Latency vs clock speed

[delshay]

Which one do you prefer?

[mc6809e]

I prefer predictability.

[MagerValp]

Originally Posted by mc6809e

I prefer predictability.

With that username I would expect that you prefer to have the clock controlled externally.

[delshay]

I was the first to bring the worlds fastest 128MB '28ns' EDO Memory x4.

Now I am about to bring the worlds fastest 1GB DDR memory with the 'lowest latency'.

This is preliminary, as there are faults still to be cleared, but it is up and running @227Mhz. This is the slowest memory module of 4 & only 1 memory module has qualify as a true PC-3700.

The worlds first PC-3700 1GB with 2-2-2-5 timing to be completed within a few months. 233Mhz (466Mhz) is the maximum speed allowed for this this project and if it get's any faster it will automatically drop into 1 clock timings.

Note above DDR(s) already has CL1.5 profile.

Does anyone know if the SAM computer can handle 1 clock cycle DDR? I already have ultra fast 100 pin DDR (custom built).

........Ultra Low Latency DDR......

Next project when completed SDRAM...

[meynaf]

Low latency vs higher clock speed ? I don't prefer one over the other. For me the best is a nicely balanced mix of both.

[delshay]

I prefer low latency before raising the clock. This way the processor/GFX spend less time waiting & wasting clock cycles, which also keeps chipset at a lower clock with little or no overclocking.

This is why there is no DDR5, their are saying it has poorer timing and is not worth it at this time.

HBM memory on GFX card(s) has shown you don't need high clock speed.

Raising the clock is a easy way to increase performance, but focus here is about getting latency timing as low as possible, then raise the clock. The highest clock frequency does not always give best performance & I have shown in the pass that low latency can sometimes outperformed higher clock frequency.

Don't know if I can do the same with this project, but early indication shows it's going to be somewhere between PC-4000 & PC-4400, but this is preliminary and can change. This project is also stuck at 227MHz until the slowest memory module improves.

Here there has already been 1 clock cycle testing, but with too many errors at this time, but it is improving, so I can't rule this out.

This is why I need help in someone editing the motherboard BIOS so I can test 'CL1'.

[tom256]

Originally Posted by delshay

I prefer low latency before raising the clock. This way the processor/GFX spend less time waiting & wasting clock cycles, which also keeps chipset at a lower clock with little or no overclocking.

This is why there is no DDR5, their are saying it has poorer timing and is not worth it at this time.

HBM memory on GFX card(s) has shown you don't need high clock speed.

Raising the clock is a easy way to increase performance, but focus here is about getting latency timing as low as possible, then raise the clock. The highest clock frequency does not always give best performance & I have shown in the pass that low latency can sometimes outperformed higher clock frequency.

Don't know if I can do the same with this project, but early indication shows it's going to be somewhere between PC-4000 & PC-4400, but this is preliminary and can change. This project is also stuck at 227MHz until the slowest memory module improves.

Here there has already been 1 clock cycle testing, but with too many errors at this time, but it is improving, so I can't rule this out.

This is why I need help in someone editing the motherboard bios so I can test 'CL1'.

Benefits from latency depend from what access it. GFX cards doesn't care about it much and mostly speed is critical here, because GPU is moving hundreds of megabytes data at one time. Critical here is throughput so Memory speed and data path wide. This is why HBM is so fast, it has data path 4096b what means in one clock it can sent 4kb of data. Old memory has mostly 128-512b patch, what means sending less data in one clock.

Old AthlonXP was very sensitive to CL latency. It was lot of fun to bring lowest latency with possible high frequency (Winbond BH5 334F with Corsair "Black Widow" Rules). Anyway below CL2 there was also no benefits.
For modern CPU you need to find sweet spot between CL and memory MHz. Both values influence CPU performance, but not significantly.

About video memory DDR5, DDR4 RAM already including this technology.
VideoDDR4 is not the same like normal DDR4.

[delshay]

Winbond BH5 334F I do believe only allows for 512MB per module. This project is 1GB per module, never being done at these speed.

The world record was held by Corsair PC-4400 2GB kit clocking the lowest latency at 223MHz with 2-2-2-5 timings. These custom memory modules has obliterated that record & should get faster.

CL1 is there, but does not show up in software, but will show up in benchmark. Depending on set-up sometime CL1 can be slower, it seems to be hardware/software set-up and CL1 can sometimes show CL2.5 (slower) as it is not detected correctly.

It's not just CL that's increasing performance, it's "TRCD", this is inflicting extreme heavy damage and driving up performance.

NOTE: be aware these memory modules also has "Enhanced Bandwidth" built into the SPD this is also inflicting extreme heavy damage and driving up performance.

To take advantage of this all timings must be set to "Auto" and will boot automatically into 2-2-2-5 timings with Enhanced Bandwidth kicking-in.

Here is a review of corsair pc-4400 2GB kit which did hold the world record with the lowest latency.

http://www.pcstats.com/articleview.c...id=1862&page=7

[delshay]

I have encountered some more problems on this project & one of two problems is now fixed.

need to fix last remaining problem before I can even think about 1-1-1-x timings.

4GB is up & running at 225MHz with 2-2-2-5 timings, no issues whatsoever & is "cleared" at this speed, but speed has taken a setback from the previous 227MHz & is limited to the new lower speed due to another problem being detected.

Just when you think everything is ok, up pops another problem.

it also seems these memory modules will operate at 2.65v or even 2.6v somewhere between 215MHz & 225MHz, but at PC-3700 speed for sure it needs 2.7v. I will look into this further when full speed is up and running & bring the voltage requirements down.

[delshay]

The world fastest 1GB DDR module with "2-2-2-5" timing is starting to show itself.. Awaiting it's top end speed.

[delshay]

(Preliminary) Awaiting Next Updates DDR PC-3700.

The above screenshot you just set Bus Speed & Voltage (2.6V-2.7v Max).. Nothing else must be touched.

http://www.crucial.com/usa/en/memory...-speed-latency

SDRAM & DDR2 Next in-line (in-order of updates)

[delshay]

DDR SPECIFICATION CHANGE UPDATE:

http://www.amiga.org/gallery/index.php?n=3946