What is the highest HTT link...?

[Tahir2]

What is the highest HTT link acheived on the BF4 platform??
Whilst doing this does not offer the best performance I was able to acheive a fairly stable (it ran Sisoft Sandra and other applications but not tried Toast or Folding with it) and ran pretty well.
Here is a screenshot

This is a Winchester 3000+ Skt 939 CPU on a DFI LanParty NF4 motherboard with OCZ DDR400 Rev 3 1024 Dual Channel Kit.

It is all done on air and the first overclock you see of 2.8GHz is not the main point but rather the HTT/FSB overclock of 437 MHz.

 

[BRiT]

I believe someone (MrIcee/overcrash86/cpulloverclock) from XtremeSystems has hit over 500Mhz HTT, and also hit a 1:1 HTT/RAM Ratio at 300Mhz. Cant seem to find the link now, but still digging around for it.

EDIT: Found the link. A thread here shows a few people have gotten beyond the 500Mhz HTT barrier, hitting 520Mhz and 543Mhz.

 

[Guden Oden]

Aren't some A64s and O's coming with 1000MHz (2x500) HT links as standard these days?

 

[BRiT]

Standard is 200Mhz HTT to the CPU.

 

[The Postman]

Isn't it 200*5 rather then 500*2?

 

[DudeMiester]

^^ Yes.

Well it's all nice and cool that you can get such a high HTT clock, it does nothing for your performance. The A64 already has loads of FSB bandwidth stock, even after overclocking you can't saturate it. Maybe if you had a 5 drive RAID 0 or something insane, but because of the integrated memory controller most of the load is taken off the FSB.

 

[Tim]

Isn't it 200*5 rather then 500*2?

No the connection it is 2x1000. The base clock is 200MHz the HT connection runs at 5x the base clock and is double data rate thus efficiently 2Ghz.

 

[The Postman]

I thought it sent data both ways rather then having a double data rate or is that what you meant? *needs to read up on HHT again*

 

[Thowllly]

I thought it sent data both ways rather then having a double data rate or is that what you meant? *needs to read up on HHT again*

It both runs at double data rate and sends data both ways at the same time (full duplex)

So the effective datarate for socket 939 is 2Ghz. The link width is 16bit in each direction for a total of 8GB/s aggregated bandwidth.

 

[Tahir2]

^^ Yes.

Well it's all nice and cool that you can get such a high HTT clock, it does nothing for your performance. The A64 already has loads of FSB bandwidth stock, even after overclocking you can't saturate it. Maybe if you had a 5 drive RAID 0 or something insane, but because of the integrated memory controller most of the load is taken off the FSB.

Absolutely, in fact at these speeds it has reduced performance for a number of reasons. But I didn't do it for extra performance. I did it because I could

 

[Guden Oden]

So with a base clock of 437MHz you have to lower the multiple, right? Because I don't believe your HT link runs at like 2200MHz DDR in either direction when standard is 1000MHz... Sounds a bit too good to be true frankly.

 

[Tahir2]

That is correct - it was at 2x I believe but this is not what hurts performance. Runing the CPU and memory at such ASYNC clocks (2:1) kills processing performance as well as the fact that you are not able to acheive high clockspeeds with a high HTT link.
That same system is running optimal at a HTT link of 280 MHz as the memory is running 200:166 the multiple of the LDT is running at 4x and the CPU is running at 2.5GHz.

 

[Tim]

That is correct - it was at 2x I believe but this is not what hurts performance. Runing the CPU and memory at such ASYNC clocks (2:1) kills processing performance as well as the fact that you are not able to acheive high clockspeeds with a high HTT link.

There are no such thing as async memory with A64s the memory is always running in sync with the CPU/memory controler.

 

[Tahir2]

Tim there are memory dividers from that work with the HTT link in some way. The memory controller is always running at whatever clockspeed your CPU is running at.

 

[martrox]

That is correct - it was at 2x I believe but this is not what hurts performance. Runing the CPU and memory at such ASYNC clocks (2:1) kills processing performance as well as the fact that you are not able to acheive high clockspeeds with a high HTT link.

There are no such thing as async memory with A64s the memory is always running in sync with the CPU/memory controler.

For whatever reason, running memory on either a 939 or 754 A64 at any setting other than 1 to 1 will result in a loss of computing power. Try it yourself. You will see......

 

[incurable]

Tim there are memory dividers from that work with the HTT link in some way. The memory controller is always running at whatever clockspeed your CPU is running at.

With the K8, you have a reference clock that is produced by a BIOS-programmable PLL on the motherboard. This reference clock, which is usually 200 MHz, is used to derive both the HyperTransport clock (through processor-indicated, BIOS-overrideable multiplier) and the core clock. (reference clock * multiplier as indicated by the processor (max_FID) or overwritten by the BIOS) The memory controller works at full core speed, as you mentioned.

From this memory controller clock, there are several integer dividers available, set in the factory by AMD, which are used to control the clock of the memory interface. Available dividers include max_FID (PC3200 setting), 2*max_FID (PC1600) as well as two dividers to support PC2700 and PC2100 memory. The latter two are selected to bring the memory clock as close as possible, while not exceeding, the specs for said memory. (133.3 / 166.7 MHz)

Summing this up, while there's no direct, 1:1 connection between the HT base clock and the memory interface speed, both are derived from the same reference clock through various multipliers and dividers. Personally, I chalk the whole K8 sync/asnyc misconception up to sloppy BIOS interface programmers.

For whatever reason, running memory on either a 939 or 754 A64 at any setting other than 1 to 1 will result in a loss of computing power. Try it yourself. You will see......

Of course! Changing the memory dividers from the one used for PC3200 (core clock/max_FID = reference clock, what you like to call "1:1") to any other divider will result in a slower memory clock, no doubt having an impact on performance.