UK BOINC Team Forum

AMD 64-core, 128-thread Threadripper 3990X

https://hexus.net/tech/news/cpu/137129- ... 2ZZR1nRFAA

Well since the 3970X (32 cores / 64 threads) is priced at $1,999 MRSP, I doubt I'll be spending £5,000 (?) just on a new CPU any time soon

Woodles wrote: ↑Mon Nov 25, 2019 4:07 pm Well since the 3970X (32 cores / 64 threads) is priced at $1,999 MRSP, I doubt I'll be spending £5,000 (?) just on a new CPU any time soon

Yeah, he mentions that he'd expect it to retail for not less than $4,000. The EPYC 7742 is already $7,000.

The math is interesting though. I had done the same calculation of base clocks and expected overall performance between a Threadripper 64 core and the 32 core 3970x. Given a 3GHz all core boost vs a ~3.9GHz all core boost, one might reasonably expect 64 core performance to be ~50% higher than the 32 core, so certainly far from double. That being said, the underlying PSU / motherboard / cooling / drives / etc costs are all meaningful and don't change much between the chips. Both have a TDP provision of 280w, so lots more available power per core in the 32 core chip.

The 2970x, as I'm sure you've seen, has translated from $1,999 to £1,900 here. We sure don't get a good deal on PC hardware in ye old Blighty …

Still, very exciting. I mean, the 16 core threadrippers were only a couple of years ago and now they seem terribly old hat!

I just posted it as its an amazing thing to see a (non-server) CPU that runs 64 cores, 128 threads, at 280w TDP.

chriscambridge wrote: ↑Mon Nov 25, 2019 4:50 pm I just posted it as its an amazing thing to see a (non-server) CPU that runs 64 cores, 128 threads, at 280w TDP.

Jaw-dropping.

I wonder how it will cope with 4 memory channels vs the 8 that presumably EPYC has. They will have to have some die / chipsets connected only indirectly to system memory.

chriscambridge wrote: ↑Mon Nov 25, 2019 4:50 pmI just posted it as its an amazing thing to see a (non-server) CPU that runs 64 cores, 128 threads, at 280w TDP.

Agreed, it's very impressive.

About that 280W TDP .... https://www.guru3d.com/articles_pages/a ... iew,5.html

485 actual Watts fully loaded! Looks like it's going to have to be a big water cooler if you want to get anything like the full potential.

A nice breakdown of current TRX40 Mobos

https://uk.pcmag.com/gallery/123749/fir ... readripper

$4,000, cheaper than i expected (of course, it'll still be £4,000 for the UK )

But 128 threads at 3GHz

AMD CEO launches the 3990X at CES

https://www.youtube.com/watch?v=t66WLU0MFAM

In truth it smashes (on freq) the 64c EPYC and most poweful Xeon(s)!

And this is the full AMD line up video at CES:

https://www.youtube.com/watch?v=8c8i3t6oIPA

Amazing. I was waiting to see this!

I know that the CPU and interconnects draw about 320w on a 3970x in order to reach 4GHz all cores under heavy load, circa 280w for ~3.85GHz. The underlying chipplets are going to be similar in this 3990x, although might be even more stringently binned / hand picked for lower voltages per frequency level.

If you’re up for exceeding the 280w default, I expect that you’d be able to get this to maybe 3.6 GHz for 500w? Just speculation of course. Definitely need chunky water cooling for that one!

I think all core default freq on the 3990X is 3.40 GHz; thats what some websites are reporting.

According to the AMD slides, base clock is 2.9GHz. Surely that's the maximum all core frequency? Otherwise, why not list 3.4 GHz as the base?

Genuine question, I've lost track of what various base/boost/turbo specifications are meant to mean these days

Or is 2.9GHz all core boost at 280W but you can get more with a higher package power?

2.9 will relate to the (Edit: minimum guaranteed) all core performance within the stated TDP. Sounds about right - 32 cores at 3.7 (it actually runs a bit higher) vs 64 cores at 2.9 to reach the stated 280w TDP.

The actual frequency and Power usage can (will) be higher, and of course more so if you give it a little overclocking nudge (e.g. my 500w comment).

Thanks Damien, sounds like I'll need to source some liquid nitrogen

I think that you'd want a good quality 360mm or 280mm radiator to get the best out of the CPU (not including graphics cooling). That would give you some room for nudging the clocks up and consuming perhaps close to 500w. The good thing about these threadrippers, is that the surface area is large and therefore the heat dissipation opportunity (into a water cooling system) is also large.

2.9 will relate to the all core performance within the stated TDP

Nooo thats not correct.. normally base is lowest freq, turbo is max eg on 1 core, and all cores is normally about middlepoint between the two; it is at least on Xeons.

Lets take the 1950x as an example to see if that math holds up on TR's.

Base is: 3.4 MHz
Boost (turbo) is: 4.0 MHz
All cores is: 3.7 MHz

(4-3.4)/2=3
3.4+2=3.6

so pretty close.

This website often lists the exact freq per core, plus many extra useful elements for most CPUs.

https://en.wikichip.org/wiki/amd/ryzen_ ... pper/1950x

Hi Chris,

I agree that the chips can run all-core higher than the base, they just need to do that at a higher TDP than they state. If you’re only able to dissipate the stated TDP heat, then the CPU won’t necessarily run any faster than the base. If you can dissipate more, then it may well run faster.

I read this Anandtech article a while back and that’s how I’ve been thinking about base clocks since then.

https://www.anandtech.com/show/13544/wh ... -tdp-turbo

I’m finding that the 3970x is religious about sticking under 280w, although it’s easy enough to give it a higher power target and let it boost everything higher.

I hope that the default 3990x is 3.4GHz. I just think it will be doing that at a TDP much higher than 280w.

Thanks for the wikichips link - that site looks incredibly useful.

Yeah I think your right on this because AMD use cTDP, which is customisable TDP so perhaps the all core freq isn't as fixed as it is with Xeons (eg non-O/C).

From what I can gather a mobo/system manufacturer can actually set the cTDP (less or more) so that it works with their cooling solution (in a given system).

The more I dig, the more murky this looks. The 'TDP' that AMD lists for the 3900x is 105w, and then the actual full CPU package power draw (the 'PPT') is more like 142w. Then AMD lists 280w as the TDP for the 3970x, but that Threadripper ‘TDP’ is actually the full package power envelope (PPT) for this chip, so a completely (more fully inclusive) different metric, with less need to overrun it.

Then, looking back over the Ryzen chips and generations, some are real sticklers about staying within their PPT limits and others are happy to overrun them for a higher performance.

Just having a play, loaded with 64 Rosetta threads and under excellent cooling, I can see the 3970x doing all core boosts as high as 4.17GHz (vs base clock of 3.7GHz, so like you were suggesting, Chris) AND it is well inside its 280w PPT limit. Some other times I've seen it butting up against 280w and ‘only’ clocking 3.85Ghz, only marginally better than the base clock. It does then fluctuate with workloads and probably ambient temps, etc.

To examine this further, I wound the power limits down to a level of 3w per core, which is what Anandtech suggested might be the actual per-core power available to a 3990x (the rest of the 280w PPT power going to the 'un-core', or all of the data transfer interconnects etc). For this particular Rosetta workload (which seems a bit of a best case) ... the chip is actually still clocking 3.66GHz!! In other workload and conditions, I have to assume that it might clock closer to its base clock?

Perhaps this architecture at 7nm is a real step-change and plays by new rules ... it seems pretty different to the way Intel works in any case. There is less need for very high watts than I had thought ...