3950X about to Redefine/Restructure the meaning of HEDT

[nicalandia]

The 3900X is already in HEDT performance territory, but extrapolating the SMT performance of the 3950X using the 3900X, this beast will outperform the 9980XE and get this, the 3950X is a main stream CPU, I simply can't fathom how much performance will a 32 Core Zen2 TR will bring, can we extrapolate such numbers using Ryzen 9 CPU now that the Central IO chiplet is equalizing the communications?

 

[Markfw]

The 3900X is already in HEDT performance territory, but extrapolating the SMT performance of the 3950X using the 3900X, this beast will outperform the 9980XE and get this, the 3950X is a main stream CPU, I simply can't fathom how much performance will a 32 Core Zen2 TR will bring, can we extrapolate such numbers using Ryzen 9 CPU now that the Central IO chiplet is equalizing the communications?

I highly anticipate more than a 32 core 3000 series TR. I hope for 64.. And they will not suffer the same problems ar 2990wx with memory access due to the IO chip design.

I can't wait. I have EIGHT X399 motherboards to repopulate.....

 

[Thunder 57]

I highly anticipate more than a 32 core 3000 series TR. I hope for 64.. And they will not suffer the same problems ar 2990wx with memory access due to the IO chip design.

I can't wait. I have EIGHT X399 motherboards to repopulate.....

I wonder what your electricity usage is on computing alone. You probably don't even need to run the heat during the Winter .

 

[nicalandia]

I highly anticipate more than a 32 core 3000 series TR..

Holy Crap..! So much performance, Intel would need at least a 48-Core Cascade Lake-AP to even think about competing with a 32 core TR3

 

[zir_blazer]

I highly anticipate more than a 32 core 3000 series TR. I hope for 64.. And they will not suffer the same problems ar 2990wx with memory access due to the IO chip design.

I can't wait. I have EIGHT X399 motherboards to repopulate.....

Given the fact that AMD bothered to release the 2990WX on the first place, which was a shoehorned design with poor Software support (Asymetric NUMA with CPUs without local RAM that was extremely dependent on OS CPU Scheduler not being stupid to perform better than the 2950X instead of worse), I see no reason for AMD to not release a halo 64C TR considering that Rome fixes all the previous multi Zeppelin die shortcomings and does everything right, as AMD could showcase something that scales properly.

Money being no issue (Which seems to be your case), I would prefer a 8 Channel - 128 PCIe EPYC monster. That depends on whenever AMD releases high clocked Workstation models instead of the current clock speed disparity than there exist between TR and EPYC.

 

[nicalandia]

Money being no issue (Which seems to be your case), I would prefer a 8 Channel - 128 PCIe EPYC monster. That depends on whenever AMD releases high clocked Workstation models instead of the current clock speed disparity than there exist between TR and EPYC.

I remember way back in the day there were a few custom/hacked bios on the G34 motherboards that allowed for OC the motherboards, could this still be done this day and age?

 

[zir_blazer]

I remember way back in the day there were a few custom/hacked bios on the G34 motherboards that allowed for OC the motherboards, could this still be done this day and age?

There were some recent EPYC overclocking results, but AMD disabled it via newer Microcode.

EPYC Motherboards are not expected to be overclocked, so even if theorically possible, they don't have the massive power delivery required to do something intense with those monsters.

 

[RichUK]

You seem really excited.

 

[Markfw]

There were some recent EPYC overclocking results, but AMD disabled it via newer Microcode.

EPYC Motherboards are not expected to be overclocked, so even if theorically possible, they don't have the massive power delivery required to do something intense with those monsters.

Thats the thing, I will take quad channel memory and a higher core clock (threadripper) over EPYC's 8 channel memory and slower speeds. and no I don't need 128 cores/256 threads in one box (cost prohibitive.)

 

[Atari2600]

Thats the thing, I will take quad channel memory and a higher core clock (threadripper) over EPYC's 8 channel memory and slower speeds. and no I don't need 128 cores/256 threads in one box (cost prohibitive.)

How sensitive are the distributed workloads to memory bandwidth Mark?

I'd imagine due to their nature, the threads are both highly independent and of low memory overhead?

If TR3 is limited to 32 cores, you might be better off with 2x AM4 3950X in both power and cost than doing drop in replacements to TR3.


{or is it really all about GPU accelerators?}

 

[LightningZ71]

It all depends on the individual Distributed workload you're talking about. Many are doing VERY well on GPUs. Others are more oriented to general purpose processors in the CPUs themselves. A lot of the different DC projects can fit their normal working set inside of a few MB of cache, and just fly on modern large cache processors. So, you'd have to be more specific about which workload that you're talking about.

 

[Atari2600]

I haven't done any of the public distributed stuff in over a decade - I've no idea of the current lie of the land!

 

[beginner99]

3950X is a main stream CPU

I disagree simply by looking at it's price and the price of x570 mobos we are clearly in HEDT territory.

 

[amrnuke]

The idea of a 64 core 128 thread TR with 4 Radeon VII running some DC protein folding stuff makes me vroom.

 

[Markfw]

How sensitive are the distributed workloads to memory bandwidth Mark?

I'd imagine due to their nature, the threads are both highly independent and of low memory overhead?

If TR3 is limited to 32 cores, you might be better off with 2x AM4 3950X in both power and cost than doing drop in replacements to TR3.


{or is it really all about GPU accelerators?}

They don't seem to need memory bandwidth, and its GPU or CPU or both depending on the project. I really only do 3, F@H, WCG and Rosetta,

 

[nicalandia]

I disagree simply by looking at it's price and the price of x570 mobos we are clearly in HEDT territory.

Not really, there has been reports of B350 boards working just fine with stock 3900X, I believe the X370 will handle the 3950X just fine, that fact alone(mainstream upgrade from 1600 to 3950X on same board) makes this chip a "Mainstream" CPU

X370 Tachi working with X390:

https://www.reddit.com/r/Amd/comments/ca60ys

$70 B350 board with X390: https://www.techspot.com/review/1872-ryzen-9-on-older-motherboards/

 

[amrnuke]

Not really, there has been reports of B350 boards working just fine with stock 3900X, I believe the X370 will handle the 3950X just fine, that fact alone(mainstream upgrade from 1600 to 3950X on same board) makes this chip a "Mainstream" CPU

X370 Tachi working with X390:

https://www.reddit.com/r/Amd/comments/ca60ys

$70 B350 board with X390: https://www.techspot.com/review/1872-ryzen-9-on-older-motherboards/

What is an X390? Isn't that a Lenovo laptop line?

 

[nicalandia]

What is an X390? Isn't that a Lenovo laptop line?

sorry I meant 3900X

edit.
Poll added.

 

[moinmoin]

I'd say an R9 3950X is as much mainstream as an RTX 2800 is.

 

[StinkyPinky]

Yeah the upcoming thread rippers will be amazing. The new design should fix some of the issues the previous gen ones had, so the sheer amount of processing power these things should have is amazing. I would love to get one but even my upcoming 3900x is overkill for my needs. May be hard to justify a 32+ core beast to browse websites and play games to the wife.....

 

[Markfw]

I can use all the cores I can get ! Give me 64 cores on TR3 ! Damn the price, full speed ahead !

 

[dnavas]

Yeah the upcoming thread rippers will be amazing..

I'm quite interested in the second "up" in "up up". Seems clear that we're not going to get much in the way of frequency improvements after we're done doubling cores (although I await the inevitable 64 5ghz all-core passively-cooled free-sex-and-beer overclock baloney we get every round), so I'm wondering what else Lisa has in store. [:cross-fingers:] I'm also wondering if they'll choose to populate 32 core TRs with 8 half-active chips (highest clocking cores) or 4 full chips [or 4 full chips plus a little extra surprise in the other four spots]. Given some bumps in the road wrt bios and node maturity, it'll be interesting to see where we are come October (if rumors turn out to be accurate).

 

[Markfw]

I'm quite interested in the second "up" in "up up". Seems clear that we're not going to get much in the way of frequency improvements after we're done doubling cores (although I await the inevitable 64 5ghz all-core passively-cooled free-sex-and-beer overclock baloney we get every round), so I'm wondering what else Lisa has in store. [:cross-fingers:] I'm also wondering if they'll choose to populate 32 core TRs with 8 half-active chips (highest clocking cores) or 4 full chips [or 4 full chips plus a little extra surprise in the other four spots]. Given some bumps in the road wrt bios and node maturity, it'll be interesting to see where we are come October (if rumors turn out to be accurate).

She has already said we were getting more cores, so at least 40-48 for round one, and maybe as much as 64 in round 2. The round one and round 2 guesses are based on what they did with AM4, 12 cores now, 16 cores in Sept.

 

[gdansk]

I would be OK with a 16 core Zen 2 chip on TR4 socket. PCIe lanes would make it worthwhile for me. On the other hand, I might be in a very small minority.

 

[dnavas]

She has already said we were getting more cores, so at least 40-48 for round one, and maybe as much as 64 in round 2.

Ah, so your suspicion is that up-up refers to a split release schedule. That's interesting, but Ryzen 9 was released that way too (as you point out), so how does that square with "up" vs "up up"? I've heard goofy speculation about dual sockets and heterogenous processing capabilities and am far more interested in the latter than the former, though if I had to guess, I'd be far likelier to suspect it's something a lot less than that (closer to your split release schedule idea). Maybe a liquid cooler that doesn't fall apart in 9 months and could actually handle high frequency 64 core heat? If not AMD, *someone* should be working on that.... Homebrewed solutions are fun for some, but that won't sell boxes to businesses.