Question AMD 3950x vs 3rd Gen Threadripper

[pmjm]

Hi all,

I'm building a machine for 4k video editing and I will also do some gaming on it. I've been going back and forth between the 3950x and the higher-end threadrippers.

Is there any downside to going with a threadripper other than price? Will I get worse single-core performance on threadripper than on the 3950x?

Looking to have my cake and eat it too, I suppose. Just curious about the downsides of threadripper if I spend the money.

 

[jojesa]

For video editing and gaming I would consider the 3950x.
Threadripper is workstation CPU, has a lower core and boost core speeds and performance will be lower compared to the 3950x.
Threadripper will not be able to keep up with the 3950x on single or multi-threaded applications (hanbrake, blender, Adobe apps, DaVinci, etc.).

I think AMD is aiming the 3950x at content creators and it also performs very good at gaming and you could keep it cool with a good air cooler.

 

[Gmoney06ss]

You may want to wait for reviews to surface. The new threadripper will have a slightly higher base clock with a slightly lower boost clock, and a price tag double the 3950x.

 

[pmjm]

Got a possible answer to my question. I watched the Full Nerd podcast with AMD's Rob Hallock, he said that the single-thread performance of the 3rd Gen Threadripper series is on par with a 3800x.

Since I'm doing a mix of single core and multi core workloads (will also use Photoshop and After Effects a lot), I think I'll stick with the 3950x.

 

[drea.drechsler]

Got a possible answer to my question. I watched the Full Nerd podcast with AMD's Rob Hallock, he said that the single-thread performance of the 3rd Gen Threadripper series is on par with a 3800x.

Since I'm doing a mix of single core and multi core workloads (will also use Photoshop and After Effects a lot), I think I'll stick with the 3950x.

What sets TR apart is it's memory capacity along with 4 channel bandwidth and connectivity with many more PCIe lanes available. That should continue to be the clear advantage vs. 3950 with Gen 3.

Gamers Nexus (Steve Burkes) clearly dislikes 3950 as excessive for gaming-only use case but does feel it's a great choice as low-end HEDT and especially if you like to fire up a game as soon as the whistle blows. But where connectivity and 4 channel memory is important, TR clearly will be the preeminent HEDT/workstation platform.

But even so nobody really knows how TR-3000 will perform, and especially so on the edge cases, until it's reviews are in. And it sounds like yours will be an edge case, so if you can wait it might be wise to do so.

 

[moltenCrazy]

What sets TR apart is it's memory capacity along with 4 channel bandwidth and connectivity with many more PCIe lanes available.

+1

I was looking at both the 3900 and 3950 as possible options for a VFX/3D/Film/Video workstation. They're both fantastic CPUs--the price/performance ratio is just crazy--but at 24 PICe lanes, they're not viable for me.

Why? GPUs. My machine will be running 3 GPUs (and possibly 4 at some point). Even with just 2 GPUs--one running at 16x and one at 8x--that's it for the PCIe lanes on the 39xx Ryzens.

If you plane to take advantage of any multi-GPU usage/acceleration for your work, keep the Threadrippers in the mix until you see all the reviews like drea.drechsler suggested.

And to be clear, when I say multi-GPU I'm not necessarily talking about SLI/NLink. There are a number of DCC packages out there that can distribute jobs across multiple GPUs, either in parallel or on a job-per-GPU basis. It's great because it's like having additional processing nodes cranking away for you while maybe a lower-end GPU handles the system tasks/monitors/etc. and you get to keep working.

For me, it's really the number of PCIe lanes (64!) that separate the Threadrippers from the 39xx Ryzens (24).

Oh, and as for memory...if you get a TR plan on an 8-module kit (8x8GB, 8x16GB, or 8x32GB). The previous generation TRs really benefited from saturating the memory channels, and the results that have already leaked on Geekbench for the 3960x show that this seems to be the case for the new ones as well.

--
mC

 

[pmjm]

Thanks guys, I am going to go with a 3970x now that I've seen the benchmarks. The 3960x is really the price/performance sweet spot for me, but it's the holidays and I'm gonna spoil myself.

 

[moltenCrazy]

Thanks guys, I am going to go with a 3970x now that I've seen the benchmarks. The 3960x is really the price/performance sweet spot for me, but it's the holidays and I'm gonna spoil myself.

Ha! Glad you're set.

Now if only I could make my choice between the 3960x and 3970x...

--
mC

 

[mdd1963]

Got a possible answer to my question. I watched the Full Nerd podcast with AMD's Rob Hallock, he said that the single-thread performance of the 3rd Gen Threadripper series is on par with a 3800x.

Since I'm doing a mix of single core and multi core workloads (will also use Photoshop and After Effects a lot), I think I'll stick with the 3950x.

Based on the reviews I've seen thus far, this clearly did not pan out....; to AMD's credit, the single core performance of 3960X/3970X is right there with the best of anything....and often on top

 

[TJ Hooker]

Threadripper will not be able to keep up with the 3950x on single or multi-threaded applications (hanbrake, blender, Adobe apps, DaVinci, etc.).

3rd Gen TR handily beats the 3950X in heavily threaded applications like handbrake and blender, and still meets or exceeds 3950X performance in adobe apps...

https://www.tomshardware.com/reviews/amd-threadripper-3970x-review/4

Edit: Oops, didn't realize I was responding to a month old comment.

 

[drea.drechsler]

Hi all,

I'm building a machine for 4k video editing and I will also do some gaming on it. I've been going back and forth between the 3950x and the higher-end threadrippers.

Is there any downside to going with a threadripper other than price? Will I get worse single-core performance on threadripper than on the 3950x?

Looking to have my cake and eat it too, I suppose. Just curious about the downsides of threadripper if I spend the money.

There is a definite downside going with 3950x vs the higher-end TR chips. First thing you'll give up with a 3950X are cores. 3960X and up have more cores so core aware apps will love the higher end...and so will you when it gets done quicker. If you're a professional with clients breathing down your neck, it's the way to go. The second thing you'll give up with a 3950X is connectivity. TR's have a lot more PCIe lanes so more high-speed NVME, more data storage options, more 16x add-in cards, more memory in more channels. If you ever needed it, you got a place for it with TR.

Also, Zen2's chiplet topology is implemented with TR 3000, and so the latencies associated with earlier TR generation is fully mitigated. Very cool! no more rebooting to change memory modes to play a game between renderings!

But that comes at a cost, of course. Not just the CPU but 3960X and up lands on the all-new TRX40 platform. They're way more costly motherboards even compared to X570, top-end for 3950X.

I suggest you watch a few of the YouTube reviews...especially HWUnboxed. Steve has a streak of common sense an Aussie mile wide. Very refreshing.

 

[jon96789]

You just have to make sure that your apps can access all the cores... For example, Handbrake runs best with eight cores. While more cores are recognized, it does not utilize all of the additional cores at full capabilities...

 

[TJ Hooker]

You just have to make sure that your apps can access all the cores... For example, Handbrake runs best with eight cores. While more cores are recognized, it does not utilize all of the additional cores at full capabilities...

A 16 core 3950X takes 62% of the time for a handbrake encode as a 8 core 3700X (vs perfect scaling of 50%). Seems like decent scaling to me.
https://www.tomshardware.com/reviews/amd-ryzen-9-3950x-review/4

Edit: Although, in comparison, a 3700X scales nearly perfectly compared to a 3600X with h.264 encoding (76% vs expected 75%). But for H.265 it's much worse for some reason (85%).

Edit2: Another thing to consider is that the all core boost speed of the 3950X might be a bit lower than for the 3700X, which would impact the scaling result.

 

[drea.drechsler]

You just have to make sure that your apps can access all the cores... For example, Handbrake runs best with eight cores. While more cores are recognized, it does not utilize all of the additional cores at full capabilities...

But you can launch multiple instances of handbrake. Assuming you have several clips to render out (not infrequently the case) each instance, abely assisted with Windows' scheduler one hopes, should find it's own 8 cores waiting in the stable.

Much faster than processing serial fashion in a queue.

 

[drea.drechsler]

A 16 core 3950X takes 60% of the time for a handbrake encode as a 8 core 3700X (vs perfect scaling of 50%). Seems like decent scaling to me.
https://www.tomshardware.com/reviews/amd-ryzen-9-3950x-review/4

does it not also depend on the codec being used? might h.265 perform differently than h.264, for instance?

 

[TJ Hooker]

does it not also depend on the codec being used? might h.265 perform differently than h.264, for instance?

That article has results for both, scaling is very similar.

 

[drea.drechsler]

That article has results for both, scaling is very similar.

hmm...I thought I'd read something about that somewhere. Never really tested it myself, except what I've observed using h.264. I have found launching two instances of HB much more seriously loads my 3700X vs. one instance, at least when I monitor thread utilization it in Task Manager and HWInfo.

It's not often I have them, but I do (or did) have occasion to encode several clips at a time. I used h.264 and my goal is uninspiring...shrinking some camera video clips of our dogs playing down to cell-phone size for sending to my wife. But it definitely ran faster doing them in two instances vs. serial fashion in a queue.

I supposed that to lead some credence to an assumption h.264, at least, had thread utilization limitations.