You Want This. - HOLY $H!T AMD Threadripper Pro 5995WX

[AdamFromLTT]

Before the launch of Ryzen 7000, AMD has slipped in one more Zen 3 product launch. Or at least, they are finally letting consumers buy it. The AMD Threadripper Pro 5995WX is a beast of a CPU that likely has Intel quaking in their boots. But with a massive price leap and the death of the non-pro lineup of Threadripper CPUs, has AMD turned to the greedy dark side? Of course they have, they’re a corporation.

 

 

[dizmo]

64 cores? And here I am struggling to make use of 12.

[Lightwreather]

I mean I would, realistically I'd run out of ram before the processor even decides to be half-used.

[sub68]

linus costs how much random crap, me oh a canon 1dx series, R3, nikion Z9, sony A1 (bassicly olympic camera) pricing. 

[John Boy]

Well such a Workstation you only buy if its really nessecary.
Because if you are a company which spent $18k on PC then they usually have a Cluster or access to a Supercomputer. Means that you usually dont run the Solver (for FEM) or your Rendering directly on the mashine. Usually you work, like LTT directly on a server and send then the Data to the Cluster or Rendering Farm.

What he also didnt mention Software cost. Because some Companys have a dramatic price increase, for thier Software if you use more then 32 Cores. Thats one of the main reason why the next Epyc gen has three 32 core variants but only one 64 core variant.
And dont underestimate Software Costs, against those, even my $10k Workstation (Intel) at work, is cheap.

[eviatar]

Hey Linus, what water cooling is being used in this demo?
It looks like an AIO and I'm desperately trying to find an AIO for my 3970x which covers the entire IHS. (Enermax cr@p is not an option)

 

TIA 

[Luscious]

When I splurged on my 8-core 5960X back in late 2014 I felt pretty good - not only was the performance awesome but I knew it would be a VERY long time before 8 cores would "feel" inadequate. For what I use it for (multimedia, gaming, multitasking) it's still incredibly capable, although I can see where newer chips would beat it in benchmarks. And that's just it. If it takes benchmarks to tell the difference and show that newer is faster, especially if it's marginal and only in certain workloads, is the new stuff really that much better and worth the cost/time? Upgrading has to be attractive for me to seriously go for it, which is why I only recently went from four 980TI's to four 3080Ti's.

 

That said, I've been eyeing the TR Pro lineup for a future build, but not the 64 core part. My pick is the 24 core 5965WX. It's higher base clock matches my 5960X but with 3X the core count and supposedly more IPC. That should translate to a tangible performance increase that's noticeable without having to resort to benchmarks. I agree though, overclocking should be a feature that's turned on out of the box, and if it's a mobo issue then guys like ASUS need to get their act together. Intel abandoned HEDT some years ago so it would seem logical that TR Pro, especially an overclockable lineup, would be the new HEDT game in town. These chips aren't just great for workstation use.

 

Besides, I can feel good at the end of the day knowing that I went from a 5960 to a 5965 

[dogwitch]

i am keeping a eye on it.

 

 

fyi if anyone watch the stream with linus ref TR builds that start to over heat.

its a power usage setting bug in windows. That seemly is to become more widespread. Only way to fix that is to set a costum curve for power.

Same with the news feed(memory leak bug). to fix that. fully disable it for now.

[hishnash]

I think intel more or less did not bother with putting much effort into the workstation market since they just sort of assumed most workstation uses wanted TB and until recently this was very much intel excessive.  For many workstation users it did not really matter if AMD shipped a 2x or even 5x faster system if they were still going to be waiting 10x the time to get data onto and off the system (workstation being the thing you have under your desk not rack mounted in the server room).   Yes there were some strange attempts by some mother board vendors to sort of support TB on older TR it was always seems as a little bit of a hack and thus not applicable to most of the market.

What will be really interesting to see is how these TR pro chips compare to the macPro once apple updates it, but then again they might well be targeting parallel but seperate subsections of the market.  

[papajo]

a) in europe you can already buy them since a few months wtf is linux talking about waiting to see them on retail? 

b) I am so disappointed that he did not follow up and run a cpu rendered crysis bench to see if this one could be able to play it at an actually decent frame rate..

Since his past video playing cpu rendered crysis withan Threadripper 3990X   I ALWAYS wondered (and looked forward to) for him to test the 5000 threadrippers that didnt exist at that time ....
 

 

\

 

Why you failed me again linus? 




P.S title of the video is kinda clickbaitty I mean it gave me the impression that there was to be a giveaway raffle/competition of some sort for this lol 

[lochlanna]

The reason that the Mozilla compile benchmark wasn't so impressive is probably due to the linker. When compiling C/C++ (and more) languages the code is compiled into small objects and then linked together. Most current linkers are predominantly single threaded.

There is a new somewhat experimental linker (linux/Mac OS only for now) called Mold which does a much much better job at taking advantage of all your cores and is much much faster than the alternatives.

 

https://github.com/rui314/mold

This CPU would be an absolute beast for some software developers. At a previous job if I had to do a fresh compile of our entire code base it was about 45 mins even though the workstation had 2x6 core Xeon cpus. A processor like this would have made my life A LOT better.

[Motifator]

I'd need to take out my whole previous water cooling gear so... no, I don't want that. That's kind of clickbait, lol.

[tim0901]

3 hours ago, lochlanna said:

The reason that the Mozilla compile benchmark wasn't so impressive is probably due to the linker. When compiling C/C++ (and more) languages the code is compiled into small objects and then linked together. Most current linkers are predominantly single threaded.

There is a new somewhat experimental linker (linux/Mac OS only for now) called Mold which does a much much better job at taking advantage of all your cores and is much much faster than the alternatives.

I doubt it. Look at the graphs on Mold's webpage - the amount of time even a 'traditional' linker takes to join everything together is only ~30 seconds when compiling Firefox, compared to the ~10 minutes the entire compilation process takes to complete. It doesn't make sense for it to be a major bottleneck in the chain here. Chances are it's a limitation of the Firefox build system to not be able to take full advantage of so many threads.

 

Personally i'd be interested to see if the results are different on Linux, since Windows is known to slow down C++ compiles a fair bit due to things like I/O bottlenecks.

[AdamFromLTT]

13 hours ago, lochlanna said:

The reason that the Mozilla compile benchmark wasn't so impressive is probably due to the linker. When compiling C/C++ (and more) languages the code is compiled into small objects and then linked together. Most current linkers are predominantly single threaded.

There is a new somewhat experimental linker (linux/Mac OS only for now) called Mold which does a much much better job at taking advantage of all your cores and is much much faster than the alternatives.

 

https://github.com/rui314/mold

This CPU would be an absolute beast for some software developers. At a previous job if I had to do a fresh compile of our entire code base it was about 45 mins even though the workstation had 2x6 core Xeon cpus. A processor like this would have made my life A LOT better.

Sadly, we didn't have enough time to sink our teeth into more compiling benchmarks. 

In other compile benchmarks that we are still in the process of standardizing, like Chromium compile, this thing would RIP. I've also seen that Unreal Engine compiles tend to scale well across cores, but we haven't looked into it much yet. Our benchmarking standards are evolving, especially with the new lab on the horizon, so we'll likely have more testing in future CPU coverage.

[AdamFromLTT]

21 hours ago, papajo said:

a) in europe you can already buy them since a few months wtf is linux talking about waiting to see them on retail? 

b) I am so disappointed that he did not follow up and run a cpu rendered crysis bench to see if this one could be able to play it at an actually decent frame rate..

Since his past video playing cpu rendered crysis withan Threadripper 3990X   I ALWAYS wondered (and looked forward to) for him to test the 5000 threadrippers that didnt exist at that time ....

 

 

 

Why you failed me again linus? 
 

A) we mention that these chips have been out for a few months, but only to OEMs. AMD only launched for retail consumers early this September, but this video got a bit delayed, sothe timeline is a bit wonky and we tried to clarify with on-screen annotations.

 

b) I'll do my best to remember this for our eventual coverage Zen 4 Epyc chips because it's a really fun demo.

[AdamFromLTT]

On 9/11/2022 at 11:58 AM, eviatar said:

Hey Linus, what water cooling is being used in this demo?
It looks like an AIO and I'm desperately trying to find an AIO for my 3970x which covers the entire IHS. (Enermax cr@p is not an option)

 

TIA 

It's a custom OEM cooler for SuperMicro, no retail version sadly.

[papajo]

34 minutes ago, AdamFromLTT said:

A) we mention that these chips have been out for a few months, but only to OEMs. AMD only launched for retail consumers early this September, but this video got a bit delayed, sothe timeline is a bit wonky and we tried to clarify with on-screen annotations.

No m8 I meant for retail sales, they have been available in long before September like I think June? not sure but they didnt come recently on the stores here in the EU that's for sure  e.g https://www.caseking.de/amd-ryzen-threadripper-pro-5995wx-2-7-ghz-chagall-pro-sockel-swrx8-tray-hpam-240.html?sPartner=185&utm_source=geizhals&utm_medium=comparison&utm_campaign=PC-Komponenten+>+CPUs+%2F+Prozessoren+>+AMD&campaign=psm/geizhals&wt_mc=preisvergleich.geizhals.feed&emid=631fb701f77a5f054230f351 

https://www.caseking.de/amd-ryzen-threadripper-pro-5995wx-2-7-ghz-chagall-pro-sockel-swrx8-tray-hpam-240.html?sPartner=185&utm_source=geizhals&utm_medium=comparison&utm_campaign=PC-Komponenten+>+CPUs+%2F+Prozessoren+>+AMD&campaign=psm/geizhals&wt_mc=preisvergleich.geizhals.feed&emid=631fb701f77a5f054230f351

In Germany 


 

 

38 minutes ago, AdamFromLTT said:

 

b) I'll do my best to remember this for our eventual coverage Zen 4 Epyc chips because it's a really fun demo.

In case you do remember please be so kind to pair it very fast low latency octachannel ram because I think it will make a very noticeable difference given that it would play the role of vram in this scenario, also OC it a little if possible  

[lochlanna]

10 hours ago, tim0901 said:

I doubt it. Look at the graphs on Mold's webpage - the amount of time even a 'traditional' linker takes to join everything together is only ~30 seconds when compiling Firefox, compared to the ~10 minutes the entire compilation process takes to complete. It doesn't make sense for it to be a major bottleneck in the chain here. Chances are it's a limitation of the Firefox build system to not be able to take full advantage of so many threads.

 

Personally i'd be interested to see if the results are different on Linux, since Windows is known to slow down C++ compiles a fair bit due to things like I/O bottlenecks.

Now that I think about it a little more, you're totally right. I've seen linking take minutes on some projects like Blender, but Firefox doesn't seem quite as heavy. The build system (MSVC especially) is more likely to be the culprit. I haven't looked at the Mozilla project, but I wonder if it's compatible with the Ninja build system. It tends to do a considerably better job at keeping the cores fed.

 

[AdamFromLTT]

1 hour ago, papajo said:

No m8 I meant for retail sales, they have been available in long before September like I think June? not sure but they didnt come recently on the stores here in the EU that's for sure  e.g https://www.caseking.de/amd-ryzen-threadripper-pro-5995wx-2-7-ghz-chagall-pro-sockel-swrx8-tray-hpam-240.html?sPartner=185&utm_source=geizhals&utm_medium=comparison&utm_campaign=PC-Komponenten+>+CPUs+%2F+Prozessoren+>+AMD&campaign=psm/geizhals&wt_mc=preisvergleich.geizhals.feed&emid=631fb701f77a5f054230f351 

https://www.caseking.de/amd-ryzen-threadripper-pro-5995wx-2-7-ghz-chagall-pro-sockel-swrx8-tray-hpam-240.html?sPartner=185&utm_source=geizhals&utm_medium=comparison&utm_campaign=PC-Komponenten+>+CPUs+%2F+Prozessoren+>+AMD&campaign=psm/geizhals&wt_mc=preisvergleich.geizhals.feed&emid=631fb701f77a5f054230f351

In Germany 


 

 

In case you do remember please be so kind to pair it very fast low latency octachannel ram because I think it will make a very noticeable difference given that it would play the role of vram in this scenario, also OC it a little if possible  

Funny, in AMD's own press release from June https://community.amd.com/t5/business/design-build-accelerate-on-the-ultimate-workstation-processor/ba-p/530661

They said that it was going out to SI's in July and then going to retail at some point(our video was delayed a few weeks). There wasn't a single outlet that was able to get their hands on one before August, and everyone who did got one from an SI like SuperMicro or Puget Systems. But I don't think any chips were available until late August, at least not in NA. so maybe the listing was posted in advance?

[IanCutress]

On 9/12/2022 at 3:17 AM, papajo said:

b) I am so disappointed that he did not follow up and run a cpu rendered crysis bench to see if this one could be able to play it at an actually decent frame rate..

 

14 hours ago, AdamFromLTT said:

b) I'll do my best to remember this for our eventual coverage Zen 4 Epyc chips because it's a really fun demo.

 

Linus used my 5995WX review video at 11:26 so I have no qualms about posting the fact that my video review does have CPU-Rendered Crysis as one of the benchmarks. I've been running CPU-Rendered Crysis on every CPU for a long while. I've created a script to make it work, because it has some really weird config limitations. But here's a link. Timestamp 14:58. 
 

 

[papajo]

6 hours ago, IanCutress said:

 

 

Linus used my 5995WX review video at 11:26 so I have no qualms about posting the fact that my video review does have CPU-Rendered Crysis as one of the benchmarks. I've been running CPU-Rendered Crysis on every CPU for a long while. I've created a script to make it work, because it has some really weird config limitations. But here's a link. Timestamp 14:58. 
 

 

Interesting! I wasnt aware of that but maybe linus has the connections amd street cred to convince Crytek to release an update on the renderer that can take advantage of more than 23 core   I mean I doubt that they need to rework the entire engine to enable more cores so I guess its feasible. 

[leadeater]

On 9/12/2022 at 11:06 AM, hishnash said:

I think intel more or less did not bother with putting much effort into the workstation market since they just sort of assumed most workstation uses wanted TB and until recently this was very much intel excessive.  For many workstation users it did not really matter if AMD shipped a 2x or even 5x faster system if they were still going to be waiting 10x the time to get data onto and off the system (workstation being the thing you have under your desk not rack mounted in the server room).   Yes there were some strange attempts by some mother board vendors to sort of support TB on older TR it was always seems as a little bit of a hack and thus not applicable to most of the market.

TB isn't a big deal on proper workstations, in fact most couldn't care about it in the slightest. You'd have 10Gb network connection at a minimum connect to a backend server which TB gives absolutely zero benefit to since you'd have to connect a 10Gb TB dongle to regardless because you are never going to be connecting directly to anything via TB in these scenarios.

 

TB is only useful for near by, same room connectivity to local devices. It's like comparing Bluetooth to Wireless Ethernet, solutions for different needs and purposes.

 

TB on Apple workstations is only important because of Apple and Apple centric accessories which aren't as widely used or wanted over on Windows/Linux workstation land.

 

I'd take a 25Gb NIC over TB on a workstation any day, so long as there is actually a 25Gb network port to plug in to of course.

 

P.S. TB on HEDT and Xeon workstation wasn't even a thing for ages while it was on consumer.

[papajo]

On 9/12/2022 at 2:06 AM, hishnash said:

I think intel more or less did not bother with putting much effort into the workstation market since they just sort of assumed most workstation uses wanted TB and until recently this was very much intel excessive.  For many workstation users it did not really matter if AMD shipped a 2x or even 5x faster system if they were still going to be waiting 10x the time to get data onto and off the system (workstation being the thing you have under your desk not rack mounted in the server room).   Yes there were some strange attempts by some mother board vendors to sort of support TB on older TR it was always seems as a little bit of a hack and thus not applicable to most of the market.

What will be really interesting to see is how these TR pro chips compare to the macPro once apple updates it, but then again they might well be targeting parallel but seperate subsections of the market.  

TIL workstations use TB as their main pipe for transferring data, lol I always assumed TB was a consumer thing to daisychain stuff or to connect an external graphics card to your laptop etc. 

Btw as far as I know Threadripper 5000 series support both thunderbolt 4 and USB 4 

The only somewhat disappointment (related to workstation stuff) I have with AMD is their DASH implementation, it exists but its more clunky and motherboard specific compared to Intel's vPro AMT  for lights out remote management (which also is far from ideal and a pain in the a* to setup!!! but comparatively to AMD its easier to implement also most intel CPUs support it but only pro series and TR or Epyc series support DASH on the amd side of things)