Zen 3 speculation - Page 4

mackerel · 2020-11-08, 20:43

There's a lot of discussion going on elsewhere about ram choice. It seems even for gaming going dual rank or 2DPC is showing a not insignificant difference, but there is a claim that 16GB single rank modules might be best? No details or test data on that yet, but could there be a new mechanism going on?

Also there's a claim MSI mobos may have difficulty running ram beyond 3200 with these CPUs currently.

I think for now we can just monitor the situation and see if there is any new understanding to come out of this.

Mark Rose · 2020-11-08, 21:21

I've seen the discussion, too. The bit about 16 GB single rank comes from Steve at Gamers Nexus in his concluding remarks:

Quote:

So after all this testing we called Wendell from Level1Techs, and we asked if he'd seen anything similar. He immediately said yes. In Wendell's experience behaviour has changed in Ryzen 5000. Four sticks of single rank 8 GB memory performs better than two. And two sticks of single rank 16 GB of memory performs best of all in Wendell's experience. We've relatively few sticks of single rank 16 GB memory laying around because we never needed higher capacity DIMMs for anything except for our production machines.

I bet there was a misunderstanding at some point. It makes sense that two sticks of dual rank memory would be fastest on a board that has daisy-chain topology between slots. On a T-topology board I don't see how it would make a difference. I guess we'll have to wait and see.

Mark Rose · 2020-11-08, 21:28

This comment on that video is worth reading:

https://www.youtube.com/watch?v=-UkG...ujcRloV4AaABAg

Xyzzy · 2020-11-08, 23:14

Our understanding is that if you are trying to get the highest frequency and don't care about performance then you want the fewest ranks possible. So you would want 2×8GB-SR or (rare) 2×16GB-SR. (Maybe even 1×8GB-SR!) Having a high frequency doesn't always translate to a faster system, although having FCLK and MCLK synced helps. This is why the best motherboards for memory overclocking have only two DIMM slots. Less signal traces means faster speeds. MSI has even gone to SMD DIMM slots on some boards to reduce signal length.

If you are looking at actual benchmarks and not just a "ludicrous" frequency speed, you want four ranks (2×DR) used. Ideally you would want eight ranks (4×DR) used but the memory controller might balk at that load when overclocking. So 2×16GB-DR would work. You get memory interleaving AND good overclocking.

If you just want things to work and be a good value and be fast, get 2×16GB-DR or 4×8GB-SR DDR4-3200C16 sticks. Take the money you saved and bump the CPU up a tier, or put that money into a better video card.

If you need lots (128GB) of memory, you can get that easily by running at stock speeds. We ran 4×32GB DDR4-3200 DR sticks with our 3950X with no issues. The memory controller can handle that signal load but probably only at stock (or in some cases JEDEC) speeds.

The numbers we have seen are (potentially) a 10-15% boost from SR to DR. This is higher than the boost from 3200MT/s to 4000MT/s in most workloads. 4000MT/s is not a trivial load for a memory controller, BTW.

To be blunt, we feel that overclocking is a hassle with new stuff. The CPUs are programmed to boost like video cards so they have all sorts of dynamic circuitry doing all sorts of fancy things. Overclocking takes time and it could change based off of ambient conditions or age or any number of other factors. The headroom that the CPU and memory have built in to deal with those factors is what you are dialing out when you overclock. Sure, it is "free", but if you are intelligent then you know that you can't rob Peter to pay Paul. Every decision has consequences.

If you did a super overclock job on both the CPU and the memory, with the FCLK and MCLK synced and all, what is the absolute most you could expect to increase in performance? Use the stock CPU and dual rank DDR4-3200C16 memory for a baseline. We suspect maybe 25-30% tops and only in certain benchmarks. In real life, you wouldn't be able to tell.

Now, we all know overclocking is fun, so do it if that is what you want to do. We like to keep current on overclocking techniques but on our own systems we don't apply them other than to just play with the settings a bit. We have a set of DDR4-3200C14 memory that will run all day long at DDR4-4000C19 but the small performance gain is not worth taking a chance on using up the memory's available headroom.

Our time is worth at least a dollar an hour, so running things stock pays off eventually. (Yes, we know XMP memory is an overclock. We make an exception for that.)

Earlier we posted that we bought a DDR4-4000C16 DR memory kit for our 5600X. We jumped on that because we knew it would sell out quick (it did) and at that point we weren't sure how things were set up. Now that we know that things are "business as usual" we will just go back to using DDR4-3200 memory. Since all the retailers have an extended return (in some cases to February) we have no issue with ordering stuff we might not use. In an ideal world we'd live near a Microcenter!

We posted before that running "ECO" mode on our 3950X made very little difference in productivity even though it greatly lowered power consumption. We plan to use this mode with our 5600X as well. AMD is pushing their CPUs hard to show a gain from generation to generation but they are running them far north of their optimum efficiency. Overclocking is the same thing, really. You pump in 50% more power to gain maybe 5-10% in speed. There is an electrical (transistor) engineering term we can't remember that refers to this point where things are balanced and most efficient. Maybe someone can remind us what that is.

Ryzen 5000 is nothing new or special WRT memory except the fact the "infinity fabric" may boost higher. They didn't change the memory controller at all. The memory differences you see are because the L3 cache is more accessible.

In the end, all this talk is about headroom and how you use it. Some people want to use it "now" and take their chances later. Some people want things to "just work" no matter what. The good news is you have a choice!

https://www.tomshardware.com/reviews...gs,6310-2.html

See chart below.

Xyzzy · 2020-11-08, 23:40

PS - Our first overclock was 200% which you could feel interactively. Over time we learned that it allowed us to make twice as many mistakes in half the time!

http://www.davebiz.com/wiki/CoCo3FPGA

Quote:

The original Color Computer 3 contained an MC68B09E processor which was rated at a maximum clock speed of 2 MHz. It had the ability to clock the CPU at the normal speed of .89 MHz for compatibility with the older CoCo models and a double speed of 1.78 MHz. BASIC CoCo programmers refer to the method of overclocking as the 'double speed poke'. Using the BASIC 'POKE' command, a couple of registers could be manipulated to change the clock speed of the CPU.

Mark Rose · 2020-11-09, 05:14

I'm still debating on getting a new system this year. Nothing I do in particular needs it, but new toys are fun.

I'm using a 6 core i7-5820K as a desktop.

There are some games I play that would benefit from higher single threaded performance, but I don't feel the pain enough to upgrade.

M344587487 · 2020-11-09, 11:23

Quote:

Originally Posted by Mark Rose

I'm still debating on getting a new system this year. Nothing I do in particular needs it, but new toys are fun.

I'm using a 6 core i7-5820K as a desktop.

There are some games I play that would benefit from higher single threaded performance, but I don't feel the pain enough to upgrade.

DDR5 is coming which puts Zen 3 in a tricky spot for those that don't upgrade often. It'll be 2 years before DDR5 is viable, roadmaps have intel supporting it in 2021 with Sapphire Rapids (at least in servers) and AMD in Q1 2022 with Zen 4. intel's roadmaps cannot be trusted. AMD is in a much more interesting position in that if they absolutely had to they could backport Zen 4 to AM4 as a stopgap for consumers if DDR5 availability is dire.

AMD or intel should do everyone a favour and add an extra RAM channel with each new generation. The other vendor will follow suit and before you know it we'll need more cores to keep up. Dual channel was enough when quad core was high end but now it's barely present even in the low end.

henryzz · 2020-11-09, 14:10

I think the whole channels thing will have a bit of a shakeup with DDR5. There are 2 32/40- bit channels in each DIMM for DDR5. Hopefully, this will lead to CPU designs being more flexible with the number of channels they support.
The voltage will also be managed by the DIMMs themselves which could mean greater flexibility for high-end chips. ECC is also built into the DIMMs as standard which could aid in overclocking and reliability. I would choose to be an early adoptor rather than tail end user of DDR4. I always regretting getting DDR2 for my Q6600 when I could have got a DDR3 motherboard. DDR2 was harder/more expensive to get later on when I needed capacity upgrades.

Xyzzy · 2020-11-14, 05:43

We have the 5600X system up and running. (We ended up with two 5600X CPUs. We are not yet sure what we will do with the second one.)

So far the CPU tops out at 55W, 55°C and ~3950MHz all-core no matter what we throw at it. There is probably a BIOS setting we set wrong.

We needed to do a CPU-less BIOS "flashback" since the board's default BIOS didn't recognize the 5600X. We were pleasantly surprised at how painless the procedure was.

The board we ended up with (C8I) "trained" to a 1T command rate for our memory. We have never gotten 1T to work on any previous board with this memory and certainly not automatically.

This particular 5600X will go to 1900 FCLK and 3800 MCLK but any faster it will not post. We didn't try adding any voltage or anything. (We left almost everything on auto.) We are currently running it at 1600 FCLK and 3200 MCLK.

As soon as we figure out why the CPU isn't boosting we will post some benchmarks. We also have a set of dual rank memory sticks scheduled to be delivered within a few days.

Current parts list:

CPU: AMD Ryzen 5 5600X 3.7 GHz 6-Core Processor
CPU Cooler: Noctua NH-U12S chromax.black 55 CFM CPU Cooler
Motherboard: Asus ROG Crosshair VIII Impact Mini DTX AM4 Motherboard
Memory: G.Skill Trident Z 16 GB (2 x 8 GB) DDR4-3200 CL14 Memory
Storage: Samsung 980 Pro 1 TB M.2-2280 NVME Solid State Drive
Video Card: NVIDIA GeForce RTX 3080 10 GB Founders Edition Video Card
Case: Fractal Design Define 7 Compact ATX Mid Tower Case
Power Supply: Fractal Design Ion+ 860 W 80+ Platinum Certified Fully Modular ATX Power Supply

Xyzzy · 2020-11-14, 16:37

Quote:

Originally Posted by Xyzzy

So far the CPU tops out at 55W, 55°C and ~3950MHz all-core no matter what we throw at it. There is probably a BIOS setting we set wrong.

We needed to set the PBO option from "auto" to "enabled". Now the system tops out at 110W, 95°C and ~4600MHz all-core.

Temporarily we have the system set up in our test bench, which greatly simplifies working on things.

PhilF · 2020-11-14, 17:06

Quote:

Originally Posted by Xyzzy

We needed to set the PBO option from "auto" to "enabled". Now the system tops out at 110W, 95°C and ~4600MHz all-core.

Temporarily we have the system set up in our test bench, which greatly simplifies working on things.

That's a cool test bench!

2020-11-08, 23:14	#37
Xyzzy "Mike" Aug 2002 3²·883 Posts	Our understanding is that if you are trying to get the highest frequency and don't care about performance then you want the fewest ranks possible. So you would want 2×8GB-SR or (rare) 2×16GB-SR. (Maybe even 1×8GB-SR!) Having a high frequency doesn't always translate to a faster system, although having FCLK and MCLK synced helps. This is why the best motherboards for memory overclocking have only two DIMM slots. Less signal traces means faster speeds. MSI has even gone to SMD DIMM slots on some boards to reduce signal length. If you are looking at actual benchmarks and not just a "ludicrous" frequency speed, you want four ranks (2×DR) used. Ideally you would want eight ranks (4×DR) used but the memory controller might balk at that load when overclocking. So 2×16GB-DR would work. You get memory interleaving AND good overclocking. If you just want things to work and be a good value and be fast, get 2×16GB-DR or 4×8GB-SR DDR4-3200C16 sticks. Take the money you saved and bump the CPU up a tier, or put that money into a better video card. If you need lots (128GB) of memory, you can get that easily by running at stock speeds. We ran 4×32GB DDR4-3200 DR sticks with our 3950X with no issues. The memory controller can handle that signal load but probably only at stock (or in some cases JEDEC) speeds. The numbers we have seen are (potentially) a 10-15% boost from SR to DR. This is higher than the boost from 3200MT/s to 4000MT/s in most workloads. 4000MT/s is not a trivial load for a memory controller, BTW. To be blunt, we feel that overclocking is a hassle with new stuff. The CPUs are programmed to boost like video cards so they have all sorts of dynamic circuitry doing all sorts of fancy things. Overclocking takes time and it could change based off of ambient conditions or age or any number of other factors. The headroom that the CPU and memory have built in to deal with those factors is what you are dialing out when you overclock. Sure, it is "free", but if you are intelligent then you know that you can't rob Peter to pay Paul. Every decision has consequences. If you did a super overclock job on both the CPU and the memory, with the FCLK and MCLK synced and all, what is the absolute most you could expect to increase in performance? Use the stock CPU and dual rank DDR4-3200C16 memory for a baseline. We suspect maybe 25-30% tops and only in certain benchmarks. In real life, you wouldn't be able to tell. Now, we all know overclocking is fun, so do it if that is what you want to do. We like to keep current on overclocking techniques but on our own systems we don't apply them other than to just play with the settings a bit. We have a set of DDR4-3200C14 memory that will run all day long at DDR4-4000C19 but the small performance gain is not worth taking a chance on using up the memory's available headroom. Our time is worth at least a dollar an hour, so running things stock pays off eventually. (Yes, we know XMP memory is an overclock. We make an exception for that.) Earlier we posted that we bought a DDR4-4000C16 DR memory kit for our 5600X. We jumped on that because we knew it would sell out quick (it did) and at that point we weren't sure how things were set up. Now that we know that things are "business as usual" we will just go back to using DDR4-3200 memory. Since all the retailers have an extended return (in some cases to February) we have no issue with ordering stuff we might not use. In an ideal world we'd live near a Microcenter! We posted before that running "ECO" mode on our 3950X made very little difference in productivity even though it greatly lowered power consumption. We plan to use this mode with our 5600X as well. AMD is pushing their CPUs hard to show a gain from generation to generation but they are running them far north of their optimum efficiency. Overclocking is the same thing, really. You pump in 50% more power to gain maybe 5-10% in speed. There is an electrical (transistor) engineering term we can't remember that refers to this point where things are balanced and most efficient. Maybe someone can remind us what that is. Ryzen 5000 is nothing new or special WRT memory except the fact the "infinity fabric" may boost higher. They didn't change the memory controller at all. The memory differences you see are because the L3 cache is more accessible. In the end, all this talk is about headroom and how you use it. Some people want to use it "now" and take their chances later. Some people want things to "just work" no matter what. The good news is you have a choice! https://www.tomshardware.com/reviews...gs,6310-2.html See chart below. Attached Thumbnails

2020-11-14, 05:43	#42
Xyzzy "Mike" Aug 2002 3²·883 Posts	We have the 5600X system up and running. (We ended up with two 5600X CPUs. We are not yet sure what we will do with the second one.) So far the CPU tops out at 55W, 55°C and ~3950MHz all-core no matter what we throw at it. There is probably a BIOS setting we set wrong. We needed to do a CPU-less BIOS "flashback" since the board's default BIOS didn't recognize the 5600X. We were pleasantly surprised at how painless the procedure was. The board we ended up with (C8I) "trained" to a 1T command rate for our memory. We have never gotten 1T to work on any previous board with this memory and certainly not automatically. This particular 5600X will go to 1900 FCLK and 3800 MCLK but any faster it will not post. We didn't try adding any voltage or anything. (We left almost everything on auto.) We are currently running it at 1600 FCLK and 3200 MCLK. As soon as we figure out why the CPU isn't boosting we will post some benchmarks. We also have a set of dual rank memory sticks scheduled to be delivered within a few days. Current parts list: CPU: AMD Ryzen 5 5600X 3.7 GHz 6-Core Processor CPU Cooler: Noctua NH-U12S chromax.black 55 CFM CPU Cooler Motherboard: Asus ROG Crosshair VIII Impact Mini DTX AM4 Motherboard Memory: G.Skill Trident Z 16 GB (2 x 8 GB) DDR4-3200 CL14 Memory Storage: Samsung 980 Pro 1 TB M.2-2280 NVME Solid State Drive Video Card: NVIDIA GeForce RTX 3080 10 GB Founders Edition Video Card Case: Fractal Design Define 7 Compact ATX Mid Tower Case Power Supply: Fractal Design Ion+ 860 W 80+ Platinum Certified Fully Modular ATX Power Supply Attached Thumbnails

2020-11-08, 20:43	#34
mackerel Feb 2016 UK 2⁴·5² Posts	There's a lot of discussion going on elsewhere about ram choice. It seems even for gaming going dual rank or 2DPC is showing a not insignificant difference, but there is a claim that 16GB single rank modules might be best? No details or test data on that yet, but could there be a new mechanism going on? Also there's a claim MSI mobos may have difficulty running ram beyond 3200 with these CPUs currently. I think for now we can just monitor the situation and see if there is any new understanding to come out of this.

2020-11-08, 21:28	#36
Mark Rose "/X\(‘-‘)/X\" Jan 2013 5561₈ Posts	This comment on that video is worth reading: https://www.youtube.com/watch?v=-UkG...ujcRloV4AaABAg

2020-11-09, 05:14	#39
Mark Rose "/X\(‘-‘)/X\" Jan 2013 29×101 Posts	I'm still debating on getting a new system this year. Nothing I do in particular needs it, but new toys are fun. I'm using a 6 core i7-5820K as a desktop. There are some games I play that would benefit from higher single threaded performance, but I don't feel the pain enough to upgrade.

2020-11-09, 14:10	#41
henryzz Just call me Henry "David" Sep 2007 Cambridge (GMT/BST) 11·17·31 Posts	I think the whole channels thing will have a bit of a shakeup with DDR5. There are 2 32/40- bit channels in each DIMM for DDR5. Hopefully, this will lead to CPU designs being more flexible with the number of channels they support. The voltage will also be managed by the DIMMs themselves which could mean greater flexibility for high-end chips. ECC is also built into the DIMMs as standard which could aid in overclocking and reliability. I would choose to be an early adoptor rather than tail end user of DDR4. I always regretting getting DDR2 for my Q6600 when I could have got a DDR3 motherboard. DDR2 was harder/more expensive to get later on when I needed capacity upgrades.