Geekzone: technology news, blogs, forums
Guest
Welcome Guest.
You haven't logged in yet. If you don't have an account you can register now.


View this topic in a long page with up to 500 replies per page Create new topic
1 | 2 
timmmay
20858 posts

Uber Geek
+1 received by user: 5350

Trusted
Lifetime subscriber

  #2503675 12-Jun-2020 14:54
Send private message

Most people probably aren't using near that amount of bandwidth most of the time. 3D rendering, sure, but I guess the high end parts are meant for people really pushing their machines and the consumer parts are for people doing web browsing and word processing. AMD looks to be better for enthusiasts and power users if single core performance isn't critical.




Handle9
11924 posts

Uber Geek
+1 received by user: 9675

Trusted
Lifetime subscriber

  #2503681 12-Jun-2020 15:07
Send private message

arcon:

toejam316: If you want more than 8x2 pcie lanes for your gpu, you'll need a Xeon or Threadripper system. TR3960X is 24c/48t and starts at $1800.


I haven't found a graphics card that can saturate more than 8X, so 8 lanes for each card is fine. But I think that still means consumer CPUs don't have enough lanes... 3 NVME drives = 12 more lanes, +4 reserved for chipset link = 32 total :/



That's only true if you are doing simultaneous operations with all the nvme drives. Practically for consumer hardware there would be very few circumstances where even sata sdd performance is the bottleneck.

arcon

423 posts

Ultimate Geek
+1 received by user: 198


  #2503767 12-Jun-2020 17:06
Send private message

Handle9:
That's only true if you are doing simultaneous operations with all the nvme drives. Practically for consumer hardware there would be very few circumstances where even sata sdd performance is the bottleneck.

 

Sata SSD is much slower in general though, if I've got an After Effects comp with many layers of 10 bit image sequences, it needs to be able to scrub the timeline back & forward fluidly, but I guess the files would all be on one drive, so as long as that drive is uber fast it should be fine.

 

So that would be 24 CPU lanes total, (i.e. 2x8 graphics +4 NVMEs + 4 chipset). Does it automatically "smartly" assign the CPU lanes to the NVME in use or does one special slot get them?

 

 




Handle9
11924 posts

Uber Geek
+1 received by user: 9675

Trusted
Lifetime subscriber

  #2503768 12-Jun-2020 17:08
Send private message

arcon:

Handle9:
That's only true if you are doing simultaneous operations with all the nvme drives. Practically for consumer hardware there would be very few circumstances where even sata sdd performance is the bottleneck.


Sata SSD is much slower in general though, if I've got an After Effects comp with many layers of 10 bit image sequences, it needs to be able to scrub the timeline back & forward fluidly, but I guess the files would all be on one drive, so as long as that drive is uber fast it should be fine.


So that would be 24 CPU lanes total, (i.e. 2x8 graphics +4 NVMEs + 4 chipset).


 



You're not talking about a typical consumer use case. That's what pro hardware is for (and why it's $$$)

arcon

423 posts

Ultimate Geek
+1 received by user: 198


  #2503789 12-Jun-2020 17:36
Send private message

Handle9:

You're not talking about a typical consumer use case. That's what pro hardware is for (and why it's $$$)

 

I'm not sure I would need pro hardware providing all media reads were from the same drive... I can already scrub 2 layers of 10 bit on my 5 year old system reasonably well, and its an i4790K with 550MB speed sata SSD. I'm thinking its not much of a stretch that 6 x faster NVME could handle many more layers.


toejam316
1516 posts

Uber Geek
+1 received by user: 888

Trusted
Lifetime subscriber

  #2503791 12-Jun-2020 17:43
Send private message

In which case any Ryzen R5 or R7 CPU would probably suit you just fine - but wait for the next batch of them.





Join Quic Broadband with my referral - no sign up fee and gives me account credit

 

Anything I say is the ramblings of an ill informed, opinionated so-and-so, and not representative of any of my past, present or future employers, and is also probably best disregarded.


 
 
 

Support Geekzone with one-off or recurring donations Donate via PressPatron.
zenourn
281 posts

Ultimate Geek
+1 received by user: 168

ID Verified
Trusted
DR

  #2503818 12-Jun-2020 18:27
Send private message

arcon:

 

I haven't found a graphics card that can saturate more than 8X, so 8 lanes for each card is fine. But I think that still means consumer CPUs don't have enough lanes... 3 NVME drives = 12 more lanes, +4 reserved for chipset link = 32 total :/

 

 

I'd agree - during testing I set the Titans to 8X and the performance was around 1-5% lower compared to the where they both run at 16X.  This is for scientific/machine learning applications which are rather demanding in terms of bandwidth.

 

In your case a Ryzen 3000 Series (or even better wait for 4000 desktop series) I think would likely give you enough bandwidth for your requirements.

 

One note - The other benefit of Threadripper is that has 8 DIMM slots (supports 256GB memory). Just wish Threadripper motherboards had better support for EEC registered memory, but they don't want to cannibalise the market for the EPYC line.


zenourn
281 posts

Ultimate Geek
+1 received by user: 168

ID Verified
Trusted
DR

  #2503831 12-Jun-2020 19:03
Send private message

Again - there are caviates with Ryzen.
CPUs are made up of 'clusters' of 4 cores - each 4 cores has 1x PCIe controller built in - if you go too cheap and get less cores - you lose PCIe lanes. (And the motherboard compensates by splitting the few available to allow all your devices to function (but not at their full potential)

So make sure the AMD CPU you chose has enough core blicks' to provide enough PCIe controllers, to provide enough PCIe lanes directly connected to the CPU.
(& of course make sure the motherboard you are looking at allows these through 'fully' & not bottlenecked.)

 

I'm not sure this is correct. My understanding is that the 24 PCIe lanes come off the cIOD die and then can have one (when <=8 cores) or two CCDs (when 12/16 cores).

 

 

 


timmmay
20858 posts

Uber Geek
+1 received by user: 5350

Trusted
Lifetime subscriber

  #2503836 12-Jun-2020 19:34
Send private message

Get plenty of RAM, it's used as a disk cache and is faster than any disk.


arcon

423 posts

Ultimate Geek
+1 received by user: 198


  #2503973 13-Jun-2020 10:28
Send private message

zenourn:

 

(or even better wait for 4000 desktop series) I think would likely give you enough bandwidth for your requirements.

 

The new Xeons are also supposed to be launching at around the same time in Q3 this year (finally with 10nm and PCIe 4.0).

 

 


arcon

423 posts

Ultimate Geek
+1 received by user: 198


  #2505599 16-Jun-2020 09:42
Send private message

I only just found out the X series of i9s exist lol. The 10900-X would be perfect... comparable in single threaded performance to a Ryzen 9 3950X, but with double the PCIe lanes from CPU means there's basically no comparison for workstation multi-GPU / NVME rigs.


1 | 2 
View this topic in a long page with up to 500 replies per page Create new topic








Geekzone Live »

Try automatic live updates from Geekzone directly in your browser, without refreshing the page, with Geekzone Live now.



Are you subscribed to our RSS feed? You can download the latest headlines and summaries from our stories directly to your computer or smartphone by using a feed reader.