Register a SA Forums Account here!
JOINING THE SA FORUMS WILL REMOVE THIS BIG AD, THE ANNOYING UNDERLINED ADS, AND STUPID INTERSTITIAL ADS!!!

You can: log in, read the tech support FAQ, or request your lost password. This dumb message (and those ads) will appear on every screen until you register! Get rid of this crap by registering your own SA Forums Account and joining roughly 150,000 Goons, for the one-time price of $9.95! We charge money because it costs us money per month for bills, and since we don't believe in showing ads to our users, we try to make the money back through forum registrations.
 
  • Post
  • Reply
Kibner
Oct 21, 2008

Acguy Supremacy

Shaocaholica posted:

Is ‘480mb’ threadripper pro cache usable by all cores?

No, not with the chiplet design.

Adbot
ADBOT LOVES YOU

Bjork Bjowlob
Feb 23, 2006
yes that's very hot and i'll deal with it in the morning


Am I reading the bottom table correctly in that the TRX50 will support both non-Pro (X) and Pro (WX) processors i.e TRX50 and WRX90 are socket compatible?

Pablo Bluth
Sep 7, 2007

I've made a huge mistake.
I believe: you can use PRO in the hedt motherboard and it'll just be limited to the hedt memory & io capabilities. Non-PRO won't work in PRO motherboard.

Kibner
Oct 21, 2008

Acguy Supremacy

Bjork Bjowlob posted:

Am I reading the bottom table correctly in that the TRX50 will support both non-Pro (X) and Pro (WX) processors i.e TRX50 and WRX90 are socket compatible?

The non-Pro boards support both Pro and non-Pro CPUs.

The Pro boards only support Pro CPUs.

BurritoJustice
Oct 9, 2012

Shaocaholica posted:

Is ‘480mb’ threadripper pro cache usable by all cores?

It's cache coherent, so yes, but it is non-uniform due to the chiplet design. Accessing the cache of a separate CCD is about as fast as accessing main memory, so it's better expressed as 12x32MB L3.

Bjork Bjowlob
Feb 23, 2006
yes that's very hot and i'll deal with it in the morning


Kibner posted:

The non-Pro boards support both Pro and non-Pro CPUs.

The Pro boards only support Pro CPUs.

That's good news then - the processor incompatibilities between TR4, sTRX4 and sWRX8 hurt any idea of longterm investment in the threadripper platform.

BurritoJustice
Oct 9, 2012

Reusing the full size IO die with memory channels disabled has the funny side effect of each NUMA node in TR non-Pro only having a single memory channel. If you put a 7995WX in a TRX50 board you'll have three whole CCDs sharing a single local memory channel.

It's a good thing the NUMA penalty is relatively small.

kliras
Mar 27, 2021
noctua with an official statement on some cooler compatibility

https://twitter.com/Noctua_at/status/1714992102294057053

quote:

Vienna, October 19th 2023 – Noctua today announced its new NH-U14S TR5-SP6 and NH-D9 TR5-SP6 4U CPU coolers for AMD’s TR5 (sTR5/sTRX5/sWRX9) and SP6 sockets for the just-released 7000 series Threadripper HEDT (High-End Desktop) and Threadripper Pro workstation processors as well as 8004 series Epyc server CPUs. While the 14cm NH-U14S TR5-SP6 provides maximum cooling performance at minimum noise levels, the 9cm NH-D9 TR5-SP6 4U combines excellent efficiency with a compact size for full 4U compatibility. In addition to the two cooler models, Noctua also introduced the new NM-TR5-SP6 mounting kit that allows upgrading existing TR4-SP3, DX-4677, DX-4189 and DX-3647 line coolers to support the new TR5/SP6 sockets of TRX50 and WRX90 motherboards.

Klyith
Aug 3, 2007

GBS Pledge Week

BurritoJustice posted:

It's cache coherent, so yes, but it is non-uniform due to the chiplet design. Accessing the cache of a separate CCD is about as fast as accessing main memory, so it's better expressed as 12x32MB L3.

threadripper X3D with 1GB of total cache WHEN

Paul MaudDib
May 3, 2006

TEAM NVIDIA:
FORUM POLICE

Pablo Bluth posted:

The fact Threadripper vanilla is a cutdown Threadripper PRO which is a variant of Epyc should bode well for it. It must cut down on the R&D costs so it's more viable long term, while benefitting from all the quality control work that the Epyc market demands.

this was always true of threadripper, though? it's always been a cut-down epyc, and development costs have always been "minimal" and has always benefited from riding on epyc's coat-tails, but they still half-assed it anyway

Paul MaudDib
May 3, 2006

TEAM NVIDIA:
FORUM POLICE
$1500 isn't an increase at least, but it's still ultimately 4x R5 7600 on a package, with half the memory channels cut off, etc. That should be a $700 product, not $1500, but it's "what the market will bear". And then you'll have the $1500 motherboard for it.

And again, regardless of it being what AMD thinks the market will bear, the fact is that the epyc boards and chips are literally better and cheaper for most people, so this is another "waste of silicon" launch. the server market "won't bear" $1500 motherboards and produces very good 8-channel motherboards kitted out with PCIe lanes and poo poo for $700, but then we apply the gamer tax and threadripper is $1500 for a board with much less capability. Chips are cheaper too, once the used market gets into gear (which is of course why they've gone after that with platform lock/etc)

https://www.asrockrack.com/general/productdetail.asp?Model=GENOAD8X-2T/BCM#Specifications

in the server market, these are $931 on provantage right now (the predecessor on SP3, ROMED8-2T, was $600). AMD wants you to pay $1500 for the threadrippper-lite or $2-3k for the full threadripper equivalent. it's just pretty much a non-starter if you care about value

Paul MaudDib fucked around with this message at 16:58 on Oct 19, 2023

Shaocaholica
Oct 29, 2002

Fig. 5E
So X3D cache is accessible by all cores with the same penalty and its slower than L3 but orders of magnitude faster than main memory? And X3D is not available on Epic/Threadripper...yet but there are plans?

Twerk from Home
Jan 17, 2009

This avatar brought to you by the 'save our dead gay forums' foundation.

Shaocaholica posted:

So X3D cache is accessible by all cores with the same penalty and its slower than L3 but orders of magnitude faster than main memory? And X3D is not available on Epic/Threadripper...yet but there are plans?

X3D has been available on Epyc for a whilez released March 2022: https://www.anandtech.com/show/17323/amd-releases-milan-x-cpus-with-3d-vcache-epyc-7003

The 3D cache is part of L3 and it makes the whole L3 a little bit slower than a non-3D part, but usually having all that extra cache helps way more than the tiny hurt you get from it being slower.

Pablo Bluth
Sep 7, 2007

I've made a huge mistake.

Paul MaudDib posted:

this was always true of threadripper, though? it's always been a cut-down epyc, and development costs have always been "minimal" and has always benefited from riding on epyc's coat-tails, but they still half-assed it anyway
I always got the impression that original run of Threadripper chips were very much a third product line, neither epyc or ryzen. Whereas Threadripper PRO are much closer to just being part of the Epyc SKU, put in desktop motherboards and rebranded.

Combat Pretzel
Jun 23, 2004

No, seriously... what kurds?!
RDIMM lol.

Subjunctive
Sep 12, 2006

✨sparkle and shine✨

Jealous that they can fill all their DIMM slots, unlike my 7800X3D!

Klyith
Aug 3, 2007

GBS Pledge Week

Shaocaholica posted:

So X3D cache is accessible by all cores with the same penalty and its slower than L3 but orders of magnitude faster than main memory? And X3D is not available on Epic/Threadripper...yet but there are plans?

X3D cache is accessible by all cores on that die / chiplet at L3 speeds (an order of magnitude faster than main memory). On a 5800X3d or 7800X3D there is just the one chiplet, so yes. It is fundamentally the same L3, just more of it.


On multi-chiplet CPUs (79_0X3D, Epyc) accessing the L3 of a different chiplet is not L3 speed, because the communication has to go through the central IO die. This has latency ranging from 19ns (ram speed) to over 30ns (worse than memory).

This is why the 7950X3D is really not the best gaming CPU: being aware of non-uniform access speeds is important and video games just don't do that because it's hard. Serious high-performance software does though. So people with 7950X3Ds who do games sometimes have to resort to manually fencing in the process to one set of cores.



to be clear, my threadripper X3D post was a joke

SlapActionJackson
Jul 27, 2006

You ain't getting to DRAM in 19ns.

Pablo Bluth
Sep 7, 2007

I've made a huge mistake.
Threadripper x3d will appear in six months. You don't blow your load in one go.

Klyith
Aug 3, 2007

GBS Pledge Week
Hmm, I may be comparing apples and oranges between a real CPU test of real latency showing 20-30ns and the supposed 16-18ns response speed of DDR memory.

Cygni
Nov 12, 2005

raring to post

These are really niche products imo, kinda like the last generation. Especially with the boards with only 4 memory slots. The first few TR generations had some more mainstream DIY appeal, but I don't really see these in the same boat.

e: just to be fair, the entire HEDT/Workstation market is niche in general, especially these days. The Sapphire Rapids Xeon W is just as, if not more, niche.

Only registered members can see post attachments!

Cygni fucked around with this message at 23:08 on Oct 19, 2023

Shaocaholica
Oct 29, 2002

Fig. 5E
My next workstation at work better have a 128core CPU (currently 5995WX 64c)

Tuna-Fish
Sep 13, 2017

Cygni posted:

These are really niche products imo, kinda like the last generation. Especially with the boards with only 4 memory slots. The first few TR generations had some more mainstream DIY appeal, but I don't really see these in the same boat.
The reason to buy a non-pro TR is that you need more PCIe slots than the AM5 platform provides.

Methylethylaldehyde
Oct 23, 2004

BAKA BAKA

Tuna-Fish posted:

The reason to buy a non-pro TR is that you need more PCIe slots than the AM5 platform provides.

Pretty much. It's great for doing things that need a lot of GPU compute AND a lot of fast local storage. Granted, there aren't many reasons you'd need a pair of 4090s AND 90TB of SSDs hanging off the same PCIe bus, but it's doable at least!

Cygni
Nov 12, 2005

raring to post

Tuna-Fish posted:

The reason to buy a non-pro TR is that you need more PCIe slots than the AM5 platform provides.

Right, but the vast majority of people who fit that need and have cash to spare are going to be using cloud/rack mount solutions, not wanting to sit next to it all day. And you likely can get that cheaper (and more performant) with an Epyc DIY box, as Paul pointed out. Also the TRX50 board weve seen so far has a grand total of 2 PCIe 5.0 x16 slots. TRX50 "only" (lol) has 48 PCIe 5.0 lanes and needs some 5.0 M.2 slots, so some slots are gonna have to be 4.0 which probably isnt a huge problem for people, but its still there.

So what youre left with for TR non-pro is people who have lots of money who either have datasets that need big parallel CPU processing but cant be accelerated and doesnt use a lot of memory bandwidth, or people that want to run lots of PCIe accelerators or PCIe storage... who also cant/wont just run that in a datacenter for cheaper for whatever reason. It just seems like a small niche, but thats the workstation/HEDT world these days. I'm sure some exist, and good for them those wacky punks, but not gonna be a lot of em.

Canned Sunshine
Nov 20, 2005

CAUTION: POST QUALITY UNDER CONSTRUCTION



If I had the money I would buy one for CFD analysis, since this would be a perfect application for it.

Desuwa
Jun 2, 2011

I'm telling my mommy. That pubbie doesn't do video games right!
While the new threadrippers don't look completely horrible they're a year late - and there's still the chance of them releasing threadripper X3D in six months. That and them actively loving over threadripper owners before multiple times, with them allowing everyone to believe TR4 would be supported like AM4 and then only giving sTRX4 a single generation despite promises otherwise. The one thing that'd really get me interested is the amount of PCIe lanes, but even that seems unnecessarily exacerbated by the horrible slot layouts of AM5 boards. 24 pcie lanes isn't great but if you allow dropping to 8 lanes to each GPU plus the chipset providing some, there should be plenty for nvme and additional pcie cards, but the AM5 motherboards just suck.

The non-PRO prices don't look nearly good enough considering you cannot count on AMD giving it more than one generation of support. If I were to have gone HEDT this generation, I'd have gone the Intel route months ago and have gotten better ddr5 speeds out of it.

Kivi
Aug 1, 2006
I care
That's pretty board and I'd get that just to have black PCB.

quote:

in the server market, these are $931 on provantage right now (the predecessor on SP3, ROMED8-2T, was $600). AMD wants you to pay $1500 for the threadrippper-lite or $2-3k for the full threadripper equivalent. it's just pretty much a non-starter if you care about value
Speaking of which, you can get the silly SP3 mITX board for cheap now on Provantage: https://www.provantage.com/asrock-romed4id-2t~7ASRI0H2.htm

In theory that board will fit NR200p :getin:

Now if any of you US goons were willing to do a customs fraud and buy and send me a https://www.newegg.com/asrock-rack-romed8-2t/p/N82E16813140044 as that's also the lowest that board has been.

BlankSystemDaemon
Mar 13, 2009



Klyith posted:

X3D cache is accessible by all cores on that die / chiplet at L3 speeds (an order of magnitude faster than main memory). On a 5800X3d or 7800X3D there is just the one chiplet, so yes. It is fundamentally the same L3, just more of it.


On multi-chiplet CPUs (79_0X3D, Epyc) accessing the L3 of a different chiplet is not L3 speed, because the communication has to go through the central IO die. This has latency ranging from 19ns (ram speed) to over 30ns (worse than memory).

This is why the 7950X3D is really not the best gaming CPU: being aware of non-uniform access speeds is important and video games just don't do that because it's hard. Serious high-performance software does though. So people with 7950X3Ds who do games sometimes have to resort to manually fencing in the process to one set of cores.



to be clear, my threadripper X3D post was a joke
NUCA is a loving mess.

I love how CPU designers looked at the issues inherit with NUMA and thought "hey, let's do that, but make it even worse".
NUMA works decently if you're bottlenecked by thread scale-out and aren't doing certain memory sensitive workloads, but it's a loving nightmare to deal with when you're working with any kind of I/O for either auxiliary storage or networking.
NUCA has all of those issues, but now it's for any workload.

BlankSystemDaemon fucked around with this message at 10:18 on Oct 20, 2023

Truga
May 4, 2014
Lipstick Apathy
it's super easy to just have process groups run on a single chiplet though, and if you spent $1000 or whatever on your cpu, surely you know how to do that

BlankSystemDaemon
Mar 13, 2009



Truga posted:

it's super easy to just have process groups run on a single chiplet though, and if you spent $1000 or whatever on your cpu, surely you know how to do that
Sure, you can do cpusets/pinning/affinity, but then you're not using all those threads you bought the CPU to have access to, so what's the point?

Truga
May 4, 2014
Lipstick Apathy
you still have access too all the threads, though? you just can't run the cache sensitive threads on the low-cache cores, and since that post was talking about games, there's no game that needs both the cache and 16 threads to run well so far

i mean poo poo, i'm still on a 3950x which has l3 cache only shared by 2 cores and it runs 8 thread games fine because even with going through the interconnect the cache is still tons faster than accessing ram

Pablo Bluth
Sep 7, 2007

I've made a huge mistake.

Desuwa posted:

The non-PRO prices don't look nearly good enough considering you cannot count on AMD giving it more than one generation of support. If I were to have gone HEDT this generation, I'd have gone the Intel route months ago and have gotten better ddr5 speeds out of it.
There's a valid concern that they might hold off on releasing more generations due to an insufficient market to justify it, but as long as they do, the socket should be tied in to the Epyc socket cycle. So far the switch to DDR5 is the only time Eypc got a new socket. Nearly all evidence is that AMD philosophy is still multi generation sockets, and the exceptions coincided with a new memory type and a broken global supply chain.

Combat Pretzel
Jun 23, 2004

No, seriously... what kurds?!
HEDT sockets have been a clusterfuck.

Shaocaholica
Oct 29, 2002

Fig. 5E
So chiplets are the new sockets. These new 1P systems are really just 4P logically. I always wanted a 4P.

BlankSystemDaemon
Mar 13, 2009



Shaocaholica posted:

So chiplets are the new sockets. These new 1P systems are really just 4P logically. I always wanted a 4P.
8S is the way to go.

Shaocaholica
Oct 29, 2002

Fig. 5E
So new desktop APUs are coming with X3D right?

FuturePastNow
May 19, 2014


On the APUs? Probably not.

Dr. Video Games 0031
Jul 17, 2004

The new AGESA version supposedly has support for phoenix APUs, so those are presumably coming to desktop soon. They probably won't be X3D, and even if they were, the cache wouldn't be on the GPU.

Adbot
ADBOT LOVES YOU

Shaocaholica
Oct 29, 2002

Fig. 5E
I just want a dumb 500W USFF APU system but no half measures.

  • 1
  • 2
  • 3
  • 4
  • 5
  • Post
  • Reply