Ryzen 3000 will have DDr4-5000 mhz support?!

Actually you were talking about replacing the entirety of System RAM.



The Names of a process have very little connection with the actual physical size of features.

Samsung seems well on it's way to 3nm. They already have an early design kit to share with customers:
https://hardforum.com/threads/anandtech-samsung-announces-3nm-gaa-mbcfet-pdk-version-0-1.1981638/

I wasnt joking (fully) I mean I get what you are saying but this is from AMD directly

https://www.tomshardware.com/news/amd-3d-memory-stacking-dram,38838.html

upload_2019-5-17_19-27-18.png


upload_2019-5-17_19-29-5.png
 
Working on full 3D stacking doesn't mean ready soon. You also have the problem of heat. The more you stack the more heat you have in one space.

That HBM photo is HBM on interposer, not HBM on CPU.
 
Working on full 3D stacking doesn't mean ready soon. You also have the problem of heat. The more you stack the more heat you have in one space.

That HBM photo is HBM on interposer, not HBM on CPU.

Stop being a naysayer and let AMD innovate. They will figure this out.
 
Sorry reality is raining on your parade.

No, thereality is that you can't define what AMD/Intel can/can't engineer. Do you work there? I am just saying that so many people say this or that can't be done over the years, yet it ends up getting done. Never fails. I said I do not see how its possible that they can get down to 3nm but I didn't say they can't. I just dont know how they are going to do it.

Youre saying stacking "Can't be done". I am saying, I dont know how it can be done, but I am not saying it can't be done.

Reality also suggests that the multiple PhD's and Masters of engineering and computer science professionals that AMD pays hundreds of thousands in salaries to design and innovate have made slides showcasing the work and direction they are headed, and you are telling us that it can't be done.

So who's reality are you suggesting is raining on who's parade here?
 
Some of the 3d stacking stuff is a bit of an exploration to woo investors ;) . In the end it is about what is practical if you seen what has been released it is just to give an impression that some problems can be solved this way that does not mean they will be solved this way :) .
 
No, thereality is that you can't define what AMD/Intel can/can't engineer. Do you work there? I am just saying that so many people say this or that can't be done over the years, yet it ends up getting done. Never fails. I said I do not see how its possible that they can get down to 3nm but I didn't say they can't. I just dont know how they are going to do it.

Youre saying stacking "Can't be done". I am saying, I dont know how it can be done, but I am not saying it can't be done.

I didn't say that? I said this exactly: "Working on full 3D stacking doesn't mean ready soon"

I am just saying it doesn't look like it is coming soon. 3D stacking will likely come eventually.

AM5 should be very interesting. Since AMD tends to keep the same socket for a long time, I really want to see what kind of new benefits their next generation of mainstream socket will bring. Since it should be very forward looking.

Will it have more memory channels? Will the socket be bigger to support more chiplets and HBM?

Will it have some fanciful 3D stacked design including system memory, which is where you seem to be going?
 
I didn't say that? I said this exactly: "Working on full 3D stacking doesn't mean ready soon"

I am just saying it doesn't look like it is coming soon. 3D stacking will likely come eventually.

AM5 should be very interesting. Since AMD tends to keep the same socket for a long time, I really want to see what kind of new benefits their next generation of mainstream socket will bring. Since it should be very forward looking.

Will it have more memory channels? Will the socket be bigger to support more chiplets and HBM?

Will it have some fanciful 3D stacked design including system memory, which is where you seem to be going?

Ahh ok you wrote in such a way it was misconstrued. Got ya

Either way we look at it these chips are getting very powerful indeed. Let's hope they dont become alive and kill us all Haha
 
While extra bandwidth may be beneficial for a 16 core CPU triple channel and quad channel comes at the price of higher latency which is why Intel 2011 platform is slower for games than 1151 despite extra memory bandwidth, It is a big part of the reason why Ryzen falls behind 1151 and why Threadripper is slower for games than Ryzen.

Yup, GDDR5 would be a relative disaster for general computing on Ryzen, and HBM would be even worse, if they were the only memory available to the CPU. Faster DDR4 and at some point DDR5 will help with more cores and with their APUs. Their APUs would only really shine if they gave them some HBM, though.

[A high-end APU with an eight-core chiplet and a GPU+HBM chiplet would be very interesting for slimmer applications; done well, this would make a great solution for everything from NUC-sized devices up to midrange DTRs]
 
Yup, GDDR5 would be a relative disaster for general computing on Ryzen, and HBM would be even worse, if they were the only memory available to the CPU. Faster DDR4 and at some point DDR5 will help with more cores and with their APUs. Their APUs would only really shine if they gave them some HBM, though.

[A high-end APU with an eight-core chiplet and a GPU+HBM chiplet would be very interesting for slimmer applications; done well, this would make a great solution for everything from NUC-sized devices up to midrange DTRs]
Why would HBM be worse? Does it have bad latency? Whats typical HBM latency?
 
Note the incredibly slow clockspeeds involved.

Sure, but what is the actual latency to get some data? DDR memory has CAS values, so you are not getting your data on the next clock cycle anyways. The main things we care about from performance perspective are latency and bandwidth. HBM already wins in the bandwidth department. I'm not familiar with memory latency values. DDR4 typical latency is like 15ns or something like that?
 
Note the incredibly slow clockspeeds involved.

Does clock speed matter as much when were dealing with 2048bit wide bus? Ddr4 is only 64 bits wide per channel. I would think that lower clock speed astronomically wider bus would mean significantly higher parallel access to the ram making throughput ridiculously higher and isn't that what counts?
 
No idea what HBM latency is like unfortunately.

5755C performance in game with a dedicated GPU was improved by ~20-30% thanks to the 128MB eDRAM L4 cache.
https://www.overclock.net/forum/attachment.php?attachmentid=264354&stc=1&thumb=1&d=1554882467
This made it faster in latency sensitive games than newer CPU with much higher clocks like 7700K, at least while they are running slow RAM.

This is what 4600c18 is capable of with tweaked sub timings
https://www.overclock.net/forum/attachment.php?attachmentid=269494&d=1557797889
The RAM is getting down to ~35ns in that test with 1151 but even if you could hit that same RAM speed on current AM4 or 2011 platforms the latency would be much higher.
Also while that looks like it is faster than the eDRAM on the 5775C and it may well be in some ways eDRAM is bidirectional while DDR4 is not.

Does clock speed matter as much when were dealing with 2048bit wide bus? Ddr4 is only 64 bits wide per channel. I would think that lower clock speed astronomically wider bus would mean significantly higher parallel access to the ram making throughput ridiculously higher and isn't that what counts?
If that was the case then quad channel DDR4 would outperform dual channel but often it is dual channel that performs better despite the loss of bandwidth simply due to lower latency.
 
Last edited:
Does clock speed matter as much when were dealing with 2048bit wide bus? Ddr4 is only 64 bits wide per channel. I would think that lower clock speed astronomically wider bus would mean significantly higher parallel access to the ram making throughput ridiculously higher and isn't that what counts?

That depends. Single-core performance is linked to access latency, and then generally in terms of how long it takes to get stuff moving. Overall CPUs aren't starved for bandwidth. The rest of the system may be, and that's where enterprise-class hardware comes into play.
 
Back
Top