes, basically the point is you just use standard PC3200 DDR RAM with any consumer-grade Athlon64.

don't bother trying to figure out HyperTransport, bus speed, MT/s ... just get some PC3200 and be done with it.


I should point out a change from my previous post. Currently all Athlon64's only perform at 1600MT/s. Two bars of DDR400's in dual channel configuration will actually give you the most optimal performance.

you are just confusing the guy, plus you don't know the K8 architecture very well.

there is no relevance between the speed of the HyperTransport bus and the speed of memory.

once again, the K8 architecture does NOT use the HyperTransport bus to communicate with memory.

(don't worry, i've made the same mistake before)

http://www.cpuid.org/K8/index.php

[goofball]

just look at it this way.

without introducing any CPU memory dividers, you would want at least PC3200.

if you start to increase the HT frequency from the base 200, the memory speed will also increase

216 = PC3500
233 = PC3700
250 = PC4000
263 = PC4200

and so on.

while HT and memory speed do not have a direct relationship, the clock generators are somewhat linked. increasing the HT will increase the memory speed.

HTT is for communicating from CPU to chipset, not CPU to memory.

[konfusion666]

just look at it this way.

without introducing any CPU memory dividers, you would want at least PC3200.

if you start to increase the HT frequency from the base 200, the memory speed will also increase

216 = PC3500
233 = PC3700
250 = PC4000
263 = PC4200

and so on.

while HT and memory speed do not have a direct relationship, the clock generators are somewhat linked. increasing the HT will increase the memory speed.

HTT is for communicating from CPU to chipset, not CPU to memory.

yes, the same BASE CLOCK is used to generate the final HT BUS FREQUENCY and MEMORY BUS frequency... sorry for yelling, lol.

On my K8N-E Dlx, I have the base clock set to 250mhz.

HT Multiplier = 3

CPU multiplier = 10.0 (default)

Memory ratio (divider) = 166 (333)

End result?

Hypertransport frequency is 250 x 3 = 750mhz

cpu speed is 250 x 10.0 = 2500mhz

memory speed (this is the tricky one) = 2500 / X

X = number which yields closest answer to (250 x 0.83 = 207.5)
X turns out to be 12!

memory speed = 2500 / 12 = 208mhz!

[walkerbowe]

just look at it this way.
HTT is for communicating from CPU to chipset, not CPU to memory.

so what's the speed from the CPU to the memory modules? I thought pc3200 was the speed of the memory itself, and how fast it can be accessed. I thought you had to buy memory according to the bus speed between the memory module and the cpu?

[walkerbowe]

yes, the same BASE CLOCK is used to generate the final HT BUS FREQUENCY and MEMORY BUS frequency... sorry for yelling, lol.

On my K8N-E Dlx, I have the base clock set to 250mhz.

HT Multiplier = 3

CPU multiplier = 10.0 (default)

Memory ratio (divider) = 166 (333)

End result?

Hypertransport frequency is 250 x 3 = 750mhz

cpu speed is 250 x 10.0 = 2500mhz

memory speed (this is the tricky one) = 2500 / X

X = number which yields closest answer to (250 x 0.83 = 207.5)
X turns out to be 12!

memory speed = 2500 / 12 = 208mhz!


Ok, I'm seeing a MEMORY BUS frequency, is this the fsb? I'm starting to follow this stuff, but where'd you get the 0.83 from?

[konfusion666]

so what's the speed from the CPU to the memory modules? I thought pc3200 was the speed of the memory itself, and how fast it can be accessed. I thought you had to buy memory according to the bus speed between the memory module and the cpu?

pc3200 is just a buzzword. the 3200 actually comes from the bandwidth achieved but that is not really relevant...

pc3200 memory runs at 200mhz and is "DDR" - Dual Data Rate - 2 "pieces" of data per signal.

see this pic: http://www.cpuid.org/K8/MC2.gif

Requests from Athlon64 CPU to memory: sent at SPEED OF CPU (2000mhz+ !!)

Actual memory "packets" from memory to CPU: sent at memory speed (200mhz x 2 = "400mhz")

[konfusion666]

Ok, I'm seeing a MEMORY BUS frequency, is this the fsb? I'm starting to follow this stuff, but where'd you get the 0.83 from?

you have to stop thinking in terms of "fsb" to understand the K8 architecture...

because I am using the 166mhz "memory speed divider" (every BIOS has a different name for it) I have to multiply my base clock by 0.83.

(166 / 200 = 0.83)

[walkerbowe]

you have to stop thinking in terms of "fsb" to understand the K8 architecture...

because I am using the 166mhz "memory speed divider" (every BIOS has a different name for it) I have to multiply my base clock by 0.83.

(166 / 200 = 0.83)

uh? You lost me? Are you starting about the speed you set the memory at, in the bios? Example I set mine at 200mhz X 2 = 400mhz

Is this what your talking about, when you get the 0.83?
In my case it'd be 200/200 = 1

[goofball]

you have to stop thinking in terms of "fsb" to understand the K8 architecture...

because I am using the 166mhz "memory speed divider" (every BIOS has a different name for it) I have to multiply my base clock by 0.83.

(166 / 200 = 0.83)

actually, the divider is not a static option setting. it's much more complicated than that with the Athlon 64.

let's take your CPU for example. with a 10x multiplier, it sounds like a 3200+?

in any case. you initiated a 6:5 memory divider for 166mhz operation (or so it should be).

your multiplier is 10.

to get your memory divider, you multiply 10 x 6, divide by 5. so 60/5 = 12. that's why you have a 12 divider from total CPU speed, to get your memory speed.

So as you can see, changing the multiplier will change the memory frequency and divider as well. It's not a static option like it used to be.

[konfusion666]

i have a different formula which seems to "work" - verified using clockgen for a variety of base frequencies from 200 to 270 (my board/CPU's limit - Newcastle CG-stepping socket 754 3000+) but i think ultimately the math works out to be the same.

yup, i've even tried lower multiplers, I ran with 9.0 for a bit and 9.5 for an afternoon... partial multipliers don't seem to be "real" though, and i think there was an article somewhere about that.

[goofball]

only integer multipliers with Athlon64. half multipliers are rounded up to the next highest integer value.

your formula works because you are using the final memory frequency as your determinant for finding the CPU Memory speed divider. This works but isn't the best way to do it if you have no idea what the resulting memory speed is going to be.

It's best to use the formula I outlined, so you can see what the resultant memory speed will be at a certain multiplier setting and HTT speed. This formula will work for all multipliers and dividers.

[walkerbowe]

I should point out a change from my previous post. Currently all Athlon64's only perform at 1600MT/s. Two bars of DDR400's in dual channel configuration will actually give you the most optimal performance.

The latest AMD boards now do 2000MT/s, equivalent to 8.0GB/s - which can support X2 processors that also perform at 2000MT/s. The closest ram to match these dual core cpus are two PC4200's (DDR533) in dual channel giving 8.4GB/s bandwidth (PC4200x2). However because of this, I do believe that current Socket939 motherboards are not ideal for dualcore CPUs because it is bottlenecked by the ram. Currently they only support max PC3200 (DDR400).

It is only ideal to buy Athlon64 cpus for the time being. By the time X2's come down in price, hopefully by then there will be AMD boards that can support DDR533 and up.

As an aside note.... if you think X2's are fast now, just wait and see how much more potential these bad boys have when DDR533 AMD boards hit the market.

Instead of using a dual channel configuration, could I just continue using a single channel, and install a OCZ GOLD EL PC2-6400 1GB memory? If I did this approach, would I still reach the 8.0GB/s bandwidth of new HT systems.

[goofball]

no, because PC2 = DDR 2, which is not supported by AMD at the moment. different pin count, would not fit the current DDR slot without a hammer.

[walkerbowe]

no, because PC2 = DDR 2, which is not supported by AMD at the moment. different pin count, would not fit the current DDR slot without a hammer.

So I guess your saying there designed for Intel based systems? What I'm asking is can you buy faster OCZ EL DDR PC-4800 and run it in single channel mode. Inside a MT/s 1600 HT bus, it's be pretty well close right?
DDR PC4800 would be running at = 600mhz, and the HT bus would be 800mhz. That'd be as close as I could get in single channel mode, right?

[konfusion666]

So I guess your saying there designed for Intel based systems? What I'm asking is can you buy faster OCZ EL DDR PC-4800 and run it in single channel mode. Inside a MT/s 1600 HT bus, it's be pretty well close right?
DDR PC4800 would be running at = 600mhz, and the HT bus would be 800mhz. That'd be as close as I could get in single channel mode, right?

i don't know how to say it nicely, but dude, you are not listening to what we are saying.

PC3200 is ALL YOU NEED unless you OVERCLOCK.

Unless you OVERCLOCK, there is absolutely NO POINT WHATSOEVER to using "greater than PC3200" memory with an Athlon64.

end of story.

[walkerbowe]

i don't know how to say it nicely, but dude, you are not listening to what we are saying.

PC3200 is ALL YOU NEED unless you OVERCLOCK.

Unless you OVERCLOCK, there is absolutely NO POINT WHATSOEVER to using "greater than PC3200" memory with an Athlon64.

end of story.

If your getting frustrated, you can always stop replying. I'm sure there's others that can answer.
I just figured instead of overclocking the memory, just buy faster memory and run it at it's stock speeds. I'm reading what you guys are writing, it's just I haven't fully put this all together in my head yet.

You guys, have said alot of the different things to me in the last 2 days. I still have to sort through it all, and find out who's feeding me false information and who actually knows what there talking about. (no offense)

[walkerbowe]

i don't know how to say it nicely, but dude, you are not listening to what we are saying.

PC3200 is ALL YOU NEED unless you OVERCLOCK.

Unless you OVERCLOCK, there is absolutely NO POINT WHATSOEVER to using "greater than PC3200" memory with an Athlon64.

end of story.

Oh but ya, I understand it more then I did 2 days ago. My goal of understanding what the MT/s is, is alot more clear.

Thanks for that anyway

[Cafe_333]

Instead of using a dual channel configuration, could I just continue using a single channel, and install a OCZ GOLD EL PC2-6400 1GB memory? If I did this approach, would I still reach the 8.0GB/s bandwidth of new HT systems.Well for starters, all AMD motherboards can only support up to max PC3200. So even if you wanted to buy higher bandwidth ram, it'll just downclock to PC3200 speeds. So currently there is no option to go any higher. :D

And yes, you are correct, high bandwidth ram currently only benefits Intel systems.

you are just confusing the guy, plus you don't know the K8 architecture very well. there is no relevance between the speed of the HyperTransport bus and the speed of memory. once again, the K8 architecture does NOT use the HyperTransport bus to communicate with memory.I am already aware that the K8 architecture does not use the HyperTransport to communicate with the memory - only to the Northbridge. What me must look at is the memory bandwidth of the CPU's integrated memory controller. And Athlon64 cpu's have a peak memory bandwidth of 3.2GB/s.

http://www.canadacomputers.com/index.php?do=ShowProduct&cmd=pd&pid=006028&cid=CPU.611
"An integrated memory controller with peak memory bandwidth of up to 3.2GB/sec"

http://www.tomshardware.com/cpu/20030923/athlon_64-11.html#integrated_memory_interface_no_northbridge _meddling
"The dual-channel interface of the Athlon 64 offers a memory bandwidth of 6.4 GB/s"

In dual channel, the Athlon64 will have a peak memory bandwidth of 6.4GB/s, meaning the only way we can get these speeds is with two DDR400 modules in dual channel (PC3200x2). Well hold on a minute..... the 6.4GB/sec bandwidth just so happens to be equivalent to the 1600MT/s speed of the HyperTransport.

So that said, the CPU communicates to the Northbridge (via HyperTransport) at 1600MT/s, AND to the RAM at 6.4GB/s.... well hey, that happens to be the same speed(1600MT/s = 6.4 GB/s). :) We see this in X2 as well - these communicate to the Northbridge via HyperTransport at 2000MT/s and to the RAM at 8.0GB/s -- same speed.

1600MT/s = 6.4 GB/s
2000MT/s = 8.0 GB/s

Conclusion: The speed of the HyperTransport is equal to the speed the Memory Controller can communicate to the RAM. I never said the CPU uses the HyperTransport to communicate with the memory -- I only said to match the ram speed to it. Hence, you must match the speed of the ram, to the speed of the bus - which has always been the way to optimize system performance. Which type of bus depends on the system--Intel and K7 using the parallel FSB, and K8 using the serial HyperTransport bus.

Simplified:
1600MT/s = 800fsb = DDR400x2 = PC3200x2 = 6.4GB/s

As for the bottleneck in X2 systems.... seeing how current AMD motherboards and K8 cpu's can only perform at a max memory bandwidth of 6.4GB/s, I believe X2's are bottlenecked when they have the potential to communicate at peak 8.0GB/s. Only they can't with the current hardware limitations. When AMD motherboards can handle PC4200 speeds, only then we will see optimal performance.

2000MT/s = 1066fsb = DDR533x2 = PC4200x2 = 8.4GB/s

[konfusion666]

If your getting frustrated, you can always stop replying. I'm sure there's others that can answer.

i am not frustrated, i just want you to save your money and just get some PC3200 ... instead of wasting money on some overpriced PC4800 OCZ crap (no offense to OCZ; i own OCZ ram myself, but I buy it for a reason)



You guys, have said alot of the different things to me in the last 2 days. I still have to sort through it all, and find out who's feeding me false information and who actually knows what there talking about. (no offense)

its a complex subject so it is to be expected that people will occasionally get their facts mixed up

[konfusion666]

In dual channel, the Athlon64 will have a peak memory bandwidth of 6.4GB/s, meaning the only way we can get these speeds is with two DDR400 modules in dual channel (PC3200x2). Well hold on a minute..... the 6.4GB/sec bandwidth just so happens to be equivalent to the 1600MT/s speed of the HyperTransport!!

not true. HyperTransport ran at 1600MT/s for socket 754 (single-channel A64)
but HyperTransport for socket 939 (dual-channel A64) runs at 2000MT/s. ;)

and the memory controller communicates to the ram at the speed of the CPU.

data from the RAM is sent back to the CPU at the speed of the RAM - which is at a rate of 6.4GB/s on socket 939,dual-channel systems -- and at a rate of 3.2GB/s on socket 754,single-channel systems.

Conclusion: The speed of the HyperTransport is equal to the speed the Memory Controller can communicate to the RAM.

no.

[Cafe_333]

Actually it's dependant on the chipset. It doesn't matter what the socket is. In fact, some of the earliest s939 motherboards only supported 1600MT/s (like the this one (http://www.giga-byte.com/Motherboard/Products/ComparisonSheet_GA-K8NSC-939.htm)) and newer boards support both 1600 and 2000MT/s - it just depends on what speed the CPU can run at. All Athlon64's, regardless of socket, run at 1600MT/s and X2's run at 2000MT/s.

From the same THG article (http://www.tomshardware.com/cpu/20030923/athlon_64-10.html) in the above link, "The serial interface with a variable bitrate allows the Athlon64 to attain a data transfer rate of 3.2 GB/s - in both directions simultaneously. This results in a total bandwidth of 6.4 GB/s." So this means all Athlon64's have a memory bandwidth of 6.4GB/s regardless. Therefore we must dual channel PC3200x2 to attain 6.4GB/s speeds.

- the THG article cited was from Sept.2003 when A64's were first introduced.

Summary:
- Athlon64's have a memory bandwidth of 6.4GB/s
- The HyperTransport runs at 1600MT/s with Athlon64 cpu's
- 1600MT/s = 6.4GB/S
- The speed of the HyperTransport is equal to the speed the Memory Controller can communicate to the RAM.
- You must match the speed of the ram, to the speed of the bus.

I think my points stand. ;)

[Cafe_333]

I consolidated my two posts into the above one, with new findings. :D

[walkerbowe]

not true. HyperTransport ran at 1600MT/s for socket 754 (single-channel A64)
but HyperTransport for socket 939 (dual-channel A64) runs at 2000MT/s. ;)

and the memory controller communicates to the ram at the speed of the CPU.

data from the RAM is sent back to the CPU at the speed of the RAM - which is at a rate of 6.4GB/s on socket 939,dual-channel systems -- and at a rate of 3.2GB/s on socket 754,single-channel systems.



no.
Oh, going from the CPU to the memory it's the cpu speed, but coming from the RAM to the CPU, it's the speed of the ram.
Which is the HT = 6.4GB/s or MT/s 2000.

Cool, it's making more sense, lol.

[konfusion666]

I think my points stand. ;)

i don't think so, but I'm getting tired, so i will let this one slide... heh.

btw that sledgehamer articles is outdated/inaccurate since it's referring to the original Opteron! DDR333 = 5.4GB/s max mem b/w to CPU!

THG sucks man. I've been reading his articles since 1996-1997 and they haven't gotten any better ;)

[Cafe_333]

The HyperTransport speed of the motherboard is in MT/s. What it'll actually run at depends on the CPU. Athlon64's run at 1600MT/s, and X2's run at 2000MT/s.

[konfusion666]

- You must match the speed of the ram, to the speed of the bus.

Nope, very wrong assumption. You're still thinking of the K8 architecture like the K7 or P4 architecture. Try reading some sites other than THG...

[Cafe_333]

i don't think so, but I'm getting tired, so i will let this one slide... heh.

btw that sledgehamer articles is outdated/inaccurate since it's referring to the original Opteron! DDR333 = 5.4GB/s max mem b/w to CPU!

THG sucks man. I've been reading his articles since 1996-1997 and they haven't gotten any better ;)I'm aware of how many people now view THG, but my points have nothing to do with it other than merely referencing.

So let's put THG aside. The fact is, single-core Athlon64's run at 1600MT/s, regardless of socket and have a memory bandwidth of 6.4GB/s. And we agree that the speed the CPU communicates *to* the ram is at the speed of the CPU. Well hey, it communicates to the ram at 6.4GB/s. So I would want to run it with ram with the same bandwidth to handle the data. When you have 6.4GB/s of Bandwidth coming *from* the CPU, you'll need RAM at the same bandwidth to receive all of it. To do this, PC3200x2 = 6.4GB/s. Because you wouldn't want PC2700x2, it'll bottleneck it. And higher bandwidth ram like PC4200x2 would be pointless.

Therefore if you can agree that 1600MT/s = 6.4GB/s, then you must agree that the speed of the HyperTransport is equal to the speed of the Memory Controller. Therefore a good rule of thumb is match the speed of the Ram to the Bus. ;)

[goofball]

Well for starters, all AMD motherboards can only support up to max PC3200. So even if you wanted to buy higher bandwidth ram, it'll just downclock to PC3200 speeds. So currently there is no option to go any higher. :D


That statement is not true. Because DDR1 has peaked at PC3200 in terms of JEDEC spec, there is nothing that anyone can truly advertise since there is no such thing as official PC3500/3700/4000/4200/4400/4800. Hence, they only claim PC3200 support because that is the official JEDEC spec. that's not to say that the support isn't there for higher speed memory modules. It's just that there is nothing "official".

I'm sure you've already read the article on Anandtech about AMD's support for PC4000 in their new rev CPU's? Sure, it doesn't do much at all but the official support from AMD is there.

[Cafe_333]

I remember that article - unless I misunderstood it, the dividers need to be enabled in the bios to run higher bandwidth ram otherwise the ram should only run at PC3200. It's been a while since I read it so I could be wrong on that. But I should clarify that AMD boards can run higher bandwidth ram, but only with Rev.E cpu's and i'm pretty sure it can't at factory settings either. Correct me if I'm wrong on that.

In any case, there seems to be a reason why AMD is targeting support specifically for DDR500/PC4000 and not any of the other unofficial bandwidth specifications that are lower or higher. Why skip PC3500 or PC3700? Why not PC4200? But AMD pin pointed to PC4000... I could hypothesize that this would eliminate the bottleneck in X2's I was describing earlier. X2 cpu's memory controller can push out 8.0GB/s of bandwidth. I would run PC4000x2 on the receiving end which can receive 8.0GB/s of bandwidth from the CPU. The fact AMD fingered out PC4000 does further support my argument of matching the speed of the memory controller to the speed of the HyperTransport, as 8.0GB/s is equal to 2000MT/s respectively.

Sure, it doesn't do much at all but the official support from AMD is there.I hope it'll do more than what Anandtech describes. :) It's hard to believe that higher bandwidth doesn't do anything, but of course demands are dependent on the application being used. In general, graphics-intensive operations demand alot of bandwidth, office applications benefit more from latency, and games tend to want both. Unfortunately Anandtech's article as I recall did not have any benchmarks in graphics intensive operations. Office apps do not benefit from bandwidth, so no surprises there when Anandtech showed only 1-2% gains. And finally, only 5% gains in gaming, which I'm not so quick to agree with benchmarks that were only done at 1024x768 resolutions - I suspect the real gains will be seen in the higher resolutions. DDR400 does seem sufficient enough for X2's but I suspect DDR500 will be much better for it in gaming.

[goofball]

I remember that article - unless I misunderstood it, the dividers need to be enabled in the bios to run higher bandwidth ram otherwise the ram should only run at PC3200. It's been a while since I read it so I could be wrong on that. But I should clarify that AMD boards can run higher bandwidth ram, but only with Rev.E cpu's and i'm pretty sure it can't at factory settings either. Correct me if I'm wrong on that.

Stock settings, yes, the Athlon64 (prior to Rev E) supported only PC3200 but with the ability to lower the multiplier available on all Athlon64 CPU's, you could have used PC4400 and gotten it to work fine, while maintaining stock CPU speed. The Rev E is the first one that has the official PC4000 support in it.

In any case, there seems to be a reason why AMD is targeting support specifically for DDR500/PC4000 and not any of the other unofficial bandwidth specifications that are lower or higher. Why skip PC3500 or PC3700? Why not PC4200? But AMD pin pointed to PC4000... I could hypothesize that this would eliminate the bottleneck in X2's I was describing earlier. X2 cpu's memory controller can push out 8.0GB/s of bandwidth. I would run PC4000x2 on the receiving end which can receive 8.0GB/s of bandwidth from the CPU. The fact AMD fingered out PC4000 does further support my argument of matching the speed of the memory controller to the speed of the HyperTransport, as 8.0GB/s is equal to 2000MT/s respectively.

It may be that, but it may also have to do with JEDEC being very close to ratifying a PC4000 spec? it's really all speculation at this point.

To be honest, you couldn't do enough to saturate the CPU to chipset HT link anyways, so I don't see the CPU to memory link as being a bottleneck at this point.

I hope it'll do more than what Anandtech describes. :) It's hard to believe that higher bandwidth doesn't do anything, but of course demands are dependent on the application being used. In general, graphics-intensive operations demand alot of bandwidth, office applications benefit more from latency, and games tend to want both. Unfortunately Anandtech's article as I recall did not have any benchmarks in graphics intensive operations. Office apps do not benefit from bandwidth, so no surprises there when Anandtech showed only 1-2% gains. And finally, only 5% gains in gaming, which I'm not so quick to agree with benchmarks that were only done at 1024x768 resolutions - I suspect the real gains will be seen in the higher resolutions. DDR400 does seem sufficient enough for X2's but I suspect DDR500 will be much better for it in gaming.

The athlon64 was never really starved for bandwidth to begin with. It's really hard to say without testing.

[Cafe_333]

It may be that, but it may also have to do with JEDEC being very close to ratifying a PC4000 spec? it's really all speculation at this point.I'm wondering why they are close to ratifying PC4000 and not any of the lower bandwidths like PC3500 or PC3700.... it is odd that they jumped straight to that, but it does coincide well with 8GB/s of memory bandwidth X2's can push out to the memory. Not saying they are ever bandwidth starved, but it does remove any kind of bottlenecks when everything can communicate in synchronously. Not saying that the bandwidth is that saturated either that bottlenecks will be a problem, but it does remain to be seen if this will give any real advantages going with PC4000. I suspect it does, but I hope there will be later reviews on that. :)

[goofball]

I think that it would be a good idea to get PC4000 that has low latency that doesn't require high voltage to keep the low latency.

as for not ratifiying the other spec's, I think it may have to do with the fact that most manufacturer's aren't making IC's based on those ns ratings. TCCD is basically PC3200 2-2-2-5 or PC4000 3-4-4-8 rated. It's not rated for in between.

There was word that JEDEC was pretty close to going to a PC3700 spec, and that Samsung was pushing hard for this. Never came around though, and not much interest was generated after the initial talk.