TechPowerUp RAM Latency Calculator Feedback

[BoggledBeagle]

So I think my confusion with UI is this for example. If I choose to calculate nanoseconds I expected to see variations of nanoseconds at a constant Data Rate or constant Latency. Does this make sense or am I just using this wrong?

View attachment 369483

No, that is calculation with two variables and one result and you get a table with combination of the two variables which produce the same result as a bonus.

 

[W1zzard]

Does this make sense or am I just using this wrong?

The table is intentionally reversed, to show you "what's the equivalent CL Latency" at various speed, because that's what I expect when entering the CL latency in the inputs. Otherwise the input that you want to provide is locked.

Happy to explore other ideas

 

[A Computer Guy]

The table is intentionally reversed, to show you "what's the equivalent CL Latency" at various speed, because that's what I expect when entering the CL latency in the inputs. Otherwise the input that you want to provide is locked.

Happy to explore other ideas

Ok I see now that makes sense. Perhaps the grid just needs a friendly label for it's purpose. I lost focus and thought the grid was my result. The users selected option is calculated by inputting the other two variables and as a bonus is the grid output. Ok got it. I guess it's fine the grid output is what confused me.

(edit) @W1zzard If you bold font the selected column and match result color I think those visual clues will make it easier to see.

Example:

My brain literally "lost focus" because the color of the output was gray.

 

[W1zzard]

and match result color

Done. Looking good

I can now see decimal points everywhere except in Edge and input field, where is a decimal comma.

This should be fixed now. Here I am in 2024, implementing my number inputs by hand ..

 

[Wirko]

I can't see when it would be useful to calculate data rate from the other two numbers. So, if others agree, consider removing that option.

 

[Count von Schwalbe]

The Up/Down arrows do not appear to work on the "Data Rate" and "Latency" fields, but the mousewheel does. The arrows work on the "Nanoseconds" field but the mousewheel does not.

Chrome/Windows.

 

[W1zzard]

The Up/Down arrows do not appear to work on the "Data Rate" and "Latency" fields, but the mousewheel does. The arrows work on the "Nanoseconds" field but the mousewheel does not.

Chrome/Windows.

Try now

 

[sneekypeet]

This may seem weird, but why the need to show all speeds and cas that will run that "real latency" rather than maybe what speed changes at that CAS do for real latency?

IE, I'd rather see 6000 C30 compared to 8200 C30 for real latency than 6000 C30 versus 8200 C41 to run that same real latency.

Neverming I found the button smh for being a bit slow. Is there a way to add a tenths place to the nanoseconds results?

 

[Count von Schwalbe]

Try now

Works now

 

[Chrispy_]

It seems 6400 for am5 is the standard at the moment at 10ns.

I'm finding 6000 CL30 EXPO kits are the best bang-for-buck over here.

I'm not sure about Zen5, but didn't AMD basically say that low-latency DDR5-6000 was the optimum (ie, the expectation) that all Zen4 Ryzens could handle at 1:1:1 ratios? IMO you might get 6400 at 1:1:1 but if you're asking for a good result from the silicon lottery at that point it's not going to work as a recommendation or (in my case) when you're buying RAM for dozens of builds at once and need ALL of them to work.

 

[Metroid]

I'm finding 6000 CL30 EXPO kits are the best bang-for-buck over here.

I'm not sure about Zen5, but didn't AMD basically say that low-latency DDR5-6000 was the optimum (ie, the expectation) that all Zen4 Ryzens could handle at 1:1:1 ratios? IMO you might get 6400 at 1:1:1 but if you're asking for a good result from the silicon lottery at that point it's not going to work as a recommendation or (in my case) when you're buying RAM for dozens of builds at once and need ALL of them to work.

6000 CL30 for Zen4, for Zen5 is 400 more, so the perfect for Zen5 = 6400 CL32.

 

[Chrispy_]

6000 CL30 for Zen4, for Zen5 is 400 more, so the perfect for Zen5 = 6400 CL32.

Is that official from AMD or just people trying higher frequencies and having success?

All I know is that AMD confirmed there are no changes to the IO die (MCLK) for Zen5 as they're 100% reusing the IO die from Zen4. Is Infnity fabric clock handled one the CCDs? I always vaguely thought that FCLK was managed and driven by the IO die (which is unchanged from Zen4) - since BIOSes put that under SoC sections rather than CPU overclocking/PBO+ sections.

 

[W1zzard]

Is that official from AMD

This is the official word:

https://imgur.com/LZep3RV

 

[D007]

You mad genius. I love it!

 

[Chrispy_]

This is the official word:

https://imgur.com/LZep3RV

Thanks for the clarification. Is that something that's published by AMD somewhere obvious, or is it hidden in the whitepaper of reviewer guidelines and only public domain because of people like you posting snippets in places like these forums?

So 6400 being the official sweet spot is nonsense, as I suspected - since anything above 6000 will default to a UCLK:MCLK ratio of 1:2, which is the rule of thumb I've been aware of ever since AM5 launched. Reusing the IO die for Zen5 means that the UCLK:MCLK behaviour was always going to be identical to Zen4. If people want to tinker with UCLK overclocking then they're entering the silicon lottery and luck of the draw, which disqualifies it from being the "sweet spot" and therefore irelevant to both this discussion, and the majority of people.

Interesting statement on them about 1:1:1 being no longer desirable, and to go auto:1:1, where "auto" is always whatever ratio results in 2000FCLK. I guess the ratio dividers are smarter for AM5's IO die compared to the AM4 IO die we had with Zen2/Zen3.

 

[Count von Schwalbe]

Interesting statement on them about 1:1:1 being no longer desirable, and to go auto:1:1, where "auto" is always whatever ratio results in 2000FCLK. I guess the ratio dividers are smarter for AM5's IO die compared to the AM4 IO die we had with Zen2/Zen3.

IIRC it's 2:3:3 for AM5 @6000 , but DDR4 wasn't fast enough to get there so it would default down to 1:2:2.

Also, 2000 FCLK and 4000 MCLK was pretty ambitious on AM4, so 1:1:1 worked best with 3600 RAM.

 

[Metroid]

Is that official from AMD or just people trying higher frequencies and having success?

All I know is that AMD confirmed there are no changes to the IO die (MCLK) for Zen5 as they're 100% reusing the IO die from Zen4. Is Infnity fabric clock handled one the CCDs? I always vaguely thought that FCLK was managed and driven by the IO die (which is unchanged from Zen4) - since BIOSes put that under SoC sections rather than CPU overclocking/PBO+ sections.

It is official, AMD said they improved Zen 5 internal memory controller by 400. So it was 6000, now is 6400. Some people said that even 6600 is getting 1/1. So it might be from 400 to 600 but this needs verification.

AMD Ryzen 9000 "Zen 5" CPUs To Offer Up To 6400 MT/s 1:1 & 8000 MT/s 1:2 4-DIMM DDR5 Memory Support, Run Cooler & Lower Power By Default

AMD's next-generation Ryzen 9000 "Zen 5" Desktop CPUs will run cooler, consume lower power & feature enhanced DDR5 OC capabilities.

wccftech.com

"AMD Ryzen 9000 “Zen 5” CPUs To Offer Up To 6400 MT/s 1:1 & 8000 MT/s 1:2 4-DIMM DDR5 Memory Support, Run Cooler & Lower Power By Default"

This is the official word:

https://imgur.com/LZep3RV

This is for Zen 4 right?

 

[W1zzard]

Is that something that's published by AMD somewhere obvious, or is it hidden in the whitepaper of reviewer guidelines and only public domain because of people like you posting snippets in places like these forums?

This is from the reviewer's guide (which is more like a "guidance", info, tips, non-binding). If the official word was higher than 6000, they would have picked other words

 

[Metroid]

DDR5 Memory Performance Scaling with AMD Zen 5

Our AMD Zen 5 Memory Scaling Review examines six different DDR5 memory speeds, including DDR5-8000, DDR5-6400 1:1, DDR5-6400 CL28 and others. We looked into how these speeds impact performance in gaming at four resolutions and a wide range of application workloads.

www.techpowerup.com

 

[W1zzard]

TechPowerUp

www.techpowerup.com

Alternate version that has 4 inputs, so you can scale a full set of timings easily

 

[Wirko]

All I know is that AMD confirmed there are no changes to the IO die (MCLK) for Zen5 as they're 100% reusing the IO die from Zen4.

Same isn't exactly same, hardware revisions happen all the time, and I'd expect manufacturing process to improve a bit in two years. If AMD is any smart, they're also doing some binning, so the best IODs end up in Ryzen 9000 processors.

 

[W1zzard]

Same isn't exactly same, hardware revisions happen all the time, and I'd expect manufacturing process to improve a bit in two years. If AMD is any smart, they're also doing some binning, so the best IODs end up in Ryzen 9000 processors.

I never heard them mention or hint at anything like that. No doubt, there's software improvements over time though

 

[Count von Schwalbe]

TechPowerUp

www.techpowerup.com

Alternate version that has 4 inputs, so you can scale a full set of timings easily

Perhaps field labels for the individual timings?

Or I guess it could be used for comparing different kits, so that would just be confusing.

 

[W1zzard]

Perhaps field labels for the individual timings?

Or I guess it could be used for comparing different kits, so that would just be confusing.

Yeah I had the labels, but then realized it could be used for a lot of other things, too

 

[A Computer Guy]

TechPowerUp

www.techpowerup.com

Alternate version that has 4 inputs, so you can scale a full set of timings easily

With some more inputs perhaps you could end up making a submission form for TPU users that helps build/update a RAM OC database based on ram kits TPU has reviewed? The best success I've had with RAM OC is looking at what others were able to do with similar ram kits to mine but often many details were lacking such as voltages and other memory options that may have been influential in the original posters success.

I often resort to using a sheet like below to see how the RAM is scaling as I change variables or trying to see what the UEFI auto behavior is trying to do. Below is an example of my DDR5 ECC kit. In this example my RAM simply was ok with scaling up to 6200MT/s without changing any voltages for a consistent % in improvement in ns across the board.