AMD Ryzen 2nd Gen Threadripper 32 Cores Confirmed

caltech31 said:
Looks like I will be replacing our remote sensing computer in the lab!

Depending on price, I may even put it in my personal machine (although the 7820x at 5ghz is a beast in its own right)
Click to expand...

What cooling are you using to hit 5ghz?
 
Chimpee said:
The CPU competition sure looks pretty bright, looking forward to see if Zen 3 will have a mainstream 12-16 core parts!
Click to expand...

we'll probably see 12 core on am4 even before zen 3, wouldn't be surprised if they try to test it out on zen 2 before the full socket change for ddr5/pcie4 with zen 3. it would at least give them a reason to bring back the 3800x naming scheme by making those two processors be the 12 core varients then the rest being the 10/8/6 core variants.
 
The important question to ask is if AMD is going to make an updated 28core i9 survival kit?
131554_De-6pEnX4AITfUD.jpg
 
Last edited:
Sold.

Will definitely get one of these. Shame about Optane.

They'll sell a ton for workstations
 
With that many cores it's going to need WAY more RAM slots. I mean, I'm interested, but solely from a VM sandbox environment perspective. 128 GB just isn't enough. They gotta pump up those numbers! Those are rookie numbers! Give me at least a 512 GB limit. In reality, 1 TB would be nice.
 
Microsoft is going to have to change their pricing model for business class operating systems. Currently it’s rated on CPU count and not core count.
 
Riccochet said:
With that many cores it's going to need WAY more RAM slots.
Click to expand...
For that, AMD has Epyc. Also Epyc supports LR-DIMMs for more memory, so 1 TB is no problem (e.g. Supermicro H11DSi with 16*64 GB is readily available).
ThreadRipper with more than 64 GB however is impractical, and more than 128 GB is not possible at this time.
 
How is twice as many cores going to do with only 4 memory channels? Rendering isn't bothered by such concerns but I imagine plenty of other workloads are.
 
chithanh said:
ThreadRipper with more than 64 GB however is impractical, and more than 128 GB is not possible at this time.
Click to expand...

More than 64 is impractical? Did you read my post? I couldn't sandbox one of my VM's with 64 GB of RAM, let alone simulate a workload on multiple.

It's an HEDT chip. I suppose if all you are doing is crunching numbers, great. But 128 GB of RAM is going to severely limit the use of all those cores.
 
Riccochet said:
More than 64 is impractical? Did you read my post? I couldn't sandbox one of my VM's with 64 GB of RAM, let alone simulate a workload on multiple.

It's an HEDT chip. I suppose if all you are doing is crunching numbers, great. But 128 GB of RAM is going to severely limit the use of all those cores.
Click to expand...


i'd say wait and see, who knows what they're going to be doing to integrate the 2 extra dies and whether or not they're able to increase the memory limitation. but i do agree that 128GB could limit uses of the processor but at the same time they need to make sure that the lower core count epyc chips have a place in the market.
 
Looking forward to how well current TR4 boards will handle this. It would be a shame to have to upgrade everything again... :D
 
R_Type said:
How is twice as many cores going to do with only 4 memory channels? Rendering isn't bothered by such concerns but I imagine plenty of other workloads are.
Click to expand...

what i heard was one die is going to have to access ram over the infinity fabric.
 
R_Type said:
How is twice as many cores going to do with only 4 memory channels? Rendering isn't bothered by such concerns but I imagine plenty of other workloads are.
Click to expand...

Technically, a CPU needs memory access, but it doesn't necessarily have to be a local memory. CPU's access remote NUMA nodes today for various reasons. Unfortunately, it isn't optimal. In fact, the penalty is quite massive. As you said, for rendering this is fine but for many applications this isn't ideal. Applications that are designed to do this are designed to work this way, but applications that are not designed for it will suffer. However, I don't think this is going to be an issue as its very likely AMD will utilize the Infinity Fabric for this rather than having the extra CPU cores access memory as a remote NUMA node.

sirmonkey1985 said:
i'd say wait and see, who knows what they're going to be doing to integrate the 2 extra dies and whether or not they're able to increase the memory limitation. but i do agree that 128GB could limit uses of the processor but at the same time they need to make sure that the lower core count epyc chips have a place in the market.
Click to expand...

Definite limitation? Are you talking about the physical 128GB memory limit or the limitations that seem apparent by only having 4x memory channels available?

trick0502 said:
what i heard was one die is going to have to access ram over the infinity fabric.
Click to expand...

As I understand it, the Infinity Fabric's design means that the amount of hops required to access the memory will be minimized and that a 32c/64t Threadripper CPU shouldn't suffer too badly. Although, I'm not sure what the latency penalty for doing this is, but its probably not as bad as the bandwidth and latency hits suffered by accessing a remote NUMA node.
 
Last edited:
NKD said:
I just like slide for 7nm GPU! 2x efficiency, 2x density, 35% performance. Seriously they can bring vega 64 performance at half the power? If those numbers are right. Looks like Lisa kicked raja out and went to the engineers said, give me 7nm working GPU in 2018 or bust! Probably the only way for them to compete with Nvidia until next gen! I say Lisa is executing the shit out of 7nm!
Click to expand...

Lisa Su is helping to keep AMD in the game while trying to get the company to remember what a few victories here and there feel like (not to mention job security, pay raises, promotions, etc...) - When (if) Lisa Su can drop a surprise launch sooner than later on nVidia of the next gen cards, I see nVidia being able to drop much larger bombs back than Intel has been able to in the CPU market (not that Intel has proven to be a slouch by any means) - But, that's exactly what we need, someone to drive nVidia, AMD to have solid leadership and provide some exciting products for team Red.... and then Intel enters the room with their Graphics Cards ;)
 
This is going to be a great workstation CPU. I'll probably replace my 2700X with one. All that we need to know now is how much heat it puts out and what the final speeds will be. Can't wait to see the review, Kyle.
 
Montu said:
This is going to be a great workstation CPU. I'll probably replace my 2700X with one. All that we need to know now is how much heat it puts out and what the final speeds will be. Can't wait to see the review, Kyle.
Click to expand...

My next CPU is likely to be a Threadripper of some sort. Unless the clocks are considerably higher than today's 32c/64t Epyc CPUs, I doubt I'll go with a 32 core CPU. I will probably opt for whatever replaces a Threadripper 1950X or a current 1950X depending on the performance of the former.
 
Riccochet said:
With that many cores it's going to need WAY more RAM slots. I mean, I'm interested, but solely from a VM sandbox environment perspective. 128 GB just isn't enough. They gotta pump up those numbers! Those are rookie numbers! Give me at least a 512 GB limit. In reality, 1 TB would be nice.
Click to expand...
People were getting 256gb of unbuffered ecc to run on an ITX X99 board. Since then I just assumed something similar would be possible on threadripper? Can it not unofficially support larger dimms?
 
Neapolitan6th said:
People were getting 256gb of unbuffered ecc to run on an ITX X99 board. Since then I just assumed something similar would be possible on threadripper? Can it not unofficially support larger dimms?
Click to expand...

AMD could absolutely increase the supported memory module density and double the amount of supported RAM in 4 channels. Whether or not they will do so officially is another matter entirely.
 
Neapolitan6th said:
The second 64 threads helps me in civilization, I would buy one. Doubt that for a long while though
Click to expand...

i have an 8 core threadripper using a 1TB M2 drive for storage, and while its smoking fast for civ.. sadly.. most cores are bored with the game. someday they will do a true 64bit multi threaded engine.
 
Okay, this is a terrible way to measure performance, it's not perfectly linear and is not the ultimate benchmark, but it is interesting anyway:

Just a quick math,with the assumption that you can multiply the frequency (in ghz) by the number of cores, if you multiply Intel's 28-core monster at 4.6ghz , the number of "combined cycles" is the same as 32-core Threadripper multiplied by 4.0ghz.

And giving the monolithic nature of Intel, the heat is much more complicated to deal with, so I doubt anyone can run it at 5ghz without phase change cooling.

Not a proper measure, but interesting though
 
trick0502 said:
what i heard was one die is going to have to access ram over the infinity fabric.
Click to expand...

cageymaru said:
Infinity Fabric I bet.
Click to expand...

Dan_D said:
Technically, a CPU needs memory access, but it doesn't necessarily have to be a local memory. CPU's access remote NUMA nodes today for various reasons. Unfortunately, it isn't optimal. In fact, the penalty is quite massive. As you said, for rendering this is fine but for many applications this isn't ideal. Applications that are designed to do this are designed to work this way, but applications that are not designed for it will suffer. However, I don't think this is going to be an issue as its very likely AMD will utilize the Infinity Fabric for this rather than having the extra CPU cores access memory as a remote NUMA node.



Definite limitation? Are you talking about the physical 128GB memory limit or the limitations that seem apparent by only having 4x memory channels available?



As I understand it, the Infinity Fabric's design means that the amount of hops required to access the memory will be minimized and that a 32c/64t Threadripper CPU shouldn't suffer too badly. Although, I'm not sure what the latency penalty for doing this is, but its probably not as bad as the bandwidth and latency hits suffered by accessing a remote NUMA node.
Click to expand...

No no guys I mean in terms of bandwidth not latency.

As far as things stand the next threadripper will double the cores but they will still be fed via just 4 memory channels giving each die a whopping 1 memory channel.

Unless on the current socket+chipset these things aren't hard and fast truths and are entirely configurable. That would mean if you put all 8 sticks in you'd have 2 channels per die.

Put it this way unless someone knows for certain what's happening with tr2 and memory channels then going from tr1 to tr2 would, in bandwidth per die terms, be like taking say a 2700x and halving the ddr4 frequency.

Again if anyone knows what the score here really is please chime in.
 
Do we know yet what the clock speeds are going to be on TR2 compared to TR1?

Not just the 32, but also the 16 core parts which should be a direct comparison to the existing TR.
 
You must log in or register to reply here.
Back
Top