Does Threadripper make sense at all? 2xEpyc?

[MadDuke]

Hi guys!   Would like some feedback from you concerning this. Mainly because I can think of a lot of professional workflow where it works like this, but of course can't think of all of them. So would appreciate from you to share some hopefully different view of the landscape and make me learn something new hehe.      Also. the reason why I'm putting the discussion here is that LTT is one of the only channels where some build like this could actually happen IMO.     

 

INTRO
Here is my train of thought.  Threadripper is all fine and dandy. For that price, Intel is just the worse option. Also, most probably when Intel goes over 10 cores (upto 18) with pushing down base clock speeds (nobody cares about boost clocks if you are doing operations which use all the cores) it will see even worse returns for money compared to their 10core parts. 

 

But, honestly. I don't see the need for threadripper in mid-term also.  I mean. It's not meant for gaming. You will basically not gain any single thing gaming on it compared to a ton cheaper 1700 OCed (or 1800x if no OC). 

 

I'm not going to count each dolar (like added for cooling) because, it's a cheap thing concerning the cost of the platform and possible uses.  Also, space constrains (Threadripper does fit in a regular PC case is a huge plus), but also. We are talking professional workflows here

 

THE MAIN QUESTION

Here is what I'm thinking

 

1xThreadripper 1950X 16C32T part costs >1.000 USD.   Overclockable to 3.9GHz on all cores (some go to 4.0, but, let's just leave 3.9 as a realistic expectation).  64 PCIe lanes

 

2xEpyc 7281 16C32T parts cost >2x600 USD = 1.200 USD with 32C64T.  Base clock, all cores go to 2.7GHz.  128 PCIe lanes

 

USEFULNESS
GPU Workloads:

So as I look at it.   Some people may buy into the Threadripper for it's 64PCIe lanes not caring for CPU performance or not needing it. (maybe their worflow is GPU specific and needs to have as many of them running on a single machine) But for them also it would be better to just buy the single 8C16T Epyc with 128 PCIe lanes for >400.  Even tough I honestly don't think price is an object when comparing what would be the cost of actually populating all 64 PCIe lanes wih GPUs (not to mention 128)  Space here doesn't play a role because with this many GPUs you are moving away from normal PC cases of course. 

 
CPU Workloads:

My logic goes like. If you are using the CPUs for rendering. Than you are using all the cores completely. In that case. The single core IPC is really not that important.  Even a slow 2.7GHz 32C64T  should be able to provide better performance in highly parallel environments than even a 4.0 GHz OCed 16C32T part.     The only workflow which could benefit from a faster TR chips could be memory speed conscious workflows (3200 MHz + for  TR vs 2666 MHz for Epyc)

 

Mixed CPU+GPU workloads
From the above mentioned scenarios, I don't see how a mixed workflow would be in any way worse since it would benefit from both of the facts mentioned above. 

 

VIRTUALIZATION

Nothing special to mention here. 

 

PLATFORM COST

Here the advantage is of course in the consumer market Threadripper part. But, my personal opinion is that for these workflows the prices aren't that big of a deal.   I mean. CPU prices themselves pale in comparison to the possible storage costs, GPU costs, memory costs...

 

FUTURE-PROOF
I have intentionally mentioned 2x Epyc 7281 16C32T parts.   Because it has a clear path of upgrade-ability towards the 2x Epyc 7601 32C64T parts for the mind boggling 64C and 128T system.   Which at the present, would require paying almost 7x the price premium for about 2x the performance.   

 

What is your opinion guys (girls included in guys for PC warriors)?

[samcool55]

Well, simple explaination. Market.

Threadripper is meant for enthousiasts who like high-end hardware and push it as fast as it can go and make it look as badass as they can.

Epyc is meant for the commercial market and comes with a lot of security advantages. But does lack cool motherboards, overclocking and other things enthousiasts care about.

 

If you want pure performance for your money than yes epyc might be a better option than threadripper but then again, it lacks features you might want.

 

Anyway, i'm glad this discussion can be a thing at all. Intel refused to do something similar like this (Xeon's weren't allowed to be a good alternative for consumers basically) and it was annoying AF because there were Xeon's that were actually a better option than an i7 (iirc it was an i7 with the price of an i5 + 20-30 bucks) they would do some stupid artificial locking to stop people doing it. (they ended up locking Xeon support to a C-chipset iirc which made them incompatible with B, H, and Z-boards)

 

So yes it might look a bit weird but imo this is a good thing. It's something i wanted for a long time and it's possible this could push intel into a position where they have to price Xeon competitive and put the prices of consumer cpu's down because if they don't everyone goes AMD or Xeon.

 

It will be interesting, the fact AMD is back in the server market is a sign it will likely be competitive for a while now. For now i haven't seen exactly a big battle between AMD and Intel but as time goes on, things will hopefully change.

[betam4x]

A lot of people are claiming that Threadripper is only worth it if you do X or Threadripper is only worth it if you do Y.  I'm going to take another approach.  Financially, can you afford to put together a Threadripper based machine?  If so, do it!   There are no disadvantages to Threadripper other than price and electricity.  Threadripper meets or exceeds intel offerings in every category, including gaming.  I am someone with workloads that will benefit from Threadripper. However, even if I weren't, I would still build one...because why not?  If you have good income and can afford it, go ahead!

[Mira Yurizaki]

Keep in mind that while Threadripper is a 16C/32T part, it's also a 2 node NUMA part (likewise Epyc is a 4 node NUMA part). So if you want to make the most out of Threadripper's performance, you need applications that aren't memory performance sensitive or at least, are NUMA aware.

[Quaker]

21 minutes ago, Mad_Duke said:

CPU Workloads:

My logic goes like. If you are using the CPUs for rendering. Than you are using all the cores completely. In that case. The single core IPC is really not that important.  Even a slow 2.7GHz 32C64T  should be able to provide better performance in highly parallel environments than even a 4.0 GHz OCed 16C32T part.     The only workflow which could benefit from a faster TR chips could be memory speed conscious workflows (3200 MHz + for  TR vs 2666 MHz for Epyc)

 

FUTURE-PROOF

 

I have intentionally mentioned 2x Epyc 7281 16C32T parts.   Because it has a clear path of upgrade-ability towards the 2x Epyc 7601 32C64T parts for the mind boggling 64C and 128T system.   Which at the present, would require paying almost 7x the price premium for about 2x the performance.   

 

What is your opinion guys (girls included in guys for PC warriors)?

CPU workload:

Keep in mind that the only time you will be using all thread is during the actual final rendering. While you are editing a video, you will only be using a few cores. The only difference that having more cores makes is in the time it takes to do the final render. Unless you are in a professional enviroment where the time is actually important, it's not economical to spend $1000 on a CPU just to save a few minutes render time.

If the number of PCIe lanes is important to you, you could use any less expensive Threadripper with fewer cores and just spend a few extra minutes drinking your coffee while it renders.

 

Future-proof:

Epyc processors are more geared to towards handling databases in large websites, etc. and I can see no advantage to them over Threadripper for video editing, etc.

 

[Bananasplit_00]

Anything that is neither AVX or GPU accelerateable and is easy to parallelarice 

[betam4x]

15 minutes ago, M.Yurizaki said:

Keep in mind that while Threadripper is a 16C/32T part, it's also a 2 node NUMA part (likewise Epyc is a 4 node NUMA part). So if you want to make the most out of Threadripper's performance, you need applications that aren't memory performance sensitive or at least, are NUMA aware.

Testing has shown this to not be an issue except for legacy applications.  Threadripper has a 'game mode' that can be used for legacy applications.

[MadDuke]

2 hours ago, samcool55 said:

Well, simple explaination. Market.

Threadripper is meant for enthousiasts who like high-end hardware and push it as fast as it can go and make it look as badass as they can.

Epyc is meant for the commercial market and comes with a lot of security advantages. But does lack cool motherboards, overclocking and other things enthousiasts care about.

 

If you want pure performance for your money than yes epyc might be a better option than threadripper but then again, it lacks features you might want.

 

Anyway, i'm glad this discussion can be a thing at all. Intel refused to do something similar like this (Xeon's weren't allowed to be a good alternative for consumers basically) and it was annoying AF because there were Xeon's that were actually a better option than an i7 (iirc it was an i7 with the price of an i5 + 20-30 bucks) they would do some stupid artificial locking to stop people doing it. (they ended up locking Xeon support to a C-chipset iirc which made them incompatible with B, H, and Z-boards)

 

So yes it might look a bit weird but imo this is a good thing. It's something i wanted for a long time and it's possible this could push intel into a position where they have to price Xeon competitive and put the prices of consumer cpu's down because if they don't everyone goes AMD or Xeon.

 

It will be interesting, the fact AMD is back in the server market is a sign it will likely be competitive for a while now. For now i haven't seen exactly a big battle between AMD and Intel but as time goes on, things will hopefully change.

 

I'm not saying that Threadripper is a bad platform or that people won't buy it. Heck. If I had space issues I would probably be buying it. Girlfriend is also doing 3D renderings so we will most likely go the 2x1700 route. She has me to build the systems and put a decent 3.9 OC on the chip and when she clicks the render button it can do distribute rendering on my PC also. So, that is the better route in my case.  I'm not going the Epyc route myself. This is just a menal/tech excercise for me.   And for example. In our and her workflow this would benefit us if we weren't working for other people and would have our own company. Heck, one Epyc server for rendering and as a storage server on a 10G uplink, 1G downling switch would just "fly". 

 

1 hour ago, Quaker said:

CPU workload:

Keep in mind that the only time you will be using all thread is during the actual final rendering. While you are editing a video, you will only be using a few cores. The only difference that having more cores makes is in the time it takes to do the final render. Unless you are in a professional enviroment where the time is actually important, it's not economical to spend $1000 on a CPU just to save a few minutes render time.

If the number of PCIe lanes is important to you, you could use any less expensive Threadripper with fewer cores and just spend a few extra minutes drinking your coffee while it renders.

 

Future-proof:

Epyc processors are more geared to towards handling databases in large websites, etc. and I can see no advantage to them over Threadripper for video editing, etc.

 

 

I agree. For solo video rendering workflow it wouldn't matter much.  For a whole office though it would.     But bear in mind that you can have a 1700, 1700x or 1800X workstation.    I'm really not proposing anyone to have an Epyc server and sit by it and work on it as a workstation PC.  You send tasks to it.   Also, believe me when I say to you. Some single "picture" renderings can go on for hours. And there is a huge difference between having something done in 8 hours, versus 2 or less.     I had one rendering on an 2600K @ 4.6Ghz that took 12 hours to complete.  That would be quite a long coffee staring at the screen doing nothing even with two Epyc CPU's   From a work perspective. Where you enabling your employees to be the most productive outweighs the cost.  One two week project earns you just with simple renders a couple of thousands USD.  If you can have technology to cut that by half. It's really worth it. 

 

2 hours ago, betam4x said:

A lot of people are claiming that Threadripper is only worth it if you do X or Threadripper is only worth it if you do Y.  I'm going to take another approach.  Financially, can you afford to put together a Threadripper based machine?  If so, do it!   There are no disadvantages to Threadripper other than price and electricity.  Threadripper meets or exceeds intel offerings in every category, including gaming.  I am someone with workloads that will benefit from Threadripper. However, even if I weren't, I would still build one...because why not?  If you have good income and can afford it, go ahead!

 

Look, honestly. For solo, high FPS 1080P whatever gaming (eSports) I would go Intel.  But there are literally more people doing 3D, video renderings than there are "pro" gamers.  I'm not counting kids who think they are "pg".  If they have rich parents they will have whatever they like.   That's why I've mentioned that Threadripper is not a bad platform. Not by a longshot. I'm here talking about a possibility of another route. And that is a dual Epyc socket one. I honestly want to hear the big downsides.   

 

 

 

[betam4x]

On 8/16/2017 at 1:24 PM, Mad_Duke said:

Look, honestly. For solo, high FPS 1080P whatever gaming (eSports) I would go Intel.  But there are literally more people doing 3D, video renderings than there are "pro" gamers.  I'm not counting kids who think they are "pg".  If they have rich parents they will have whatever they like.   That's why I've mentioned that Threadripper is not a bad platform. Not by a longshot. I'm here talking about a possibility of another route. And that is a dual Epyc socket one. I honestly want to hear the big downsides.   

 

The issue with EPYC is quite honestly, performance.  EPYC is designed for the datacenter, so clock speeds are much lower.  This is done in order to meet the cooling requirements of a rack mounted system.  You can take all that cash for EPYC and throw it at Threadripper and you'll have a faster machine at the end of the day.  EPYC scales to a higher core count, so if you had the thousands needed for the top end offering, you COULD end up with a faster rendering machine over Threadripper, but for $999 (and eventually, street prices will come down), Threadripper is a damn good value.

 

AMD chips are extremely competitive with Intel chips when it comes to gaming.  I think that AMD shot themselves in the foot with the memory issues.  Most reviews came back showing a massive disadvantage for gaming, but later on, once the memory issues were straightened out, this is no longer the case.  A Ryzen 7 1800X ($369 at my local Microcenter) with DDR4 3200 either beats the 7700k or is within 5% of it in every benchmark I've seen with the exception of Rocket League (which is likely due to a bug).  Threadripper in game mode completely stomps Intel's offerings, even when overclocked.  Check out Anandtech bench for example, the results will surprise you.

 

Keep in mind, this is coming from a guy who doesn't currently own a single AMD part.  I am looking at purchasing an AMD CPU for the first time in over a decade.  I do consulting, anything from software development in a myriad of languages to custom graphics and videos, and in my free time?  I game.  Due to my requirements. I am going to pick up a 1950x as soon as demand slows down and you can find the products in stock.  The value/performance is an absolute steal for my use case.