Threadripper pro 7000,5000, 3000 versus new AM5 9950x

Ok here is my 50k foot broad stroke questions regarding these platforms.
1)Assume I can afford a 7055 16 core TR Pro Setup. Wrx90 And CPU. I don’t feel like I need all the cores YET. This is the latest and new boards that should remain valid till I need to upgrade to more cores.
2) Locate a 5965-75TRPRO with wrx80 maybe with the 24 to32core.
3) locate a 3675 TRPRO with similar specs.
4) Build a AM5 top of line rig and forget the multiple gpu for ml for now.
I want pci lanes for hooking up multiple gpa to play around with ML and run my house Lab etc on it.

How long do you guys have the Threadripper pro rigs in service. It seems one could last me 10 years if I get brand new series and have ability to update to more cores on wrx90 in 5 years etc….

Mind you I could probably build two or three 9950x with 670e boards for the cost so there’s that. Not a gamer. Love AI Art, 3d cad modeling etc.

I currently have the MiniForum MSO1 w a 4060 I mess around with. See my profile. It’s pretty tricked out. Could use for my router and Nas and plex server when upgrading.

Thoughts and criticism welcomed. The lower core count 7k series TRPRO are doable budget wise and I don’t know if I really need more than 16 for the ML and AIArt starting out. Seems I could grow for many more years with this rig versus buying older gen with high core counts.

I Am looking to buy something before end of the year if I can make it work with the accounting.

2 Likes

If you need a write-off before the end of the year there are a few charities I could recommend.

Otherwise, you don’t seem have a clear need/usage profile that requires new hardware.

The answer is very simple: if you don’t know why you need a Threadripper platform you don’t need it!

5 Likes

So which limitations are you hitting on this system? VRAM?

You can do two GPUs quite reasonable on AM5. PCIE x8 is not too much of a limitation, for shure not so much that 3x the cost makes sense. Bang for your buck AM5 + 2x3090 is your best option IMO. If that doesn’t suffice, run experiments locally and rent a cloud GPU once in a while for a few hours. When you outgrow that, well you better be making money on this or have a lot of disposable income to justify it.

Even 2xA6000 on AM5 would be an option to consider. Gets you 96GB of VRAM, and it’s still cheaper than 4x4090 in a threadripper system. Yes, it will be slower but much more manageable (lower power, 2 slot GPU, etc.)

But I agree with @jode, unless you know you need more than 2 GPUs or more than 192GB of RAM, just stay on AM5.

2 Likes

Duly noted. If I have it in the budget is my point. At this point it looks like ill wait. My charity for the year is maxed out. I am a civil engineer and have plenty of use cases for a threadripper. The PCIE and memory ougbt to provide a longer lifecyle imo. Say 10 year versus 3 or so. I invited criticism so I guess this falls undef that category.

The AM5 system will work for a period i suspect. Civil 3D loves alot of memory on large models and does work well on TR; many offices run them.

I need to clarify my question to get informed on the difference between the 3 generations of TR. Is the new generation 7k going to be valid for a while and is it worth getting one of those with a lower core CPU and adding another later as deemed necessary? Is the 3k or 5k a better bargain assuming a 7-10 year lifecycle.

Hi there. Appreciate the comment. Ive been trying to figure out the lanes and how gen 5 speeds allow on to break up the lanes for multiple gpu use.

So I can get a max of 2 GPUs at 8 lanes each? I beleieve the cards 3090 are PCI gen 4. If my maths correct if i can split up the lanes it only takes 4 lanes to run a card. Ie i can get 4 gpus on a bifurcated 16 lane pcie 5!?

Thank you for the input and for the helpful info.

I am thinking ahead and trying to buy soemting that will last and over time ill get my money out of it meanwhile it will be fun build on if that makes sense. I mean i dont need a 911 porsche either but its alot more fun than my wifes rig.

In principle. In practice PCIe 4.0 breakout rather easily gets a bit iffy even with redrivers and retimers in place. Not much on 5.0 yet, but doubling the signal rate’s not going to make it any less fussy.

For simply slotting GPUs, ATX’s mainly a two GPU form factor at this point. If they only need 5x4 then x8/x8 bifurcating AM5 PEG’s fine and Threadripper’s 80-120 other lanes don’t add anything.

Two GPUs at 5x4 or 4x8 is also only 31.5 GB/s max unidirectional, so two DDR channels might be ok, but big models suggest to me either paralleled 9950s or Threadripper for DDR rather than lanes.

10 years I think is pretty optimistic. We don’t do FEM but our compute intensive stuff is often 2-3x faster on Zen 5 than Zen 3. So the main difference between a 5975WX a 9950X might end up being able to run single model instances that need more than 192 GB DDR. Possibly at the expense of waiting longer on smaller ones depending on the workload’s core and DDR channel scaling.

We also see 9900X typically at least tie 7950X, demolish it on AVX-512, and 7950-60-70X comparisons suggest 9950X has a decent chance at beating 7970X depending on the workload. Don’t have 9950X data yet but I suspect it’d have to be a fairly memory bound workload for a 7955WX to be advantageous. Right now we’re on a 5-6 year refresh but, for any major compute push, all the Zen 3 and 4 hardware’s already effectively obsolete.

Our workloads usually shard nicely across machines, so 3x9950X with 2x48, no contest. Maybe 2x9950X + 9950X3D depending what exactly AMD launches. 2x64 if the UDIMMs show up for that.

Me three.

2 Likes

So in your experience doing anying with zen3 or zen4 with any product is a no go. It seems we dont know enough about zen 5 capabilities to state it will run more than 2 GPU today effectively. Which that today would give me plenty but in a year or two it may not. The pcie lanes and cores will eventulally be used imo. I understand Jodes premise you all agree with and I dont disagree except I do have use cases I intend to grow into and would prefer to buy once cry once. Isnt that how you approach these purchases? Forgive me if Im not using the platform to respond to the specific comments. Im learning to use it.

Are we getting up to 192GB on AM5 now? Thought its still TBD. In theory but needs tweaking to get it all. I do know the future use cases where ill grow into where i need the pci lanes and the memory but today it would be more than enough. Id like to have some room to grow and explore some new projects etc. Is this not the approach you guys use?

This is an interesting solution. The cards still go for around 4k it seems but thatd alot of vram. One of my ideas was to use multiple used high VRAM cards for a while or 3090s.

No. We just add a current gen desktop at the top of the waterfall. Our fiduciary standards require I reanalyze Ryzen-Threadripper annually and Threadripper always loses due to lower price-performance. By how much depends mostly on whether the analysis date happens to fall in the window where Threadripper’s caught up to Ryzen’s arch gen.

No. We usually reconfigure hardware every few months. Nice to keep motherboard, processor, and system drive together but splitting them or changing up the system drive isn’t that big of a hassle with TPM keys and domain records. Expensive monolithic’s more a high uptime, large OEM service contract, or workstation-server config fragility approach, it seems to me. Desktop compatibility’s high enough self-insured, self-maintained use can be more flexible.

192’s been supported for 19 months, 256 for 10. We’ve been running 192 for over a year and will move to 64GB UDIMMs once they’re available. In the several 192 threads here there’s a lot of attention to 2DPC versus 1DPC clocks and just about no attention to 24 and 32 core X or 48 and 64 core WX Threadrippers having the same number of cores per channel as 12 and 16 core Ryzen. Socket power per core and core clocks are also often neglected.

I don’t know of WX benches useful here and X data is minimal. The CFD thread looks at Comsol’s memory bound behavior and finds essentially linear 7950X-7960X core scaling (+50%) with +18% from 7950X-7950X3D and +30% 7950X-9950X, which is an order of magnitude lower marginal return on Ryzen-Threadripper than Zen 4-5. Depth of overclock, timings, and OSes vary but it appears plausible 9950X3D might outperform 7960X.

They’re fine. What I’m running on a 9900X most weekends would take all week on a 5950X, though. For other workloads 9900X and 5950X is at least a tie. Depends mainly how SIMDy the hot paths are, how L3 resident the data is, and how linear the thread scaling is.

4 Likes

I, personally, would mostly get in line with the other comments here, they are not bad advise. I however see the reason differently. There are very valid use cases for having a Threadripper system. Be it more or higher bandwidth memory, simply more cores, or more PCIe lanes.

Instead of buying one system that is supposed to last 10 years rather buy a new system every four to five years and sell the old one. The reason for that is simply that over the generations new features get added. Like modern GPUs now also do AV1 encode, more modern processors have more modern versions of AVX that increase the speed for certain workloads. When you upgrade in a more regular interval you will be able to make use of the speedups these new features bring. Instead of anticipating what the world will look like and how your use case will look like in 10 years evaluate that again after like half the time.

That does not mean it is a bad idea to go with Threadripper. I myself bought a used 3995WX with 64 cores about one and a half years ago. My use case is virtualization and in this case more cores are more better. Plenty of PCIe lanes come in handy too. Used Threadripper and Epyc systems can be much more affordable if you can strike a good deal. Be aware though that many offers are priced much too high. For example I could sell my Threadripper for the same price now that I bought it for one and a half years ago, because back then I found a good deal and many people today are still wanting more money then it is actually worth.

The difference between the 3995WX and the 5995WX is mostly in the single core performance if only a few cores are in use. You can have a look at the link I’ll attach to this message an see how the single core performance of the 5995WX is much better, while the multi core performance of both processors is almost identical.

For me the 3995WX is a great processor that does it’s job well. But when I am gaming on it you can definitely feel the, compared to today, low single core performance of those Zen 2 cores. Also I am missing all those nice AVX enhancements of Threadripper 7000, or soon to come Threadripper 9000, that benefit AI workloads.

It is really also a matter of taste, I prefer older workstation and enterprise hardware over the newest consumer or workstation hardware. Modern consumer hardware does not fulfill all my virtualization and PCIe needs and newer Threadripper or Epyc hardware would not be a financially sound decision for me since this machine is my personal test bed, but not an income generating machine.

Alright, I rambled a lot, but maybe these thoughts are a help to your decision making.

https://www.cpu-monkey.com/en/compare_cpu-amd_ryzen_threadripper_pro_5995wx-vs-amd_ryzen_threadripper_pro_3995wx

1 Like

If you think you could profit from a little more cores, a moderately higher memory bandwidth and more PCIe, maybe TRX50 would be a good middle of the way solution for you to not break the bank, get a suitable machine that does the job and will last you well a couple years and then you upgrade a little sooner then 10 years.

1 Like

for production in enterprise: 3 years
for less mission critical: 6 years
Just be sure to replace all drives and rethermal paste every 3 years

My favorite charity is “cash”
hollar and we can get the address you send that check to…

1 Like

This is awesome info. Im a little behind the curve on some of the technical terms here but it seems you prefer the 9900- over threadripper and are making a case that it performs better than TR. Do you co sider TRPRO when reassessing your firms needs? Just curipus because so far youve mentioned TR systems.

Regarding memeory i guess i misunderstood Wendell when he did the video or perhaps it was an old one. I think at the time he was sayi g to use just 2 48gb sticks of ram. Sounds like you took it to another level.

So as of not they outperform in many ways and simply lack the pci lanes one might need. It would seem to me that i could go with the 9050x and setup an other less expensive rig to do the VM, home server sort of tasks. Maybe even use my ms01. I have already put proxmox and linux on it and it screams for a little machine. I appreciate the time you put into responding. You sound like youve been doing this a while and im sure youre busy. Again thank you.

Tbh it seems the value I thought a TRPRO all in one setup for a long lifespan isnt there.

Yes. We look also at Xeon. Though it’s mostly just my extended workgroup as our computational needs are somewhat atypical. Other groups tend to either be fine with decent desktops or laptops or are doing rackmount compute.

This’ll probably prompt some objections but I’d suggest when workstations show reasonable return on investment is when sharding tasks into desktop size requires enough effort labor costs approach the riser in hardware cost. Since hardware is cheap compared to people some workloads hit that crossover instantly. But a lot of them never do. It’s the ones in the middle where the decision gets complicated.

2x48GB will pretty well own 1DPC until 64s are available, yes, and we definitely see diminishing numbers of workloads benefitting from 2x48, 4x32, and 4x48. Requirements vary widely with the application space though so, as always, know your workloads DDR demands and response. Often 192’s overkill, other times it’s not even entry level, sometimes you really want 1DPC, a lot of the time 2DPC is an adequate compromise.

But most of the lanes are unusable in a slotted dual GPU build unless you spend for two slot workstation blower cards. For example, with three slot transverse finned dGPUs

  • ASRock TRX50 WS gives 5x16 to one GPU, 5x8 to the other, and 4x12 are blocked by the GPUs. The remaining 5x16 is mechanically accessible but using it blocks the upper dGPU’s intake. So you spend for an 80 lane IO die pretty much to get +8 lanes for one GPU.
  • ASRock WRX90 WS gives 5x16 to two GPUs. The other 88 lanes are blocked or in the intake slot. So you spend for 120 lanes to again get +8 lanes to one GPU.

With four slot GPUs returns are potentially lower. There is some M.2 to MCIO and SlimSAS conversion on workstation boards but that’s mostly not so different from desktop lane allocations. So to some extent it starts to look like a really complicated way to get a board with a 10 GbE port.

If it’s slot one GPU and riser the other that’s quite a bit better for slot access by other cards. But maybe like half the risers advertised as 4x16 actually run 4.0, I don’t know of any 5.0 ones yet, and actual 5.0 support with retimers is going to be expensive. So buying up the IO die to riser potentially ends up an expensive way to plug in a GPU at 3x16.

FWIW, we don’t do machine learning past small LLM scale so our solution’s just been to stop at a single high end dGPU. I don’t think X870E dual’s necessarily out of the question but we’d have to get the electricians in to sort building power first.

2 Likes

i started looking these over as well. I have learn alot chatting here and did some research on the zen series and realized that it matters alot what architechture i use. it seems to me HEDT and TRPro are still on zen 4 series and an update may be coming in near term. it really doesn’t appear wise to go in deep now and i see your point. Appreciate the input.

1 Like

Thank you.

I (finally) answered the Threadripper efficiency question over here:

1 Like

Thanks again for the detailed responses. I did a deep dive on the architecthures of zen to understand some of it. the consumer series are using Zen 5 while the TR series are still on Zen4. I think your successes with your systems is related to that if im not mistaken and i believe the TRPRO series ot TR hardware will roll out Zen 5 soon and it would be a waste to buy today even if i can justify the TR series.

1 Like