For those that have been waiting for more Threadripper options, TR Pro has been announced!
tj_1ca9500b
Posts: 2,057
https://www.amd.com/en/processors/ryzen-threadripper-pro
Short form, 128 PCIe lanes, 12-64 cores, 8 memory channels, 7nm, 280W TDP, PCIe4
Variants:
3945WX: 12 cores/24 threads, 4.0/4.4 GHz, @ 70.75 MB Cache (L1-3)
3955WX: 16 cores/32 threads, 3.9/4.4 GHz, 73MB Cache(L1-3)
3975WX: 32 cores/64 threads,: 3/5/4.35GHz, 146MB Cache(L1-3)
3995WX: 64 cores/128 threads, 2.7/4.3 GHz, 292 MB Cache(L1-3)
Note that my cache totals are the total of L1-3 (L1&2 add 6.75-36MB to the totals) , hence a bit higher than in the press release below.
The rumormill indicates that this will be on a new socket (I'm not seeing official confirmation of this on the AMD page), but since they've doubled the memory chanels, yeah that's probably a thing. Also, no indication if you could run these on existing TRX40 boards, at least not that I've seen as of yet.
Here's the associated press release mentioning the new Lenovo Threadripper Pro Workstations:
Anyways, for those of us looking for lower core count options on a workstation class Threadripper motherboard... with 128 PCIe lanes and 8 memory channels just for fun, and with higher clocks than EPYC Rome...
Now to see how fast these hit the retail channels, and if these will be 'wide release' or 'OEM only' for the time being, not to mention prices.
Daz 3D is part of 
Connect
DAZ Productions, Inc.
7533 S Center View Ct #4664
West Jordan, UT 84084
Licensing Agreement | Terms of Service | Privacy Policy | EULA
© 2025 Daz Productions Inc. All Rights Reserved.
Comments
This tweet on the Anandtech feed indicates OEM only (for now?):
https://twitter.com/anandtech/status/1283079671811276802?ref_src=twsrc^google|twcamp^serp|twgr^tweet
It mentions a September release date.
And the Anandtech article:
https://www.anandtech.com/show/15910/amd-announces-ryzen-threadripper-pro-workstation-parts-for-oems-only?utm_source=twitter&utm_medium=social
Your next computer!
It has to be on a new socket.
While it is 8 channels the question is how many slots per channel will be supported. One of the big failings of the current TR3000 is that it doesn't support the higher density memory that would allow it to actually reach the sorts of memory capacities needed for the tasks that could use 64c/128t.
If it TR pro is 8 channels but 1 slot per channel and it also doesn't support LRDIMM's then it won't increase the total RAM capcity and won't change the issue.
It supports LRDIMMs.
Just finished watching the video. Thanks Sevrin!
The Lenovo option only as 1 dimm slot per channel as noted in the video, for a total of 1 TB of capacity, even though the new Threadrippers can theoretically support up to 2TB. Hopefully we get another option with more memory slots and roomier PCIe expansion later.
Kinda disappointed that the first offering only has space for 2 'double slot' cards. They do mention using 4 Quadro 4000's, which are single slot cards apparently. Looking on Newegg...
https://www.newegg.com/p/1FT-0015-000U3?Description=nvidia quadro 4000&cm_re=nvidia_quadro 4000-_-1FT-0015-000U3-_-Product
Single slot, 8 GB GDDR6, $1000.
Not sure how it stacks up against other cards performance wise. I'll have to google that and see. It's a Quadro, but not seeing any mention of NVLink... checking Nvidia's site page:
2304 CUDA cores...
https://www.nvidia.com/en-us/design-visualization/quadro/rtx-4000/
As for NVLink, took a bit more digging, aaaand NO!
https://www.nvidia.com/content/dam/en-zz/Solutions/design-visualization/documents/quadro-rtx-8000-line card-us-nvidia-1001494-r1-web.pdf
So with just 8 GB of VRAM and no NVLink to pool the memory, that is probably a deal breaker for some people. I like the single slot thing though! Hopefully there's a single slot 'RTX 3000 equivalent' card with more VRAM when those eventually get released, later this year maybe?
Hopefully Lenovo does a followup that supports a 'true Quad double slot' GPU setup. Seems like a waste of PCIe lanes otherwise...
The Anandtech article noted that people should let AMD know if they really want these new WX Threadrippers sold separately in the retail channel. If you are one of those people, might be time to give AMD a shout... the 16 core one is the one I'd look pretty hard at, as I do mostly GPU only renders and 24 cores is a bit of a waste and out of my price range at the moment, and AM4 comes up a bit short on the PCIe front. Stupid COVID...
In any case, late September, yay...
More threads in an application doesn't mean it requires higher memory capacity.
Yeh I got an email from AMD about it; shame they didn't send some cash too.
Yes, actually it does.
Every thread is a seperate process. Every process has it's own stack, a stack is a segment of RAM usable by that process.
In very general terms a program has two kinds of memory stack and heap. Stack memory is where local variables are stored, int x or string foo. Heap memory is where dynamic variables are stored like string *foo = new string; You can pass pointers from the heap across processes in the same program without issue. Passing stack variables out of their context can get you in real trouble.
And that is just the very most basic issue.
right as i was getting ready to rebuild my machine with a TR 3960x too
guess i am gonna wait for performance/pricing/availability/socket/mobo information on these ones now. LTT/GN/JZ2C will no doubt do run downs on them when they can get their hands on them, should tell me everything i need to know
Does it have OpenCL support, ie. can you use dForce?
For now, Wendell has shared a few thoughts on this. We'll have to wait a couple of months for launch day though...
The 'usual suspects' are no doubt petitioning Lenovo right now for review rigs...
Having a Threadripper, meh (for me). Of course it depends on what you're doing with it.
Daz Studio still only uses 1 core for many things, and my TR 1950x is slower loading scenes than my Intel 6700K. Identical libraries. Some things it is faster, but unfortunately most of my "work" is in Studio.
What I'd like to see from AMD is a higher level board with more PCIe lanes for the high tier Ryzen boards. That would really get my attention.
Based in the performance increase from an intel i7-3630QM to a AMD Ryen 7 2700 when a 5th Generation AMD Ryzen 32 core or greater Zen CPU does come about it would be very worthwhile to buy when those components get much cheaper when even newer Zen generations come about. I definately can't afford that HW new.
With Threadripper, for the most part if you are doing Iray rendering/GPU only, as others and myself haveI noted previously, the extra cores are kind of a waste. For me, I'm mainly after the extra PCIe Gen 4 lanes. Sure, we may only be talking a couple of percentage points improvement in performance, but if one is building a new system anyways...
That's why the 12 and 16 core parts interest me. As I noted, the AM4 boards are a bit anemic on the PCIe for quad GPU setups, assuming you can even find an AM4 board with 4 PCIe 16 slots (MSI has one), but those will be running at much slower PCIe speeds, and sharing those lanes with storage and other things. THAT is the point of Threadripper, for us 'power user' types...
Sure, EPYC Rome is out there, but those CPUs are a bit slower than Threadrippers, and the EPYC motherboard options for a workstation setup are less inspiring ATM...
Quad GPU is just over 3.6x faster than a single GPU in Daz Studio, according to the Iray benchmarking thread, and of course in other similar rendering benchmarks. I'd rather direct my pennies to an extra GPU versus more CPU cores given the option, but I'm not interested in the older Threadrippers ATM. That 8 core 1900X Threadrippper part is a lot of bang for the buck though, even if it is 'just' 1st Gen Threadripper, but the earlier Threadrippers are significantly slower than the 7nm parts, which comes into play when using other apps that rely on the CPU more, say when converting animations into different, more efficient video formats.
I'd probably use animation more in Daz Studio if rendering all those individual frames didn't take so long. With a quad GPU setup, well cutting down the time to render all those frames becomes very significant...
TRX 40 otherwise would be just fine for me, but as I type this, while the 24 core Threadripper is getting cheaper, it is still well over $1000. And as I and others have noted, a lot of those cores essentially go to waste most of the time in Daz studio. It's nice if your render goes CPU only, but usually at that point I just quit the render and do some scene optimizations, or set up two render passes, as two passes on GPU will still be much faster than a single pass CPU only render.
the 1000 and 2000 series Threadrippers are, well, dated at this point. I'm looking forward a bit with this purchase, to when PCIe gen 4 becomes more of a big deal with the next gen GPUs, and of course faster storage options. Sure, SSD's are plenty fast now, but shaving a few seconds off here and there adds up over time.
If I was doing 3Delight, then those 24 cores on the 3960X would be very nice to have, but I'm on the Iray bandwagon, and pinching my pennies for additional video cards, soooo....
Those 12 and 16 core 7nm Threadripper parts should be fairly affordable, IF they ever end up in the retail channel. Hopefully enough people are able to convince AMD to do this eventually. And of course eventually in the next year or so there will be new and improved Threadrippers hitting the market, as the 'Vermeer' CPUs are expected shortly, as are their EPYC counterparts.
As I noted above, I'm really interested to see if Nvidia comes up with a single slot GPU option for the next gen CPUs later this year/early next year. If they could squeeze say a 3080 Ti into a single slot config, or otherwise make a single slot card with say 12GB or more of VRAM, yeah I'd be looking really seriously at said card. There's NVLink as well, but so far the single slot card (RTX Quadro 4000) doesn't do the NVlink thing, as I noted above.
As for why not Intel, yeah they've been overcharging for CPUs for well over a decade now, and are just now lowering their prices to counter a resurgent AMD. And I've enjoyed woking with all of my previous and current AMD systems, so brand loyalty and all that...
Enough ramblling. Just explaining my logic as to why 64 or 128 PCIe lanes matters to me, and why the 24 core parts seem like a bit of a waste for my usage case, which is why the 16 core 128 PCIe lane Threadripper part is so intriguing...
BTW, if anyone was wondering if there is a waterblock option for a single slot RTX Quadro 4000, yes there is...
https://www.ekwb.com/shop/ek-pro-gpu-wb-rtx-4000-ni-inox
So if there is a next gen single slot successor to the RTX Quadro 4000, with say 11-12 GB or more VRAM, and EKWB or some other company makes a water block for it... well I think you can see where this is going!
No, it actually does not.
This statement could not be any more incorrect. From the Windows Dev Center:
Yes, that's true, but the practical effect is negligible. The default size is 1 megabyte, which is enormous. Not that long ago, the default stack size was 4 kilobytes. But the point being missed by the above statement is that the memory is virtual and the kernel only actually backs one 4K page of it with actual, physical memory until the stack pointer actually references non-resident memory, at which point the kernel will map another page, 4K. This means that for most threads, the stack will consume an insignificant amount of actual physical memory.
When you think about this, it has to be this way. Consider every single process in the Task Manager, all the threads that each might have, and all the hidden processes, and all the threads they might have. It would be an unacceptable waste if each used a megabyte of physical memory just to use 0.39% of it.
These statements are completely irrelevant.
@tj_1ca9500b I'm in complete agreement, and my experience has been the same. Threadrippers are a dream come true if you actually throw a parallel task at it, like compiling a large C++ project, but they just aren't any better suited for, say, running Daz Studio.
As you said, the real benefit is in the generous pool of PCIe lanes, and my (in)experience has taught me that it is best not to end up being your motherboard vendor's QA team by testing a configuration that approaches the limits of the motherboard, or areas the vendor did not test well: when I put 4 GPUs into my system at 16x 8x 16x 8x, on paper it should have just barely worked. But at boot time, PCIe negotiation killed my SATA ports and the vendor could not resolve the problem. If I had a TR with 128 lanes, my configuration would have represented a much smaller fraction of the available PCIe lanes.
I was unaware of the 16 core, 128 lane product. It would make a lot more sense for Daz Studio and my particular setup.
Thanks, @tj_1ca9500b this information was extremely valuable... I know what to do next year :)
I'm confused. Why did you post statements that disproved your own claims?
You are most welcome! COVID killed my upgrade budget, otherwise I'd have pulled the trigger by now, probably with some EPYC Rome system I cobbled together, so in this case the extra wait will help in the long run. Obviously, I've been giving this a lot of thought.
BTW, the reason the single slot GPUs interest me is because of the other add in cards I may need besides 4 GPUs. Sure, I can fall back on ribbon cables or something, but that introduces another point of failure, plus then there is the question of how to secure those devices attached to ribbon cables, so I'd rather just use the motherboard PCIe slots directly if possible. I have other uses in mind besides rendering for my next system, so extra flexibility will be helpful. I'd also rather not have to uninstall 'extra' GPUs just to install some other card if I otherwise have the PCIe slots.
I just want the moon and the stars. Don't mind me...
If the rumors are true, the greatest problem you, I, and a few others will face is the 1600W power budget before you have to upgrade your panel and run dual PSUs or something similarly exotic. It might be cheaper to just run two systems with two GPUs each. It'd certainly be simpler to build.
Well, with 2 PSU's you can always run a suitably beefy extension cord from a nearby circuit group that is underutilized, so that you aren't plugging both power supplies into the same 15A or 20A circuit...
Fortunately, I already have some pretty beefy extension cords that I used to power up some 1200 watt amplifiers back when I was doing the live music thing. I still needed someone to wire in a 100 amp drop from the main box for the outdoor gigs, but yeah...
I am hoping that Nvidia's 7nm parts are more power efficient as well as speedier, but I guess we'll find out soon enough. Not looking to overclock/not my thing, I'm already underclocking the1080 Ti I'm currently using to manage heat. This actually works out better, as at 100% wattage/load, it throttles pretty heavily, so dropping the power by say 20% or so actually results in higher clock speeds after the heat load peaks. It's still hotter than I'd like at 78c (note that I don't have AC, and my HTPC box isn't the most ideal cooling situation), but at least I get my renders done in the meantime.
I also do most of my rendering at night and during the early morning, when the indoor temps are cooler... I might invest in a new central air conditioner someday, but the heat doesn't really bother me/doesn't get above say the mid-80s (F, not C, obviously, otherwise I'd be dead!) indoors on the hot days. Not really relevant to this discussion, but different strokes for different folks and all that!
Is it really that simple? I've read online some horror stories about differing ground levels on different circuits damaging the dual PSU. I don't know enough about the subject to say anything about it.
I run 4 blowers and they don't throttle at all; performance has been a straight line. I was told by an engineer at System76 that blowers are strictly required for more than 2 GPUs. I'm not sure why blowers are not more popular, I love mine.
And I've heard that while the 7nm parts are of course more power efficient, NVidia "spent" the power savings to crank up the clockspeeds to increase performance and the draw is actually higher. These are rumors, of course, and we won't really know until we... know.
It has been years since I tried a multi PSU setup outside of a server rack but back in the day there were no end of issues. Even in servers the redundant PSU's don't run at the same time. They are fail overs.
You can exceed 1600W though. You'll just need a higher amp circuit. In the US standard household voltage is 120V (in reality it is 115 but call it 120 to stop certain people from freaking out). 1600W is roughly the draw from a 15A circuit with a margin for the efficiency loss for the PSU. But there is no reason not to go to a 20A circuit. which can safely support a 2000w PSU (assuming nothing else is on it). There are a few 2000W PSU's out there if you really want to build a rig like that. Just get an electrician to come in check over your wiring and do the wiring needed. You don't want a fire.
Sort of a side note, but I ran a 7 x 2080ti setup with 2 x 1600w PSU's.. power is the least of the problem with that config :)
Did a bit of reading. Thermaltake does make a 650 supplementary power supply, that 'slaves' to your main PSU. But yeah you'd be co-mingling power as you'd be pulling up to 75W through the PCIe slot, from the main power supply, and the 650 W only provides power to a pair of GPUs via two pairs of 6+8 pin connectors. I think the Thermaltake solution assumes that you'd still be working through the same power strip/wall socket.
So yeah, as you noted, weird things could happen as you'd have two separate circuit breakers co-mingling power by powering the slave PSU via a second circuit. I don't think those like that. It was a thought anyways...
This idea might work if you had a separate power supply for say water pumps or case fans via a separate controller maybe, but yeah I don't think pumps and fans pull that much power compared to GPUs...
My idea might work for an eGPU setup via Thunderbolt 3, where the only 'connection' would be a thunderbolt cable, but that kinda defeats the purpose of using lots of PCIe lanes in the first place. As long as it's just a signal cable that is the link, and not a full power cable... Thunderbolt 3 can provide some power though, so that's an interesting question...
The Quadro RTX 4000's pull only 160 watts according to the specs, but those aren't my first choice (only 8 GB of VRAM and fewer CUDA cores). As a comparison, the Titan RTX has a TDP of 280 Watts... Throw in another 200-300 watts for CPU, then there's memory, etc...
Obviously those quad GPU workstations that you see occasionally do this somehow! Just looked up a 1600W Power supply, it indicates a 16 amp power draw. That may vary slightly from model to model, but it's a ballpark at least. Might burn my house down though!
Things that make you go hmmm...
Is that a USA thing? I run 4 GPU's and everything else of a single 1600w PSU (admittedly through a UPS, but still)
They don't have warning on the wall sockets for "Do not stick your tongue into the socket", therefore they have only 115/120volts => higher amperage.
On this side of the pond, we do not need such warnings because we have tougher tongues and therefore we can have higher voltage.
They are definately suited for MD & Blender & Visual Studio & UE4 & Unity and becoming more and more suitable and relevant for more and more apps.
DAZ 3D on GitHub has created a project for a pipeline into Blender/Unity/and so on and I've like to see them find an opensource Blender plugin that uses USD format and make it easy on all to back and forth pipeline from DAZ Studio to 3rd party 3D apps, material, morphs, rigs & all..
It's more of an older house thing. Theoretically I have a 20 amp circuit or two to work with, but the wiring is older, plus other things are on those circuits as well... Tearing the older wiring out would require an extensive remodel, and yeah I'm not doing that...
Possibly I could run a new power drop (I know my way around such things electrician wise). The issue is finding parts for such an old breaker box, and whether I want to risk stressing the main any further... sometimes it's better to leave well enough alone.
It's a household voltage thing. In North America standard voltage is 115V (Do not start) and the usual household circuit is 15A so a 1600W PSU is what a household circuit can handle. In other places 230V is the household standard, probably where you live. A 15A circuit could handle a much higher wattage PSU on your supply.