Zekk Pacus' AMD Ryze(n) hardware thread, July 2017

**Shaftoes** · October 13 2020, 03:07:25 PM

Originally Posted by Daneel Trevize

Originally Posted by Shaftoes

No discounts on RAM

https://www.amazon.co.uk/dp/B083TRRT16/
£45 for 2x8GB 3200Mhz CL16 Black Crucial Ballistix

Saw it. Not the ram I am looking for sadly.

**Daneel Trevize** · October 13 2020, 04:34:29 PM

The 3600Mhz was apparently also listed for £47 while it was in stock. Some are reporting it reappearing for pre-order at that price, ETA Oct 20th.

**Shaftoes** · October 14 2020, 11:52:11 AM

Finally some fast RAM on offer:

https://www.amazon.co.uk/gp/product/...?ie=UTF8&psc=1

**Daneel Trevize** · October 14 2020, 12:38:00 PM

You call 3600Mhz CL17 fast?? £100 for 16GB?? Please, link something where this kit reliably goes >4200Mhz and that is worth doing...

3200 CL16 (1/3200 *16) is 5.0 insert-correct-time-unit-here, 3600 CL17 is 4.72. <6% difference of a single primary timing. Not of bandwidth, or total latency, or noticable load time or fps. It's over twice the price of that actual deal I linked yesterday. Oh and has RGB. wut

**halka** · October 14 2020, 12:52:24 PM

Originally Posted by Daneel Trevize

You call 3600Mhz CL17 fast?? £100 for 16GB?? Please, link something where this kit reliably goes >4200Mhz and that is worth doing...

3200 CL16 (1/3200 *16) is 5.0 insert-correct-time-unit-here, 3600 CL17 is 4.72. <6% difference of a single primary timing. Not of bandwidth, or total latency, or noticable load time or fps. It's over twice the price of that actual deal I linked yesterday. Oh and has RGB. wut

Incidentally, what was the highest RAM freq that didn't prompt AMD chipsets to downclock? I'm not really familiar with the correct detailed terminology, my knowledge stops somewhere around "FSB", it's just that I recall I've seen something like that discussed over at Steve's channel.

**Shaftoes** · October 14 2020, 01:11:58 PM

Originally Posted by Daneel Trevize

You call 3600Mhz CL17 fast?? £100 for 16GB?? Please, link something where this kit reliably goes >4200Mhz and that is worth doing...

3200 CL16 (1/3200 *16) is 5.0 insert-correct-time-unit-here, 3600 CL17 is 4.72. <6% difference of a single primary timing. Not of bandwidth, or total latency, or noticable load time or fps. It's over twice the price of that actual deal I linked yesterday. Oh and has RGB. wut

So I am fairly ignorant when it comes to RAM.

Can you explain why 3200MHz C16 is comparable to 3600MHz C17? I was told the general rule of thumb is to get RAM that gets as close to your CPU clock speed as possible, and that anything above 3600 is overkill? Was there a step change in architecture between DIMMs that upped the efficiency?

**Daneel Trevize** · October 14 2020, 02:31:57 PM

Do you want the wikipedia rabbit hole https://en.wikipedia.org/wiki/CAS_latency and https://en.wikipedia.org/wiki/Dynami...#Memory_timing
or the Buildzoid/ActualHardwareOverclocking ramble real-world vids? https://www.youtube.com/c/ActuallyHa...locking/videos
Maybe specifically ~9mins and beyond of https://www.youtube.com/watch?v=xSAFs-t0DCg

If your CAS, also known as CL, is a cycle count, and you're doing 3.2billion "transfers" per second, your timings latency (as a wall-clock-time period measure) is the product of cycle period and cycle count.
Because the physical silicon & circuits can only realistically manage a certain speed by design and 1m&1 constraints, you end up fitting that true value (e.g. 47ns, again could be wildly wrong in my order of magnitude but it's a time unit) into whatever pairing of frequency and count that produces a compatible period. E.g. 3.2Ghz * 16 cycles = 50ns, or 3.6Ghz * 17 cycles = 47.2ns. 3.6 is greater than 3.2 but not the ~12% you might think it is overall.
Generally "faster" as in GHz or GTransfers RAM has slower latency timings, and you end up also having to have "looser" subtimings, giving an extra cycle or few here and there (AKA extra small time slices) to ensure physical signals & shit actually have real time to get done before the rest of the hardware is assuming they are ready for another round.

DDR generations tend to trade clock GHz for latency periods, so maybe 100% higher GT as a base standard, but also slower timings (just not fully 2x slower, so there's actual improvement made here too).
DDR3-800 is 0.8GT @ 12.5CL, DDR4-1600 is 1.6GT @ 15CL, DDR5 3.2GT @ 17.5. Remember 3.2GT/s is running at 1.6GHz, because DoubleDataRate. It's confusing because everyone markets RAM falsely using "GHz" for effective transfer rate, not actual clock.
They set the new base speed, lower V for general operation, add in actual logical improvements to maximise theoretical bandwidth and timing usage, and then try claw back from a slower initial latency to get all this clever new stuff done per cycle to also be faster in the end.

As to ideal target speed vs processor, talk of FrontSideBuses kinda went away with Integrated Memory Controllers, and more multipliers of your base nominal 133MHz clock are involved.
For Ryzen, you now consider the Infinity Fabric clock (FClk, ~1766-1900Mhz) and the synchronised ratio with the Memory clock. DDR means 1.8 GHz MClk matches 3.6GT/s, and so whether you're achieving that on both sides(RAM & CPU) or not interacts with having a 1:1 or 1:2 ratio, or some other async combo with extra latency to allow for the mismatch.
I am no expert on this, follow links above. Also maybe check pages 5 & 6 here https://www.corsair.com/corsairmedia...ckingGuide.pdf

**rufuske** · October 14 2020, 02:56:27 PM

Originally Posted by Daneel Trevize

Do you want the wikipedia rabbit hole https://en.wikipedia.org/wiki/CAS_latency and https://en.wikipedia.org/wiki/Dynami...#Memory_timing
or the Buildzoid/ActualHardwareOverclocking ramble real-world vids? https://www.youtube.com/c/ActuallyHa...locking/videos
Maybe specifically ~9mins and beyond of https://www.youtube.com/watch?v=xSAFs-t0DCg

If your CAS, also known as CL, is a cycle count, and you're doing 3.2billion "transfers" per second, your timings latency (as a wall-clock-time period measure) is the product of cycle period and cycle count.
Because the physical silicon & circuits can only realistically manage a certain speed by design and 1m&1 constraints, you end up fitting that true value (e.g. 47ns, again could be wildly wrong in my order of magnitude but it's a time unit) into whatever pairing of frequency and count that produces a compatible period. E.g. 3.2Ghz * 16 cycles = 50ns, or 3.6Ghz * 17 cycles = 47.2ns. 3.6 is greater than 3.2 but not the ~12% you might think it is overall.
Generally "faster" as in GHz or GTransfers RAM has slower latency timings, and you end up also having to have "looser" subtimings, giving an extra cycle or few here and there (AKA extra small time slices) to ensure physical signals & shit actually have real time to get done before the rest of the hardware is assuming they are ready for another round.

DDR generations tend to trade clock GHz for latency periods, so maybe 100% higher GT as a base standard, but also slower timings (just not fully 2x slower, so there's actual improvement made here too).
DDR3-800 is 0.8GT @ 12.5CL, DDR4-1600 is 1.6GT @ 15CL, DDR5 3.2GT @ 17.5. Remember 3.2GT/s is running at 1.6GHz, because DoubleDataRate. It's confusing because everyone markets RAM falsely using "GHz" for effective transfer rate, not actual clock.
They set the new base speed, lower V for general operation, add in actual logical improvements to maximise theoretical bandwidth and timing usage, and then try claw back from a slower initial latency to get all this clever new stuff done per cycle to also be faster in the end.

As to ideal target speed vs processor, talk of FrontSideBuses kinda went away with Integrated Memory Controllers, and more multipliers of your base nominal 133Mhz clock are involved.
For Ryzen, you now consider the Infinity Fabric clock (FClk, ~1766-1900Mhz) and the synchronised ratio with the Memory clock. DDR means 1.8 GHz MClk matches 3.6GT/s, and so whether you're achieving that on both sides(RAM & CPU) or not interacts with having a 1:1 or 1:2 ratio, or some other async combo with extra latency to allow for the mismatch.
I am no expert on this, follow links above. Also maybe check pages 5 & 6 here https://www.corsair.com/corsairmedia...ckingGuide.pdf

Tl;dr Get 3600, but not that Kingston pair as it's CL17 so in reality pretty slow. As a rule of a thumb Kingston HyperX were never fast but throughout my life I haven't found a single motherboard that didn't happily run with them. If you want fast you want 3600 CL16, Daneel's previous link of Crucial Ballistix had them but you're paying almost double. If you know a bit about overclocking you can grab the 3200 CL16 pair and set them to run at 3600 CL16 and there's a pretty big chance they'll do so, as these are pretty good. If you're unlucky you will have to raise voltage and do some more timing tweaks.

**Daneel Trevize** · October 14 2020, 03:04:58 PM

Honestly, memory- (and most) overclocking is a relic from when companies didn't fully realise the marketing angle of the vocal DIY enthusiasts and halo products, nor have the motivation to push official clocks close to the hardware's limits (thanks to Moore's Law, they just shipped what worked, Intel locked out overclocking, and planned for the next release cycle that'd officially support the same product running faster).

Modern CPUs and GPUs have complex self-management for clocks, power, thermals, sleep states, etc. In fact, as witnessed by the latest Nvidia 30 series, and ClockTuner for Ryzen, you can underclock the things to reduce the power draw and trade ~1% performance for easy double digit % less power, heat, noise. Sometime actually even wiggling the performance slightly up by slightly reducing clocks and V, because the 30 series are power/heat limited out of the box, and some CPU benchmarks are few-cores-limited enough that uneven thermal budgets via Boost speeds do better for a given cooling solution.

P.S. Consider that CL14 vs CL15 is a >7% improvement at a given clock frequency. 16 vs 15 is 6.67%, it grows larger the lower CL gets because 1 cycle less is always a fraction of the total number needed for this timing, until you magically get it from 2 to 1, a 100% speedup.

See these 10900k (not Ryzen) numbers on the leftmost (1st column of Primaries is CAS/CL) and rightmost (benchmark score) columns for a probably-easier-to-digest format. https://www.youtube.com/watch?v=25NW8cHNrgA
Tightening timings by 2 or 3 matched a ~400MHz (10%) RAM clock increase. But the overall takeaway is about Single vs Dual Rank (which is not directly the same as stick count or dual memory channels)...

**rufuske** · October 14 2020, 03:16:00 PM

Originally Posted by Daneel Trevize

Honestly, memory and most overclocking is a relic from when companies didn't fully realise the marketing angle of the vocal DIY enthusiasts and halo products, nor have the motivation to push official clocks close to the hardware's limits (thanks to Moore's Law, they just shipped what worked, Intel locked out overclocking, and planned for the next release cycle that'd officially support the same product running faster).

Modern CPUs and GPUs have complex self-management for clocks, power, thermals, sleep states, etc. In fact, as witnessed by the latest Nvidia 30 series, and ClockTuner for Ryzen, you can underclock the things to reduce the power draw and trade ~1% performance for easy double digit % less power, heat, noise. Sometime actually even wiggling the performance slightly up by slightly reducing clocks and V, because the 30 series are power/heat limited out of the box, and some CPU benchmarks are few-cores-limited enough that uneven thermal budgets via Boost speeds do better for a given cooling solution.

Pretty accurate when we're talking about CPUs, completely true about GPUs. But memory is still the realm of either load xmp profile or spend whole day figuring out if that CAS, tRCD, tRP, tRAS etc combination is just stable enough to post, just stable to start windows or something that you can use all the time (or maybe just raise voltage a bit and start over). And then after few days of this you realize you could have paid that 30 extra and just stick to xmp.

**Shaftoes** · October 14 2020, 03:27:03 PM

Originally Posted by rufuske

Originally Posted by Daneel Trevize

Do you want the wikipedia rabbit hole https://en.wikipedia.org/wiki/CAS_latency and https://en.wikipedia.org/wiki/Dynami...#Memory_timing
or the Buildzoid/ActualHardwareOverclocking ramble real-world vids? https://www.youtube.com/c/ActuallyHa...locking/videos
Maybe specifically ~9mins and beyond of https://www.youtube.com/watch?v=xSAFs-t0DCg

If your CAS, also known as CL, is a cycle count, and you're doing 3.2billion "transfers" per second, your timings latency (as a wall-clock-time period measure) is the product of cycle period and cycle count.
Because the physical silicon & circuits can only realistically manage a certain speed by design and 1m&1 constraints, you end up fitting that true value (e.g. 47ns, again could be wildly wrong in my order of magnitude but it's a time unit) into whatever pairing of frequency and count that produces a compatible period. E.g. 3.2Ghz * 16 cycles = 50ns, or 3.6Ghz * 17 cycles = 47.2ns. 3.6 is greater than 3.2 but not the ~12% you might think it is overall.
Generally "faster" as in GHz or GTransfers RAM has slower latency timings, and you end up also having to have "looser" subtimings, giving an extra cycle or few here and there (AKA extra small time slices) to ensure physical signals & shit actually have real time to get done before the rest of the hardware is assuming they are ready for another round.

DDR generations tend to trade clock GHz for latency periods, so maybe 100% higher GT as a base standard, but also slower timings (just not fully 2x slower, so there's actual improvement made here too).
DDR3-800 is 0.8GT @ 12.5CL, DDR4-1600 is 1.6GT @ 15CL, DDR5 3.2GT @ 17.5. Remember 3.2GT/s is running at 1.6GHz, because DoubleDataRate. It's confusing because everyone markets RAM falsely using "GHz" for effective transfer rate, not actual clock.
They set the new base speed, lower V for general operation, add in actual logical improvements to maximise theoretical bandwidth and timing usage, and then try claw back from a slower initial latency to get all this clever new stuff done per cycle to also be faster in the end.

As to ideal target speed vs processor, talk of FrontSideBuses kinda went away with Integrated Memory Controllers, and more multipliers of your base nominal 133Mhz clock are involved.
For Ryzen, you now consider the Infinity Fabric clock (FClk, ~1766-1900Mhz) and the synchronised ratio with the Memory clock. DDR means 1.8 GHz MClk matches 3.6GT/s, and so whether you're achieving that on both sides(RAM & CPU) or not interacts with having a 1:1 or 1:2 ratio, or some other async combo with extra latency to allow for the mismatch.
I am no expert on this, follow links above. Also maybe check pages 5 & 6 here https://www.corsair.com/corsairmedia...ckingGuide.pdf

Tl;dr Get 3600, but not that Kingston pair as it's CL17 so in reality pretty slow. As a rule of a thumb Kingston HyperX were never fast but throughout my life I haven't found a single motherboard that didn't happily run with them. If you want fast you want 3600 CL16, Daneel's previous link of Crucial Ballistix had them but you're paying almost double. If you know a bit about overclocking you can grab the 3200 CL16 pair and set them to run at 3600 CL16 and there's a pretty big chance they'll do so, as these are pretty good. If you're unlucky you will have to raise voltage and do some more timing tweaks.

Thanks, this is great advice.

I don't really want to go down the hardware rabbithole. I'm already autistic enough thanks.

**Daneel Trevize** · October 14 2020, 03:33:24 PM

Originally Posted by rufuske

you could have paid that 30 extra and just stick to xmp.

Even then XMP is just an official way to market the headroom the hardware always had, while board makers can still shirk honouring anything above JEDEC base standard or processor generation base support speeds.

Yeah, just (check the Qualified Vendor List) get 3600Mhz ~CL16 DDR4, check your mobo is actually running it at the proper speed & timings (some don't have or honour all XMProfile subtimings, for better or worse stability), and get on with using the system.

Unless you're running databases, or professionally gaming as a job, you're not going to notice or benefit from a few % memory performance, as an even smaller % app performance, as an even smaller impact on win rate or fun.

I will repeat though that it's appreciably better to slightly underclock & undervolt GPUs to be able to tune your fan curves for a much quieter and/or cooler system. Given they they usually have <90mm fans, unlike the 140mm+ you can be running everywhere else.
Mr sleek SFF setups (https://www.youtube.com/c/OptimumTech/videos) showed removing the GPU shroud and using bottom-mounted 120mm case fans can be significant in this regard too.

**Overspark** · October 14 2020, 08:04:32 PM

Originally Posted by rufuske

Pretty accurate when we're talking about CPUs, completely true about GPUs. But memory is still the realm of either load xmp profile or spend whole day figuring out if that CAS, tRCD, tRP, tRAS etc combination is just stable enough to post, just stable to start windows or something that you can use all the time (or maybe just raise voltage a bit and start over). And then after few days of this you realize you could have paid that 30 extra and just stick to xmp.

Or just download the Ryzen DRAM Calculator, enter your hardware (you do need to know which chips are on your DIMMs), pick the safe settings, enter those into your BIOS and get something far better than the garbage settings that the XMP usually provides. XMP timings are often geared more towards Intel CPU's than AMD CPU's too.

**Smegs** · October 14 2020, 08:37:55 PM

Originally Posted by Overspark

Or just download the Ryzen DRAM Calculator, enter your hardware (you do need to know which chips are on your DIMMs), pick the safe settings, enter those into your BIOS and get something far better than the garbage settings that the XMP usually provides. XMP timings are often geared more towards Intel CPU's than AMD CPU's too.

Spent about a week dialing in DRAM frequncies and timings to get to the fabled 1900 F/U/Mclk on the 3900X, after many, many hours of crying and nashing of teeth, got it to something approaching stable. A week after that got introduced to DRAM calculator and not only improved the timings but improved stability as well in about 15 minutes.... FML.

Now have some of the Patriot Memory Viper Steel Series DDR4 16GB 4400MHz (PVS416G440C9K) after Buildzoid ran the numbers and ... guess i'm going down the rabbit hole again. If you are happy with fiddling with timings and not just XMP'ing, for approx £100 this ram is great, although not really needed for the non-enthusiast.