after 32 cores?

Epyon

[H]ard|Gawd
Joined
Oct 25, 2001
Messages
1,185
I work offshore and will be building a small 32 core box that I can run 24/7 for free but I want more. I'm greedy. What's after 32 cores? Rumors? More cores on a TR chip?
 
Zen 2 is rumored to scale out to 12c dies, which will make 48c TR/EPYCs possible. There's some debate on how that will occur, another 4c CCX or 6c CCXs. Beyond that, Zen 2 will scale out to 16c dies/64c CPUs (unknown if that's quad 4c CCX or dual 8c CCX). At this moment, 64c is only confirmed for EYPC/Rome and not TR. Quad channel ram is going to be a bottleneck on a 64c machine for most tasks that can utilize that many cores. IMHO, we should see 48c TR3 next year, but won't see 64c TR4 until the socket is changed for at least six channel ram, but given Zens architecture eight channel makes the most sense.

On the Intel front, LCC will likely scale out to the existing HCC layout. HCC will scale out to ECC and ECC should scale up to 42c. When is uncertain and it doesn't seem like Cascade Lake will make that happen. Likely it won't happen due to yields until Intel gets 10nm working. Ice Lake in late 2019, but probably more like 2020 seems like the first real chance for an architecture change. Until then, it's probably just a few more cores if 10nm works out.

Edit: my guess is 64c TR4 isn't until 2020/2021 on a new octa-channel socket. Or version 2.0 of TR4 which would really just be the current EYPC socket variant. 2020 if 7nm enhancements work out, 2021 if they stall.

Edit 2: At some point DDR5 will bail TR out, but that is also a new socket. It's just that with DDR5, TR could remain quad-channel.
 
Last edited:
my guess is 64c TR4 isn't until 2020/2021 on a new octa-channel socket

With DDR5 it might not need 8 channels. Although 64 cores is getting a little crazy.
 
With DDR5 it might not need 8 channels. Although 64 cores is getting a little crazy.
LOL. Looks like I was editing that while you were replying. Based on the DDR5 and Zen 3 projections, we probably won't see a TR variant until 2021, maybe 2022. Z3 EPYC in 2021 with TR within the year.
 
i have a rig that has 68 cores on 8 channle memory i can tell you the one thing i would wish it has more of would be better ipc. this is especially important for any software the lisences on a per thread/core basis.
 
i have a rig that has 68 cores on 8 channle memory i can tell you the one thing i would wish it has more of would be better ipc. this is especially important for any software the lisences on a per thread/core basis.

thats a lot to ask for just 215W.. 272Threads at 1.5ghz it's A LOT.. however without virtualization it doesn't have too much usage for me.. I think it would be great if they increase the TDP up to 300W in order to reach higher clocks..
 
I could imagine an X499 chipset that supports Octo-Channel memory. Essentially it will be backwards compatible with TR1 and TR2, but they will run in QC mode with two dims-per-channel when installed. The TR3 chips will support 8ch with 1DPC, but will fall back to 4ch 2DPC on older X399 boards.

My guess.
 
I could imagine an X499 chipset that supports Octo-Channel memory. Essentially it will be backwards compatible with TR1 and TR2, but they will run in QC mode with two dims-per-channel when installed. The TR3 chips will support 8ch with 1DPC, but will fall back to 4ch 2DPC on older X399 boards.

My guess.

That shouldn't be too difficult. The abysmal failure that was KabyLake X worked on that principle. Instal a real HEDT chip on the board and get quad channel, but instal a jumped up Kaby and get dual channel.
 
Zen 2 is rumored to scale out to 12c dies, which will make 48c TR/EPYCs possible. There's some debate on how that will occur, another 4c CCX or 6c CCXs. Beyond that, Zen 2 will scale out to 16c dies/64c CPUs (unknown if that's quad 4c CCX or dual 8c CCX). At this moment, 64c is only confirmed for EYPC/Rome and not TR. Quad channel ram is going to be a bottleneck on a 64c machine for most tasks that can utilize that many cores. IMHO, we should see 48c TR3 next year, but won't see 64c TR4 until the socket is changed for at least six channel ram, but given Zens architecture eight channel makes the most sense.

On the Intel front, LCC will likely scale out to the existing HCC layout. HCC will scale out to ECC and ECC should scale up to 42c. When is uncertain and it doesn't seem like Cascade Lake will make that happen. Likely it won't happen due to yields until Intel gets 10nm working. Ice Lake in late 2019, but probably more like 2020 seems like the first real chance for an architecture change. Until then, it's probably just a few more cores if 10nm works out.

Edit: my guess is 64c TR4 isn't until 2020/2021 on a new octa-channel socket. Or version 2.0 of TR4 which would really just be the current EYPC socket variant. 2020 if 7nm enhancements work out, 2021 if they stall.

Edit 2: At some point DDR5 will bail TR out, but that is also a new socket. It's just that with DDR5, TR could remain quad-channel.

64c is coming to EPYC in 2019 with sampling underway or shortly by now.
I wonder though, how do they get 64c? It has to be 8 dies of 2x4 core ccx each or 6 dies x 12 cores with weird uneven disabled cores, 1 or 2 depending. They must have less memory channels, or a lot of that socket is unused already. If they planned ahead, I guess they have enough memory channels. They are not changing socket till 2020, so this might be the last hurrah and planned for.

I work offshore and will be building a small 32 core box that I can run 24/7 for free but I want more. I'm greedy. What's after 32 cores? Rumors? More cores on a TR chip?
Wait for 64 Core 7nm Epyc 2 it's out in '19
 
64c is coming to EPYC in 2019 with sampling underway or shortly by now.
I wonder though, how do they get 64c? It has to be 8 dies of 2x4 core ccx each or 6 dies x 12 cores with weird uneven disabled cores, 1 or 2 depending. They must have less memory channels, or a lot of that socket is unused already. If they planned ahead, I guess they have enough memory channels. They are not changing socket till 2020, so this might be the last hurrah and planned for.


Wait for 64 Core 7nm Epyc 2 it's out in '19

it'll be on the 7nm process so maybe they'll have 8x8 x 4. as far as epyc goes it already has 8 channel DDR4 vs 4 channel ddr4 on TR4.
 
it'll be on the 7nm process so maybe they'll have 8x8 x 4. as far as epyc goes it already has 8 channel DDR4 vs 4 channel ddr4 on TR4.
Don't you mean 8 cores per CCX, 8x2 x4 dies?, 8 per CCX would be amazing for latency related stuff and gaming boosts. Well, to be honest, it was never that bad even inter-CCX at higher ram speeds. Inter-die was/is the only main limitation.
I see WTFshit is thinking amd will make two different dies, 12 and 16 core, that's so un-AMD to do this, bet it's just 16 core cut to 12.
 
thats a lot to ask for just 215W.. 272Threads at 1.5ghz it's A LOT.. however without virtualization it doesn't have too much usage for me.. I think it would be great if they increase the TDP up to 300W in order to reach higher clocks..

The new chips have virtualization :p and yea I would like them to up the tdp to like 500w as the hardware is all specilized anywhay
 
I work offshore and will be building a small 32 core box that I can run 24/7 for free but I want more. I'm greedy. What's after 32 cores? Rumors? More cores on a TR chip?

Well, once somebody comes out with a CPU that has 2^32 -1 threads I will have to update the code for my prime number generator. And that update would only be a change from 32-bit integers to 64-bit integers for a few of the variables.
Super-awesome code that uses 0 locks except to wait to spit out the final results after all the worker threads finish at about the same time.
 
Last edited:
The new chips have virtualization :p and yea I would like them to up the tdp to like 500w as the hardware is all specilized anywhay

this is interesting indeed. I may put aim on some of the higher clocked 7290F or are strictly needed 72x5 family? i have some home work to do =) thanks..
 
this is interesting indeed. I may put aim on some of the higher clocked 7290F or are strictly needed 72x5 family? i have some home work to do =) thanks..

From what I have found the 90s are not worth the extra money. I have abunch of es 7210s that peg all cores at 1.5ghz under load. I do want to get me a 72x5 chip strictly for the VM capabilities but I can't find any cheap one
 
If you're gaming you are good with 8 cores for the foreseeable future. Don't waste your money unless you NEED the extra cores.
 
I mean you gotta be able to use your cores whether its 4 or 64 cores. Do you have a usage scenario (s)?

I couldn't imagine anything smaller than Enterprise running big VM boxes needing even 32 cores much less 64.

I may get a 2990x if they have good reviews and have improved in some desirable areas since the 1950x I owned and sold.
 
I mean you gotta be able to use your cores whether its 4 or 64 cores. Do you have a usage scenario (s)?

I couldn't imagine anything smaller than Enterprise running big VM boxes needing even 32 cores much less 64.

I may get a 2990x if they have good reviews and have improved in some desirable areas since the 1950x I owned and sold.


Mmmmmiiiiinnnnniiiinnnnnggggggg
 
I use it for keyshot 3d rendering. It only uses cores. I do not game or mine.
 
You can always buy 2 machines? It depends a little on the price of the server parts.
Is rendering software these days flexible enough to run 1 project on 2 physical machines?
 
Pieter. I am building a small custom box that I can fit in my offshore bag. Or something that can fit in a very small back pack. Thin and light is key. Also you do not want multiple computers. I don't want to pay rendering node prices. They can keep that bull shit. I do this as a hobby not a job. I just love the fact I can run 32 cores 24/7 for 28 days at a time free.
 
Pieter. I am building a small custom box that I can fit in my offshore bag. Or something that can fit in a very small back pack. Thin and light is key. Also you do not want multiple computers. I don't want to pay rendering node prices. They can keep that bull shit. I do this as a hobby not a job. I just love the fact I can run 32 cores 24/7 for 28 days at a time free.
 
Mmmmmiiiiinnnnniiiinnnnnggggggg

Why? You get an order of magnitude higher performance via GPUs.

That's the big problem here: GPUs are simply much more efficient at parallel workloads, and if AMD and Intel aren't careful, they are going to make NVIDIA king. GPUs are basically massively-parallel CPUs at this point anyways.
 
Why? You get an order of magnitude higher performance via GPUs.

That's the big problem here: GPUs are simply much more efficient at parallel workloads, and if AMD and Intel aren't careful, they are going to make NVIDIA king. GPUs are basically massively-parallel CPUs at this point anyways.

A Vega gpu does 2000h/s on monero a Intel phi CPU does 3000h/s on monero. 3000>2000
 
That part maybe true but how much Watt per system :)

xeon phi it's about 215W up to 260W depending on which exact model, VEGA64 is AT LEAST 295W, so if someone still undervolt and underclock to keep power controlled Xeon Phi still will win by a large margin in every aspect, performance, power and efficiency in such case scenario (mining monero). and 1000h/s mean about 30 extra dollars monthly without factor efficiency from 215W to 295W 24/7 the entire month, of course everything will be the opposite depending on much the phi and Vega64 cost, phi tend to be more expensive at about 1200 -1500$ for the 7120Model while you can buy right now a Vega64 for less than 600$
 
Last edited:
With DDR5 it might not need 8 channels. Although 64 cores is getting a little crazy.
it got crazy with 8 cores. 64 is just nuts. We're going the wrong direction with this. 4 cores and 8 threads with higher clocks is still more useful most of the time. BUT if we can go past 5Ghz add all the cores you want. We just need better per core performance making its way out.
 
If you're gaming you are good with 8 cores for the foreseeable future. Don't waste your money unless you NEED the extra cores.
even in gaming most games are throttled by the GPU far before reaching 8 cores.
 
it got crazy with 8 cores. 64 is just nuts. We're going the wrong direction with this. 4 cores and 8 threads with higher clocks is still more useful most of the time.

It's easier to add more cores (well if the heat/power can be kept in check) than to improve the performance per core.
 
it got crazy with 8 cores. 64 is just nuts. We're going the wrong direction with this. 4 cores and 8 threads with higher clocks is still more useful most of the time. BUT if we can go past 5Ghz add all the cores you want. We just need better per core performance making its way out.

no we need more programs/games to better use the threads available to them, clock speed means less and less once an application can scale across multiple cores, we need to stop supporting the half ass lazy coding that is still going on because "oh no people in Brazil or Russia are still playing games on a Pentium 3's so lets screw over everyone to make it fair for them".
 
no we need more programs/games to better use the threads available to them, clock speed means less and less once an application can scale across multiple cores, we need to stop supporting the half ass lazy coding that is still going on because "oh no people in Brazil or Russia are still playing games on a Pentium 3's so lets screw over everyone to make it fair for them".
Do you even code? there are tons of things that just don't thread very well. Everything that DOES thread well belongs on a GPU It has nothing to do with lazy coding. I've run into some things running SLOWER by adding threading. No matter what you have a main thread which all the other threads report back to. 90% of the time I'd gladly take 4 cores running @ 5ghz over 6 running at under 4Ghz. Adding cores leads to dropping clock speed and thats simply doing things wrong for the sake of some oddball benchmarks that don't indicate real world usage.
 
Do you even code? there are tons of things that just don't thread very well. Everything that DOES thread well belongs on a GPU It has nothing to do with lazy coding. I've run into some things running SLOWER by adding threading. No matter what you have a main thread which all the other threads report back to. 90% of the time I'd gladly take 4 cores running @ 5ghz over 6 running at under 4Ghz. Adding cores leads to dropping clock speed and thats simply doing things wrong for the sake of some oddball benchmarks that don't indicate real world usage.

Pretty much this.

Case in point: There was I project I was working on that got ported to a "modern" (post-80's) processor. The software continued to run horribly. A massive effort was undertaken to make the program more threaded, which ended up costing the taxpayers (you) mid-7 figures.

Performance dropped 50%. We found the software kept grinding to a halt due to threads constantly waiting on eachother.

After that debacle, an actual analysis of the code was done. One minor code change to the original (pre-threaded) code base resulted in a 500% increase in performance.

Threading is not a magical salve for performance increases; the majority of workloads do not scale to an increased number of threads. All that adding more and more cores to the CPU is going to do is clamp down on clock speeds and maximum OC, resulting in lower performance and higher cost to consumers.
 
Back
Top