AMD’s 64-Core Threadripper 3990X, only $3990! Coming February 7th

[Guest]

Just now, GOTSpectrum said:

Folding? BOINC?

 

get on it man!

NO. It's $3,990. My budget for CPUs is $3,880.

[Letgomyleghoe]

On 1/7/2020 at 12:42 AM, porina said:

I'm gonna need a $/core-GHz chart or something at this rate... this is exactly the product AMD needs to keep turning the thumbscrews on Intel. Few (but not none) will buy it, but solidifying their leadership position will help grow their lower products also.

 

I'd still throw caution in that, depending on the workload, a monolithic CPU can work better than multiple fragmented ones. 

i was looking at buying trx40 board just for the purpose of having a crazy upgrade path, when 64 cores in a system becomes common (which it *hopefully* wont for a while) i'll still be able to have a top of the line system without switching platforms.

[GOTSpectrum]

31 minutes ago, Den-Fi said:

NO. It's $3,990. My budget for CPUs is $3,880.

You'll have to settle for a 2nd gen then ....

 

F

 

 

[Guest]

11 minutes ago, GOTSpectrum said:

You'll have to settle for a 2nd gen then ....

 

F

 

 

I have a 3970X tho

[xtroria]

Why the hell are people comparing this to mac pro?

 

people who are buying mac pros are mostly digital artists and their workflow are generally very mac specific and / or they don’t want to deal with the hassle of working with custom rig + windows + drivers

 

mac pros still have the accelerator card as well as metal.

 

this cpu will be mostly used by people who are more into big data computation or work where the mac software doesnt have any benefit over windows

[leadeater]

52 minutes ago, GOTSpectrum said:

Folding? BOINC?

 

get on it man!

Speaking of, wonder what the most power efficient highest PPD distribution of hardware would be for 2kw (that's my peak solar output). For both F@H and BOINC, like the overall best mix.

[GOTSpectrum]

1 minute ago, leadeater said:

Speaking of, wonder what the most power efficient highest PPD distribution of hardware would be for 2kw (that's my peak solar output). For both F@H and BOINC, like the overall best mix.

RTX 2060 where best price to performance I believe. @Gorgon did some serious research into this so better to ask him

[Guest]

1 hour ago, GOTSpectrum said:

RTX 2060 where best price to performance I believe. @Gorgon did some serious research into this so better to ask him

Yes, that's how he cheated and made me mildly nervous in the beginning of folding month.

[Gorgon]

31 minutes ago, Den-Fi said:

Yes, that's how he cheated and made me mildly nervous in the beginning of folding month.

Lol. The RTX 2070 Supers are about the best bang for the buck. The 2080 Supers do yield higher throughput but a lot of WUs don’t use all the CUDA cores so they tend to underperform but that will likely change as things progress.

 

The 2070 Supers run at about 190W with a slight overclock. They yield about 1.5MPPD and the EVGA hybrid models are reasonably priced.

 

ill probably pick up a 3950x for my x570 board in the spring

[GOTSpectrum]

7 hours ago, Gorgon said:

Lol. The RTX 2070 Supers are about the best bang for the buck. The 2080 Supers do yield higher throughput but a lot of WUs don’t use all the CUDA cores so they tend to underperform but that will likely change as things progress.

 

The 2070 Supers run at about 190W with a slight overclock. They yield about 1.5MPPD and the EVGA hybrid models are reasonably priced.

 

ill probably pick up a 3950x for my x570 board in the spring

@leadeater there you go, the boffin has spoken

[leadeater]

29 minutes ago, GOTSpectrum said:

@leadeater there you go, the boffin has spoken

Yea but that only covered things that can run on GPU, I'd be looking for the best mix with the CPU also fully utilized which would eat in to the 2kw power budget. Systems that would be highly useful for both folding month and BOINC pent.

 

Since you can power cap the CPU there must be something that is the best at like 150W or 100W, also would dual CPU give the best total combined or single CPU. I think 2kw is a fair amount so would be more than one system but then again that 8 GPU server I was using could do that itself... not that I'm looking to do that lol.

[porina]

1 hour ago, leadeater said:

Yea but that only covered things that can run on GPU, I'd be looking for the best mix with the CPU also fully utilized which would eat in to the 2kw power budget. Systems that would be highly useful for both folding month and BOINC pent.

 

Since you can power cap the CPU there must be something that is the best at like 150W or 100W, also would dual CPU give the best total combined or single CPU. I think 2kw is a fair amount so would be more than one system but then again that 8 GPU server I was using could do that itself... not that I'm looking to do that lol.

A difficulty is that performance can and will vary with the project choice/mix. However if we look at it in very generic terms, if the limit is only CPU power budget, this usually goes towards more cores at lower clock, than fewer cores at higher clock. Also, Zen 2 seems to be leading the way in performance per watt, so the highest core count Zen 2 solution would probably be best in any power budget (don't know how much you'd want to dedicate to GPU side).

[MageTank]

On 1/7/2020 at 3:19 PM, Deli said:

With 128GB RAM, even with Zen 2, it won't be easy to run them at 3600MHz.

This is a challenge I'll gladly undertake. Sadly I only have a 3970X on hand and a Gigabyte TRX40 board that I have never used in my life. Do I get bonus points for pushing 3800mhz?

[leadeater]

16 minutes ago, porina said:

A difficulty is that performance can and will vary with the project choice/mix. However if we look at it in very generic terms, if the limit is only CPU power budget, this usually goes towards more cores at lower clock, than fewer cores at higher clock. Also, Zen 2 seems to be leading the way in performance per watt, so the highest core count Zen 2 solution would probably be best in any power budget (don't know how much you'd want to dedicate to GPU side).

Yea I just assume at some point you'll hit too low power limit for the number of cores e.g. 80W for 64 cores, don't think that will go too well. Not like I'm getting a 3990X any time soon to find out lol

[porina]

16 minutes ago, leadeater said:

Yea I just assume at some point you'll hit too low power limit for the number of cores e.g. 80W for 64 cores, don't think that will go too well. Not like I'm getting a 3990X any time soon to find out lol

As two random data points, Anandtech's look at the earlier models suggested as low as using 3W/core under load.

https://www.anandtech.com/show/15044/the-amd-ryzen-threadripper-3960x-and-3970x-review-24-and-32-cores-on-7nm/2

 

Also, the recently announced mobile CPU is 8 cores at 15W TDP. Now this is more APU-like so wont suffer the package overhead the IOD brings.

 

Still... we're looking at low single digit W/core possibility here. Spending your 100W or 150W budget on a 32-core should still give great performance at great efficiency. 64 cores might be pushing it a bit... but look at all the money I saved you, you "only" need the 32 core  (or maybe the 48-core when they decide to fill that very obvious hole).

[Kilrah]

I really want to see long-term benchmarks (like 1h tasks) and power consumption figures...

 

Since all 3 parts have the same TDP one can imagine the lower core count parts would just boost higher and give similar performance to the higher core count ones that have to throttle down for total power reasons. On a cinebench run like what they used to show a 2x increase of power that'll happen so fast that the power supply/cooling solution don't get time to start saturating.

 

That would mean that the high core count models would only be a real advantage in loads that result in short bursts and the few situations where you really want the number of cores like if you're running 15 quad core VMs...

[porina]

Just now, Kilrah said:

Since all 3 parts have the same TDP one can imagine the lower core count parts would just boost higher and give similar performance to the higher core count ones that have to throttle down for total power reasons.

It will reduce the gap that might be expected if you only look at the core count, but generally speaking more cores at lower clock are more efficient than fewer cores at higher clock, for a given power budget. More performance per watt at higher core counts. This also depends on the workload having good scaling to more cores.

[Kilrah]

Yep true, didn't consider that aspect.

 

Will also be interesting how core usage/frequency will work with that many cores.

The traditional "1-core boost" makes absolutely no sense when you've got 64 since it'd never happen, so I wonder up to what average load you'll still be able to have one core running full boost for a single threaded task, and how OSes will be able to deal with that.

 

[leadeater]

24 minutes ago, porina said:

64 cores might be pushing it a bit... but look at all the money I saved you, you "only" need the 32 core  (or maybe the 48-core when they decide to fill that very obvious hole).

lol thanks 

 

Also I could have just saved myself some time and just looked at this 

Quote

The 7452 has a TDP of 155 W with a base frequency of 2.2 GHz and a boost frequency of up to 3.35 GHz

https://en.wikichip.org/wiki/amd/epyc/7452

 

Based on that and parts around it should be possible to power cap 3970X to 150W and manually tune voltages to get ~3.0GHz+ on all cores, performance wise that should be excellent.

[porina]

11 minutes ago, Kilrah said:

The traditional "1-core boost" makes absolutely no sense when you've got 64 since it'd never happen

It hardly happens even now even on say 6 core systems. There's so much background activity it is hard for a single thread to be the only thing running at a time. On intel side, I think they don't publish it any more but there are places where you can look up the turbo tables for however many cores active. On AMD side this isn't so easy as they make much more use of power limiters, so clock is all over the place depending on the load.

 

1 minute ago, leadeater said:

Based on that and parts around it should be possible to power cap 3970X to 150W and manually tune voltages to get ~3.0GHz+ on all cores, performance wise that should be excellent.

Just do the power cap, don't touch voltages or clocks manually. The boost/limiting going on is far in advanced of old methods, and you're more likely to get the most out of more workloads than with old school tinkering. Have you played with any Zen 2 yet? The ECO mode is very interesting for crunching. This sets a lower power limit than default on the socket. I need to do more detailed testing, but for all core loads, it doesn't affect the clocks that much but it makes a bigger dent in the power consumption. And because it is a power limit, the clock still has the freedom to go up or down depending on how big a stress the load is. Weaker load, higher clock, without you having to manually tweak it.

[leadeater]

22 minutes ago, porina said:

Just do the power cap, don't touch voltages or clocks manually. The boost/limiting going on is far in advanced of old methods, and you're more likely to get the most out of more workloads than with old school tinkering. Have you played with any Zen 2 yet? The ECO mode is very interesting for crunching. This sets a lower power limit than default on the socket. I need to do more detailed testing, but for all core loads, it doesn't affect the clocks that much but it makes a bigger dent in the power consumption. And because it is a power limit, the clock still has the freedom to go up or down depending on how big a stress the load is. Weaker load, higher clock, without you having to manually tweak it.

Yea I would start with just the power cap then just lower the voltage to see what that does, wouldn't change anything other than that. Only other thing I'd potentially look at is IOD power and IF power, that to me seems like the best area to actually look at if at all possible due to that 75W near static load so even just 5W-8W less would be a big help.

 

Saying that I would use both a current clamp and a scope just to go full nerd to really know what is going on with each change.

[Jeppes]

13 hours ago, xtroria said:

Why the hell are people comparing this to mac pro?

 

people who are buying mac pros are mostly digital artists and their workflow are generally very mac specific and / or they don’t want to deal with the hassle of working with custom rig + windows + drivers

 

mac pros still have the accelerator card as well as metal.

 

this cpu will be mostly used by people who are more into big data computation or work where the mac software doesnt have any benefit over windows

The comparison is quite obvious as the price of Mac pro is so ridiculous in comparison. You could pay 1000 dollars for someone to build a tr system for you and get 300 dollars per year tech support and mac would still be more expensive. Support your local small businesses and spend on stuff that puts money into circulation instead of buying overpriced products from hundred billion dollar companies.

[porina]

53 minutes ago, leadeater said:

Saying that I would use both a current clamp and a scope just to go full nerd to really know what is going on with each change.

You have or have access to that kinda kit? Nice. I could do similar but I'm too lazy. I forgot the exact model, but I think it was a NZXT PSU I bought that can monitor the usage between the connectors e.g. EPS, mobo, PCIe... might be "good enough" in the first instance. The Corsair PSUs I have with monitoring don't split it by connector, only by rail.

[RejZoR]

1 hour ago, leadeater said:

Yea I would start with just the power cap then just lower the voltage to see what that does, wouldn't change anything other than that. Only other thing I'd potentially look at is IOD power and IF power, that to me seems like the best area to actually look at if at all possible due to that 75W near static load so even just 5W-8W less would be a big help.

 

Saying that I would use both a current clamp and a scope just to go full nerd to really know what is going on with each change.

Yeah, this way you can sort of cheat the boosting algorithm. Lower voltage means less heat and less total power consumption, meaning it'll crank things up more. RX Vega was similar, where undervolting it actually boosted things.

[leadeater]

15 minutes ago, porina said:

You have or have access to that kinda kit? Nice. I could do similar but I'm too lazy. I forgot the exact model, but I think it was a NZXT PSU I bought that can monitor the usage between the connectors e.g. EPS, mobo, PCIe... might be "good enough" in the first instance. The Corsair PSUs I have with monitoring don't split it by connector, only by rail.

Yea I have a current clap and they are actually pretty cheap, the scope I can borrow but there are some actually really good ones you plug in to a PC for cheap which means you can also do data logging, should look in to it yourself.

 

But the power monitoring from PSU's are pretty bad btw, like often so far off to not actually be useful if you need anything at all accurate. My EVGA NEX1500 has it and it's a joke when you compare it to something proper.