Likely yes. I have 4x16GB 3200MHz CL16 E-Die kit and I was easily able to OC it to 3600MHz CL14. Unfortunately my 3900X won't boot with FCLK at 1900MHz (haven't tried tweaking voltage yet). I will be attempting to do 3733MHz CL14 when I get time.
Guaranteed to be very little if any. Any time you see a test of new higher bandwidth memory with higher latency, the results are always extremely small gains if any, sometimes its even slower depending on the latency.
True, but given that APUs have traditionally benefitted more from high bandwidth memory than other platforms, I think there could be some scope for AMD CPU's.
The APUs do see an increase with faster RAM, but the increase is not substantial. Usually single digit FPS increases. Instead of spending hundreds of dollars on fast RAM it's better to buy a dedicated GPU.
This is false. The gains are ridiculously substantial! With say a Ryzen 5 3400G, you are looking at near 1-1-1 gaming performance increase to memory bandwidth increase. Just going from 2133MHz to 3200MHz increases performance by like 20-30%.
Of course, there are substantial gains to be had going from 2133 to 3200, however, the gains start to diminish rapidly after that. The CU count and speed on the APUs is tailored to the expected memory bandwidth. I expect that the next APU, Renoir, will retain 11-12 VEGA2 CUs, perhaps even a few more with LOWER clocks for power savings. They will likely target supporting 3200-3600 speed DDR4, which is still highly restrictive.
Going forward, I suspect that very high speed DDR4 sticks will be the same situation as very high end DDR3 sticks were, since the standard is going to be on the way out, they will never hit high enough production numbers to make them cost effective. Maybe we'll see DDR4-4200 hit decent volume and decent prices, and there will be an APU that supports those speeds, but that's still not a whole lot of GPU bandwidth.
I wonder when AMD will put a L4 onto the CPU package. It seems with all these cores, memory fetch is becoming a very real issue again. Something like 1-2GB as an L4/L3 on its own silicon, and with their new standalone IO chip, it could be done.
Intel had Crystalwell thing on some CPUs and although it was mainly intended to support the iGPU, it was also shared with the CPU cores as a last layer cache. In CPU benchmarks, it didn't make a lot of difference. Obviously we are talking about different processor architectures and the chiplet configuration is different too, but performance from caches rapidly diminish with capacity. I don't think there would be a need to even go as big as 1GB to find a sweet spot in performance versus cost. 128-256GB of fast/low latency memory would probably do the trick, but that's assuming the existing caches are already a bottleneck. I recall AMD went pretty big with cache on the Zen-derived cores so it may not even be a performance problem as-is unless you're also suggesting said large cache also support integrated graphics as Intel did with the Iris/Crystalwell combination. iGPUs tend to be starved for memory bandwidth (nevermind the shared thermal/cooling situation) and with AMD usually packing a comparably quick graphics package compared to Intel, giving said GPU more bandwidth may help. It also could endanger sales of dGPUs if performance reaches high enough so there might not be an incentive to implement such a feature when AMD would likely rather sell you a dedicated Radeon card.
I loved the Broadwell desktop CPUs because of the L4. In ram intensive compute tasks, it provided practically unlimited bandwidth and they ended up faster than Skylake, which has higher IPC but was unable to be kept fed adequately unless you had serious ram for the time. I kinda see similar happening with the generous L3 in Zen 2 CPUs. It helps a lot, although currently with the split L3 per CCX that hinders performance than if it were unified per CCD. The cache also helps mitigate against the crippled write bandwidth. I can't see either company offering more channels at consumer level, so big caches will be the way to go to scale to higher core counts.
Maybe, but that's unlikely. The restrictive IF bandwidth will limit it's performance impact per CCD, it would only hide a little latency. In fact, it'll add a layer of latency as the IO die will need to decide if the requested memory area has been locally cached, or is in the DRAM itself. It would have to be massive to make a measurable impact.
Well, the zen2 I/O dies have a large amount of real-estate with no apparent need. AMD is pretty quiet on what is consuming all those transistors. It might not qualify as an actual cache but I suspect there is a lot of block RAM in there.
I'd like to see whether or not the core-to-core latency changes when zen2's memory controller is attached to 3700 RAM vs. +3700 RAM. As in, does the IF slow down when the RAM is at a speed above the rated speed of the memory controller and the memory controller has to switch to a 2:1 ratio.
You don't need to wait and see, this has already been tested. It does indeed switch to a 2:1 ratio. Unless you OC the fabric, anything above 3733 results in a latency penalty. You'd likely need this fast of RAM just to overcome it.
People with Logic don`t buy Ferrari... there Are allways peoples who has money to do whatever They like. So somebody will buy these and have something that most other people don`t have... yet. Three four years from now and these speed can be normal enthusiast speeds in Intel 13000 series and amd 6000 series ;)
Kind of amusing and absurd...all they have to do is sell maybe one or two sets, call it quits, and it makes "good" advertising PR for them for awhile...;) My current ram is 2x8GBs, Patriot Viper Elite PV416G320C6K @3200mhz 16 16 16 36 1T @ 3733Mhz 18-19-19-19-38-68 1T @1.35v, x570 Aorus Master, 3600X. Ram is P95-tested stable--100%. I actually have my BCLK slightly overclocked to 100.59MHz, which pushes the ram bus to 3757MHz. My jaw dropped when I discovered I could do this kind of ram OC successfully. The difference in performance is noticeable and measurable--I used the AIDA 64 benches to chart the performance increase over stock 3200 with much tighter timings.
The price difference from regular DDR4 3200 1.2V modules is enough to upgrade to a Threadripper platform instead, which will provide significantly more bandwidth due to quad-channel and shorter latency values, probably at a much lower power consumption.
This really isn't usefull for anything besides record-setting overlcocks.
I would like to see how DDR4 5000 compares to lower spec'd memory with a lower cas ratings. Its tough to rationalize higher speed memory when they all seem to work at a higher cas rating = more clock ticks to transfer data. Is there a way to calculate the improvement of a faster memory as it pertains to a lower speed memory but at a (for example) cas 7? If I have a cas 8 3200 memory and a 4000 cas 15 memory is available, is it worth the upgrade? Does the faster speed overcome the higher cas?
We’ve updated our terms. By continuing to use the site and/or by logging into your account, you agree to the Site’s updated Terms of Use and Privacy Policy.
Facebook
Twitter
RSS
38 Comments
Back to Article
PeachNCream - Friday, October 11, 2019 - link
I guess if you don't like your bank account balance, buying RAM that costs that much is a good way to get some "Vengeance" on it.Targon - Friday, October 11, 2019 - link
So, Ryzen 9 3900X for $499(MSRP), plus a $350+ motherboard, plus $1225? 5000CL18 is very good, but then you have to wonder if it can do 3800CL14.TheWereCat - Friday, October 11, 2019 - link
Likely yes.I have 4x16GB 3200MHz CL16 E-Die kit and I was easily able to OC it to 3600MHz CL14.
Unfortunately my 3900X won't boot with FCLK at 1900MHz (haven't tried tweaking voltage yet).
I will be attempting to do 3733MHz CL14 when I get time.
regsEx - Sunday, October 13, 2019 - link
What is your motherboard?TheWereCat - Sunday, October 13, 2019 - link
X570 Crosshair Hero VIII Wi-Fimrvco - Friday, October 11, 2019 - link
I'd just be curious to see what impact this RAM has on real world performance vs 3x00 and 4x00 DDR4 RAM.goatfajitas - Friday, October 11, 2019 - link
Guaranteed to be very little if any. Any time you see a test of new higher bandwidth memory with higher latency, the results are always extremely small gains if any, sometimes its even slower depending on the latency.ads295 - Friday, October 11, 2019 - link
True, but given that APUs have traditionally benefitted more from high bandwidth memory than other platforms, I think there could be some scope for AMD CPU's.goatfajitas - Friday, October 11, 2019 - link
Yeah, but who would be spending money on high end RAM just to get better integrated graphics ? Much better to buy a discreet VC.mooninite - Friday, October 11, 2019 - link
The APUs do see an increase with faster RAM, but the increase is not substantial. Usually single digit FPS increases. Instead of spending hundreds of dollars on fast RAM it's better to buy a dedicated GPU.Cooe - Saturday, October 12, 2019 - link
This is false. The gains are ridiculously substantial! With say a Ryzen 5 3400G, you are looking at near 1-1-1 gaming performance increase to memory bandwidth increase. Just going from 2133MHz to 3200MHz increases performance by like 20-30%.Korguz - Monday, October 14, 2019 - link
is there an article or review you can refer to ?lightningz71 - Tuesday, October 15, 2019 - link
Of course, there are substantial gains to be had going from 2133 to 3200, however, the gains start to diminish rapidly after that. The CU count and speed on the APUs is tailored to the expected memory bandwidth. I expect that the next APU, Renoir, will retain 11-12 VEGA2 CUs, perhaps even a few more with LOWER clocks for power savings. They will likely target supporting 3200-3600 speed DDR4, which is still highly restrictive.Going forward, I suspect that very high speed DDR4 sticks will be the same situation as very high end DDR3 sticks were, since the standard is going to be on the way out, they will never hit high enough production numbers to make them cost effective. Maybe we'll see DDR4-4200 hit decent volume and decent prices, and there will be an APU that supports those speeds, but that's still not a whole lot of GPU bandwidth.
MrEcho - Friday, October 11, 2019 - link
I wonder when AMD will put a L4 onto the CPU package. It seems with all these cores, memory fetch is becoming a very real issue again. Something like 1-2GB as an L4/L3 on its own silicon, and with their new standalone IO chip, it could be done.PeachNCream - Friday, October 11, 2019 - link
Intel had Crystalwell thing on some CPUs and although it was mainly intended to support the iGPU, it was also shared with the CPU cores as a last layer cache. In CPU benchmarks, it didn't make a lot of difference. Obviously we are talking about different processor architectures and the chiplet configuration is different too, but performance from caches rapidly diminish with capacity. I don't think there would be a need to even go as big as 1GB to find a sweet spot in performance versus cost. 128-256GB of fast/low latency memory would probably do the trick, but that's assuming the existing caches are already a bottleneck. I recall AMD went pretty big with cache on the Zen-derived cores so it may not even be a performance problem as-is unless you're also suggesting said large cache also support integrated graphics as Intel did with the Iris/Crystalwell combination. iGPUs tend to be starved for memory bandwidth (nevermind the shared thermal/cooling situation) and with AMD usually packing a comparably quick graphics package compared to Intel, giving said GPU more bandwidth may help. It also could endanger sales of dGPUs if performance reaches high enough so there might not be an incentive to implement such a feature when AMD would likely rather sell you a dedicated Radeon card.porina - Saturday, October 12, 2019 - link
I loved the Broadwell desktop CPUs because of the L4. In ram intensive compute tasks, it provided practically unlimited bandwidth and they ended up faster than Skylake, which has higher IPC but was unable to be kept fed adequately unless you had serious ram for the time. I kinda see similar happening with the generous L3 in Zen 2 CPUs. It helps a lot, although currently with the split L3 per CCX that hinders performance than if it were unified per CCD. The cache also helps mitigate against the crippled write bandwidth. I can't see either company offering more channels at consumer level, so big caches will be the way to go to scale to higher core counts.Rudde - Saturday, October 12, 2019 - link
There are rumors that Zen 3 has unified L3 per CCD.FreckledTrout - Friday, October 11, 2019 - link
If they move from 14nm to TSMC's 7nm for the IO die they would have a lot of extra die space I could see 256MB of L4.lightningz71 - Tuesday, October 15, 2019 - link
Maybe, but that's unlikely. The restrictive IF bandwidth will limit it's performance impact per CCD, it would only hide a little latency. In fact, it'll add a layer of latency as the IO die will need to decide if the requested memory area has been locally cached, or is in the DRAM itself. It would have to be massive to make a measurable impact.evanh - Saturday, October 12, 2019 - link
Well, the zen2 I/O dies have a large amount of real-estate with no apparent need. AMD is pretty quiet on what is consuming all those transistors. It might not qualify as an actual cache but I suspect there is a lot of block RAM in there.lightningz71 - Tuesday, October 15, 2019 - link
Remember, the IO die is ALSO used as the chipset on the x570. Some of the dark circuits are used for that.Hul8 - Saturday, October 12, 2019 - link
Threadripper 2990WX's 64 threads weren't memory starved even with four slower memory channels.3000 series added full and equal memory access from any chiplet and allows for higher memory frequencies, so I doubt there should be any problems.
The unified 32MB chiplet L3 on Zen 3 (instead of 16MB L3 exclusive to each CCX) will also help.
peevee - Monday, October 14, 2019 - link
"Threadripper 2990WX's 64 threads weren't memory starved even with four slower memory channels."In any task? Because different tasks use memory differently.
Great_Scott - Friday, October 11, 2019 - link
The best memory for a cheap IGP system is too expensive for said system.That said, I'm curious how Ice Lake / Raven Ridge graphics performs with this RAM.
ads295 - Friday, October 11, 2019 - link
Haha. Nail on the headballsystemlord - Friday, October 11, 2019 - link
I'd like to see whether or not the core-to-core latency changes when zen2's memory controller is attached to 3700 RAM vs. +3700 RAM.As in, does the IF slow down when the RAM is at a speed above the rated speed of the memory controller and the memory controller has to switch to a 2:1 ratio.
evernessince - Friday, October 11, 2019 - link
You don't need to wait and see, this has already been tested. It does indeed switch to a 2:1 ratio. Unless you OC the fabric, anything above 3733 results in a latency penalty. You'd likely need this fast of RAM just to overcome it.haukionkannel - Saturday, October 12, 2019 - link
We need ddr4 7200 so that we can run fabric at 1800 ;)Waiting for that! And lotto winnery...
jabber - Saturday, October 12, 2019 - link
So a jump from 134FPS to 136FPS?Amazing.
Ironchef3500 - Monday, October 14, 2019 - link
+1mjz_5 - Saturday, October 12, 2019 - link
People with logic don’t buy products like this.haukionkannel - Sunday, October 13, 2019 - link
People with Logic don`t buy Ferrari... there Are allways peoples who has money to do whatever They like. So somebody will buy these and have something that most other people don`t have... yet. Three four years from now and these speed can be normal enthusiast speeds in Intel 13000 series and amd 6000 series ;)WaltC - Sunday, October 13, 2019 - link
Kind of amusing and absurd...all they have to do is sell maybe one or two sets, call it quits, and it makes "good" advertising PR for them for awhile...;) My current ram is 2x8GBs, Patriot Viper Elite PV416G320C6K @3200mhz 16 16 16 36 1T @ 3733Mhz 18-19-19-19-38-68 1T @1.35v, x570 Aorus Master, 3600X. Ram is P95-tested stable--100%. I actually have my BCLK slightly overclocked to 100.59MHz, which pushes the ram bus to 3757MHz. My jaw dropped when I discovered I could do this kind of ram OC successfully. The difference in performance is noticeable and measurable--I used the AIDA 64 benches to chart the performance increase over stock 3200 with much tighter timings.fotcorn - Monday, October 14, 2019 - link
der8auer already did a test with these modules and a Ryzen 3900X: https://www.youtube.com/watch?v=MLja1q-M4SUfotcorn - Monday, October 14, 2019 - link
TL;DR: Not really better than some 3600/3733 Mhz modules with good timings.ToTTenTranz - Monday, October 14, 2019 - link
The price difference from regular DDR4 3200 1.2V modules is enough to upgrade to a Threadripper platform instead, which will provide significantly more bandwidth due to quad-channel and shorter latency values, probably at a much lower power consumption.This really isn't usefull for anything besides record-setting overlcocks.
peevee - Monday, October 14, 2019 - link
Ryzen would need to be set into 2:1 mode for it to work, so the on-package IF is going to be 2500 instead of the normal speed (say, 3200), isn't it?Blaab1 - Wednesday, October 16, 2019 - link
I would like to see how DDR4 5000 compares to lower spec'd memory with a lower cas ratings. Its tough to rationalize higher speed memory when they all seem to work at a higher cas rating = more clock ticks to transfer data. Is there a way to calculate the improvement of a faster memory as it pertains to a lower speed memory but at a (for example) cas 7? If I have a cas 8 3200 memory and a 4000 cas 15 memory is available, is it worth the upgrade? Does the faster speed overcome the higher cas?