Is the future of the Ryzen line dependent on memory bandwidth?

Grimlaking

2[H]4U
Joined
May 9, 2006
Messages
3,245
I think that will be the sweet spot for the Ryzen line of CPU's going forward. Right now the better memory speed you can eek out the better CPU performance gains are as compared to other manufacturers memory solutions.

So as the Ryzen line matures and we see 12 and 16 core Ryzens come out..

Do you think a 12 core solution will be...

4+4+4 at 3 memory channels.

And the 16 core at...
4+4+4+4 at 4 memory channels.

And more importantly do you think the additional available memory bandwidth will directly impact the IPC for the CPU, OR is it all about bus/memory speed and not total throughput?

I personally hope that the 12 and 16 core solutions really blow the doors off and equal or exceed Intel CPU IPC performance. This will put the Ryzen CPU's into more gaming and other high performance logic systems and not just ruling the roost for systems that need high Physical I/O because of increased PCIe lanes.

I think that AMD by making a server class chip for workstations will shoot itself in the foot by not extracting the additional performance to show how it bests intel on more fronts.
 
When you have different numbers of memory channels, you need different sockets. It does not make sense for AMD to have 3 different sockets. If there is to be 12 and 16 core models, the 12 core will most likely be the 16 core with one core disabled per module.

From what I have read, it appears that CCX performance is tied to memory speed. CCX performance is generally irrelevant until data starts crossing between module caches, at which point it becomes a huge bottleneck. That scenario happens mostly in games and not on workstation/productivity workloads. It would be better if CCX performance could be decoupled from memory performance and tweaked independently.

Ryzen IPC isn't going to increase. What will happen is that it will be extremely competitive with the high core count, low clocked locked Xeons. Haswell-EP tops out at 3.5 ghz, well within the 4 ghz limits of Ryzen, and IPC is essentially equivalent.
 
I don't think memory bandwidth matters all that much. X99 has quad channel DRAM, but anytime those memory bandwidth figures are brought up, the pat answer is that it doesn't make any difference over dual channel.
 
If they can get their Infinity Fabric to run at 2x the memory speed I think there would be far fewer concerns about RAM speed.

This may have to wait for their Zen+ (Zen 2?) iteration, but asking a bus inside the chip to run at ~2.13 - ~3.6ghz doesn't seem unreasonable for the current process node. It may increase power consumption and TDP a bit.
 
The value for Ryzen in the server and high-end workstation space will be price. The CCX concept makes scaling up to more cores a less expensive proposition. Ryzen will offer similar performance to a number of Xeon CPUs, at a much lower price point (though still high compared to what desktop CPUs cost). Also, Ryzen's power usage drops dramatically at lower clock speeds (say, 3GHz or less), due as much to process as to anything else. So lower-clocked, high core count Ryzen parts are likely to be an excellent choice for servers. For high-end workstations, you may see 12 and 16 core parts at a higher clock speed, and these would probably be excellent for on-site rendering, or render farms, again likely far cheaper than Intel equivalents.
 
When you have different numbers of memory channels, you need different sockets. It does not make sense for AMD to have 3 different sockets. If there is to be 12 and 16 core models, the 12 core will most likely be the 16 core with one core disabled per module.
Correct me if I'm wrong but that's not true... For the AM4X (I'll just call it X for Extreme to denote the 12+ Core chips), they could easily have all of these chips be equipped with 4-channel memory controllers, all use the same sockets, but leave it up to the manufacturers (like it has been) whether or not to utilize them that way. I can offer you no more perfect of an example than what we've already seen in the past!
AMD Chips, they have been dual-channel designs. Motherboard manufacturers offer boards that have 4 DIMM slots, thus allowing for 2-DIMMs per channel.
HOWEVER, that has not stopped them from making 2 DIMM slot boards when it comes to smaller ATX or lower end systems, as well as laptops.
Furthermore, that has not stopped them (well, system builders like HP/Dell/etc) from offering laptops with only ONE solitary DIMM slot. My damn HP ENVY x350 is an example of that stupidity. There's provisions for onboard DDR4, but they didn't offer it, and also left the BIOS configured in dual-channel mode which hampered performance even further... but I digress.

So we have a Dual Channel chip that is offered with 4-DIMM Dual-Channel, 2-DIMM Dual-Channel, and 1-DIMM Single Channel configurations.

Therefore, I see no reason why we can't have the exact same instances here with an AM4X system that comes in variants just the same. Workstation/Top Enthusiast boards with 8-DIMM Quad-Channel, Enthusiast/Ultra-Gamer boards with 4-DIMM Quad-Channel, and even High-End Gamer mATX or ITX boards with 2-DIMM Dual-Channel (thought, I think that'd be hampering the system).

What I don't know, but would really like to think possible would be the ability to run in Triple-Channel mode as well. That would provide far far more flexibility, particularly in pricing, by being able to offer 8-DIMM Q-Ch, 6-DIMM T-Ch, 4-DIMM Q-Ch, 3-DIMM T-Ch, 2-DIMM D-Ch.

Again, I just don't know enough about memory controllers to know if that sort of thing is possible, but given how things work on Graphics Cards it seems plausible. We've had 512bit Bus, cut down to 384bit, cut down to 256bt, down further to 128bit, and even 64bit on the "this is just a video card, not a graphics card" models. Ignorance provides for fun speculation, though :)


I think that will be the sweet spot for the Ryzen line of CPU's going forward. Right now the better memory speed you can eek out the better CPU performance gains are as compared to other manufacturers memory solutions.

So as the Ryzen line matures and we see 12 and 16 core Ryzens come out..

Do you think a 12 core solution will be...

4+4+4 at 3 memory channels.

And the 16 core at...
4+4+4+4 at 4 memory channels.

And more importantly do you think the additional available memory bandwidth will directly impact the IPC for the CPU, OR is it all about bus/memory speed and not total throughput?

I personally hope that the 12 and 16 core solutions really blow the doors off and equal or exceed Intel CPU IPC performance. This will put the Ryzen CPU's into more gaming and other high performance logic systems and not just ruling the roost for systems that need high Physical I/O because of increased PCIe lanes.

I think that AMD by making a server class chip for workstations will shoot itself in the foot by not extracting the additional performance to show how it bests intel on more fronts.
I think it's safe to assume that EVERY Threadripper (AM4X, AM44, whatever you want to call it) would likely be a Quad-Channel, be it 12 or 16 (or possibly even 8 core, though that'd be a waste of a chip since 1/2 would be disabled and so I think unlikely...) It just seems like a lot more work than is needed and would only hurt themselves by limiting a chip in hardware to Triple-Channel.

AMD likes flexibility as we've seen. Logically, if we consider those previous fake designs, there could indeed be two different sockets, but I would think that would simply come down to a matter of Single and Dual socket motherboards where the extra pin counts are for chip-to-chip communication like the server models. Yet, I again can't help but point out previously what AMD had done. Their dedicated "Workstation" boards were Socket C32 (LGA1207) and the "Server" models were Socket G34 (LGA1974), the latter of which was intended for servers but due to single socket being part of that design it did become featured on E-ATX boards for Workstation usage. (Like the Supermicro H8SGL board)

At any rate, some of this is ignorant speculation, some of it is hope, but either way I'd like to see AMD break the mold here by giving us something that DOES have that wide range of DIMM/Channel configurations and all through same pincount socket. :)
 
Theoretically possible, yes. However, no motherboard manufacturer will shoot themselves in the foot making a gimped board like that. They will make all of the channels available, and I imagine that adding channels is not very expensive.
 
I personally would love to see a quad channel 128gb supported 16 core chip. Basing on current limits based on dual "cpu" chips with 4 cores per.. Upping that to a quad CPU chip with 4 cores and 4 memory controllers could make a lot of sense..

Or we could have a real freaking problem with it because of latency when having to deal with 4 independent memory banks.. That does kind of confuse me on dual channel systems.. but I'll leave it be and trust the chip makers because I don't want to compare dual channel memory to SLI configurations on video cards. ;)
 
because I don't want to compare dual channel memory to SLI configurations on video cards. ;)
lmao I feel like that was a shot at me, but I'm not offended either way and found it humorous :p

Theoretically possible, yes. However, no motherboard manufacturer will shoot themselves in the foot making a gimped board like that. They will make all of the channels available, and I imagine that adding channels is not very expensive.
And yet... they shoot at their feet quite often it seems? (These are all AM4 and MicroATX, but I can easily find AMD and Intel from past generations, and pretty certain in full ATX as well)
ASRockAM4_06.jpg
ImgW.ashx


I actually had a couple other AM4 boards ready but then, by what I can only deem fate... I came across this X58 board (I had found a bunch of 2-slot ones as well but this just drives my point home best lmao)
Intel_BLKDX58SO_Chipset-Intel_X58_LGA-1366_16Gb_DDR3-1600MHz_ATX_Motherboard.jpg


That's right... It's an x58 with 4 DIMMs :) Three are each on a channel I presume, and then the 4th is piggybacking like usual.
And you'll also notice that, while it's not a motherboard I've ever seen before, those are not shitty components. Those caps on the DRAM and CPU are all Panasonic OSCON. Which after looking at the image name and then back at the image it hit me.... INTEL is who made that. o_0
Here's a much higher res version.
While they may have sourced different capacitors instead of going will nearly an entire array of OSCON, that's definitely the same board! lol


Honestly though, I don't mean it to come across like I'm trying to start shit. You have a valid point, one that I agree with in the sense of compromises like that shouldn't generally be made, but it's the world we live in. Everything comes with different option packages. We have a Camaro with a V8. A Camaro with a V6. A Camaro SS with a high end V8. A Camaro ZL-1 with an even HIGHER performance V8. Back in the 80s then even came with 4cyl, if you can imagine that... Same damn 2.5L I4 "Iron Duke" engine that came in the Fieros! lmao Point is, we all want something, and sometimes in order to get that, sacrifices are made. We may want a nice shiny brand new Camaro SS but... budgets n' all... so we have to settle for a just as nice and shiny brand new Camaro with a V6. To further draw parallels... just because it's a V6 in that new Camaro, doesn't mean it's going to be a slouch... The 5th Gen's (2010-2015) put out 312HP! That's more than my 1993 Firebird Formula's (and early C4 Vette's) LT1 V8 which put out 305HP! ;)
 
Honestly though, I don't mean it to come across like I'm trying to start shit. You have a valid point, one that I agree with in the sense of compromises like that shouldn't generally be made, but it's the world we live in. Everything comes with different option packages. We have a Camaro with a V8. A Camaro with a V6. A Camaro SS with a high end V8. A Camaro ZL-1 with an even HIGHER performance V8. Back in the 80s then even came with 4cyl, if you can imagine that... Same damn 2.5L I4 "Iron Duke" engine that came in the Fieros! lmao Point is, we all want something, and sometimes in order to get that, sacrifices are made. We may want a nice shiny brand new Camaro SS but... budgets n' all... so we have to settle for a just as nice and shiny brand new Camaro with a V6. To further draw parallels... just because it's a V6 in that new Camaro, doesn't mean it's going to be a slouch... The 5th Gen's (2010-2015) put out 312HP! That's more than my 1993 Firebird Formula's (and early C4 Vette's) LT1 V8 which put out 305HP! ;)

OT, but man that Iron Duke motor was a dog. Drove one of those for a while, and you'd get better performance out of a f*cking lawnmower. I hear you on the car budgets. Went with a '15 Mustang GT 5.0... but stripped to the bone. No options. Manual transmission. V8 and a stick, basically. Lot of folks thought that was dumb. "Dude, you can get a premium with the leather and sh*t for only $3k more!" *shrug* That's $3k more for engine upgrades.
 
OT, but man that Iron Duke motor was a dog. Drove one of those for a while, and you'd get better performance out of a f*cking lawnmower. I hear you on the car budgets. Went with a '15 Mustang GT 5.0... but stripped to the bone. No options. Manual transmission. V8 and a stick, basically. Lot of folks thought that was dumb. "Dude, you can get a premium with the leather and sh*t for only $3k more!" *shrug* That's $3k more for engine upgrades.
That's what's called the "Stripper Model"... and not because it has a nice ass, but because all the luxuries are 'stripped' off. Those who are truly into performance love to find those models just as much as someone likes to find a fully-optioned out model. While I sincerely doubt you can find a car with OUT power windows these days, in the early 90s that was like 30-50lbs of saved weight!

Unless you had the Fiero GT with the shitty 2.8 V6 that was in the Camaro's and the S-10's, that I regretfully owned.
Which is what made the Fiero such a lil screamin machine lol Came with the V6, but at a whole bunch of weight savings.
 
Unless you had the Fiero GT with the shitty 2.8 V6 that was in the Camaro's and the S-10's, that I regretfully owned.

...and I had one of those to, in my old '84 S-10 Blazer >.<. You folks are finding all the sh*t engines I've ever owned in one thread.
 
That's what's called the "Stripper Model"... and not because it has a nice ass, but because all the luxuries are 'stripped' off. Those who are truly into performance love to find those models just as much as someone likes to find a fully-optioned out model. While I sincerely doubt you can find a car with OUT power windows these days, in the early 90s that was like 30-50lbs of saved weight!

You got it. But yeah... still came with power windows. You know, the stripper models these days are actually pretty lavishly equipped compared to the old days. I mean, the center stack has little computer screen with a damn 1/4 mile and 0-60 track app built in. And from tests I did at the dragstrip, it's pretty accurate. Crazy to think that's a *BASE* model.
 
Many motherboard manufacturers thought 2 dimms per channel were unnecessary for X79. That proved to be an unpopular decision as many chose the 8 dimm boards over the 4 dimm boards. They learned their lesson and when X99 came around, almost all boards were launched with 8 dimm slots. The ones that didn't were either bottom barrel boards or had size constraints (i.e. mATX). The board real estate needed for 3 vs 4 dimms is not that large, so it makes little sense to design a board for what is essentially an extremely niche market (extreme budget HEDT).

Edit: going with the OT topic… I am extremely tempted to LS swap a Fiero at some point in the future. Probably when I have more money than sense :D
 
Last edited:
Which is what made the Fiero such a lil screamin machine lol Came with the V6, but at a whole bunch of weight savings.

Yeah mabey when it decided to run on any given Tuesday and didn't manage to puke it's guts out in the driveway on the following Wednesday, it was a little pocket rocket.
 
Yeah mabey when it decided to run on any given Tuesday and didn't manage to puke it's guts out on the following Wednesday, it was a little pocket rocket.

In the far heavier S-10 Blazer, it was a dog. Lived in LA at the time, and there's a grade up on I-5, north of the city, going up toward Lancaster. I'd have to get in the slow lane and chug up the grade at 35-40 mph with the semi trucks. Motor couldn't handle it. Sometimes the semis would pass me. So much lol. And it never ran right. Sometimes to get it to start, I'd have to stick a clothespin on the choke valve, to hold it open a second.
 
In the far heavier S-10 Blazer, it was a dog. Lived in LA at the time, and there's a grade up on I-5, north of the city, going up toward Lancaster. I'd have to get in the slow lane and chug up the grade at 35-40 mph with the semi trucks. Motor couldn't handle it. Sometimes the semis would pass me. So much lol. And it never ran right. Sometimes to get it to start, I'd have to stick a clothespin on the choke valve, to hold it open a second.

I was living in Fl. at the time so the only hills I saw were driveways and bridges, and it didn't like those at all. Ripped the front airdam off twice in the first month of ownership.
It was great for the beach though being all fiberglass till you hit a puddle then it would stall out.
 
Do you think a 12 core solution will be...

4+4+4 at 3 memory channels.

And the 16 core at...
4+4+4+4 at 4 memory channels.
While possible, the limit will likely be set by the socket. I doubt it's worth adding a three channel design over just using the four and not populating all the slots. The only way I see it happening is if a lot of faulty memory controllers show up and they attempt to salvage dice.

more importantly do you think the additional available memory bandwidth will directly impact the IPC for the CPU, OR is it all about bus/memory speed and not total throughput?

I personally hope that the 12 and 16 core
More cores will obviously want more bandwidth for certain workloads. Contention is always harmful to IPC.

From what I have read, it appears that CCX performance is tied to memory speed. CCX performance is generally irrelevant until data starts crossing between module caches, at which point it becomes a huge bottleneck. That scenario happens mostly in games and not on workstation/productivity workloads. It would be better if CCX performance could be decoupled from memory performance and tweaked independently
We still need to wait and see Naples or the 12/16 core designs. Theory being fabric scaling should distribute the load better with a mesh.
 
The IMC on Ryzen probably only second to what we have seen on Kaby Lake. AMD did an excellent job in the IMC arena.
 
Do you think it will ever do the ultra high speed DDR4 before Zen+? I mean, if we can get 3200 C14 running on most boards it would be nice but I'd like to see some of the really fast stuff benched.
I bet we'll have 3600 multi by end of may, and maybe even SubTiming support. We already have people overclocking to DDR4-3600 at CL12s (granted, they have DDR4-4000+ CL14 kits... lol). And those are using the much lower memory multis as a basis, for the tighter subtimings! I'd even dare say we'll have 3800 by years end. Whether or not that'll be on B1 stepping is to be seen, but I sure hope we will, as this was supposed to be a multi-year build for me and MSI already screwed me out of BCLK on my Titanium I think (I've seen it hidden in the 1.41 BIOS, but I dunno if it'll be for v1.1 revision of motherboards), so I'll need that official support if that's the case.

My 3200 15-15-15 kit works just fine from the v1.1 shipping BIOS on the Titanium so if you (people in general) have Samsung B-die memory, then I feel like 3200 CL14 (if your modules support it) is a gauranteed doable... It'll just take motherboard makers to get the BIOS jiggered around each board in particular.
 
Do you think it will ever do the ultra high speed DDR4 before Zen+? I mean, if we can get 3200 C14 running on most boards it would be nice but I'd like to see some of the really fast stuff benched.
Not sure what "ultra high speed DDR4" is.
 
Well we've got what, 4000+ Mhz sticks out now? I had no problem running 3600 on Skylake. Just seems like Ryzen scales well with memory speed so it would be interesting to see how far it can go.
I do not know the upper end of the IMC's abilities right now. Most of the issues I THINK we are seeing currently pertain to the UEFI and AGESA microcode. We are still testing, but I have not seen anything stable over 3200 on my test bench.
 
Ok, that was what I thought. Probably have to wait for refresh but considering this is the first iteration of the architecture hopefully it will improve like it has on the intel side. I'm hoping to get 3200 stable on my rig, my 1700X seems to do 2933 perfectly fine but the board doesn't go any higher (B350). I just got in my new Gigabyte K7 yesterday so I'm hoping I can push for more with that board.
Dude you need to read Chews thread at XS :| lol They're running RAM up to 3400 on Dual-Rank DIMMs. Here's one at lower, but still... (that's overclocked mind you, so even better on the timings with the lower divider)

screen525.jpg
 
Dude you need to read Chews thread at XS :| lol They're running RAM up to 3400 on Dual-Rank DIMMs. Here's one at lower, but still... (that's overclocked mind you, so even better on the timings with the lower divider)

screen525.jpg

Yeah, but I'm talking plug and play for all boards not something you have to fiddle with to get it to post. It seems like a lot of people especially with the 1700 are struggling to even hit 3000mhz. This was the same way when Skylake came out if you remember and after a year or so we had people running 3800-4000mhz with some tweaks and 3200-3666 on plain XMP.
 
Yeah, but I'm talking plug and play for all boards not something you have to fiddle with to get it to post. It seems like a lot of people especially with the 1700 are struggling to even hit 3000mhz. This was the same way when Skylake came out if you remember and after a year or so we had people running 3800-4000mhz with some tweaks and 3200-3666 on plain XMP.
That's why I said give it till the end of the year and I suspect we'll have 3800, as that's rumored to be what Ryzen's IMC is capable of, or I guess at least the capability of the one they licensed.
 
Back
Top