Threadripper: The inside story

[Dylanc1500]

3 minutes ago, samcool55 said:

2 things.

1) it's not a coincidence the amount of cores is just a bit over the maximum of threadripper (16 vs 18)

2) The die of Skylake-X only has 12c, the only die Intel has that's last-gen and more than 12 cores is their server stuff so they are going to use that for the top-end of the X299 platform.

 

And at launch there was almost no information about those cpu's. We knew they would be launched later and the core count but that's it, no core speeds, TDP, nothing.

18 cores was done because of the different silicon dies (10,18, 28). I don't know what you mean by 12 core, if you could explain that a little better, I apologize.

 

like I had said it was released earlier than planned, they had plans to release but not when they  ended up. I can't really go to in depth with that.

[goodtofufriday]

5 minutes ago, mynameisjuan said:

First of all you bring price to performance in to the equation which is not my point. Then you compare two different performance brackets, compare core count and totally ignore the processors that intel has that will demolish AMD's Epyc line. Ignoring price, intel is king in any use case scenario 

 

2 minutes ago, Dylanc1500 said:

That's very dependent on the workload, clock speed can have a great effect on multicore workloads. (Example of a workload I'm personally familiar with) If you have a database running transaction records, which can saturate ever core you give it, you increase the clock speed and every core can finish its individual workload quicker and move on to the next task.

Windows licenses its server software at a per core basis. When buying multiple servers this makes TR and Epyc a better buy. And heat output matters a lot. I am not looking to put multiple 90C servers into my server room, or even at a work bench. Tempuratures matter a lot as you add more computers/servers to a room even with efficient cooling. And no one in the work station or server space is looking to OC either. We need as stable as possible machines. No matter how stable an OC may seem, for most use cases its just not a viable long term option. 

 

The only computers staying on intel going forward are my CAD computers, as they are still largely single core applications, for some reason.

[WereCat]

6 minutes ago, Dylanc1500 said:

That's very dependent on the workload, clock speed can have a great effect on multicore workloads. (Example of a workload I'm personally familiar with) If you have a database running transaction records, which can saturate ever core you give it, you increase the clock speed and every core can finish its individual workload quicker and move on to the next task.

But wouldn't in that case more cores make as much sense as high clock speeds?

You finish X amount of tasks fast and then move on Y amount of tasks or you finish X+n amount of tasks slower and then you move on Y+n amount of tasks.

[TOMPPIX]

3 minutes ago, Dylanc1500 said:

like I had said it was released earlier than planned, they had plans to release but not when they ended up. I can't really go to in depth with that.

got a source on that?

[mynameisjuan]

1 minute ago, goodtofufriday said:

And no one in the work station or server space is looking to OC either.

Where did I say OC? I am talking pure stock speeds on Intel vs AMD, both top tier, intel wins hands down. And apparently cores only matter up to 18 or are we purposely forgetting intel's 24 core xeon?

[mynameisjuan]

3 minutes ago, WereCat said:

But wouldn't in that case more cores make as much sense as high clock speeds?

You finish X amount of tasks fast and then move on Y amount of tasks or you finish X+n amount of tasks slower and then you move on Y+n amount of tasks.

Well yeah....But still IPC is more important in general because programs still dont utilize all core 99.9999% of the time.

 

But if it comes to IPC vs cores, intel still wins both. 

[Dylanc1500]

1 minute ago, WereCat said:

But wouldn't in that case more cores make as much sense as high clock speeds?

You finish X amount of tasks fast and then move on Y amount of tasks or you finish X+n amount of tasks slower and then you move on Y+n amount of tasks.

 

3 minutes ago, goodtofufriday said:

 

Windows licenses its server software at a per core basis. When buying multiple servers this makes TR and Epyc a better buy. And heat output matters a lot. I am not looking to put multiple 90C servers into my server room, or even at a work bench. Tempuratures matter a lot as you add more computers/servers to a room even with efficient cooling. And no one in the work station or server space is looking to OC either. We need as stable as possible machines. No matter how stable an OC may seem, for most use cases its just not a viable long term option. 

 

The only computers staying on intel going forward are my CAD computers, as they are still largely single core applications, for some reason.

okey doaky. I will respond in an hour or two as I can't give a worthwhile response on my phone, and I'll be back at my desk. I shall be back!!!

[WereCat]

1 minute ago, mynameisjuan said:

Where did I say OC? I am talking pure stock speeds on Intel vs AMD, both top tier, intel wins hands down. And apparently cores only matter up to 18 or are we purposely forgetting intel's 24 core xeon?

Or 32 core EPYC which is clocked higher than 24 core Xeon?

[Dylanc1500]

3 minutes ago, TOMPPIX said:

got a source on that?

Not legally.

[DildorTheDecent]

1 hour ago, WereCat said:

Under water at 90*C with at least twice the price.

Actual custom loop or shitty aluminium AIO? 

 

240mm rad or 1000mm+ rad ?

 

What do you mean exactly? 

[mynameisjuan]

7 minutes ago, WereCat said:

Or 32 core EPYC which is clocked higher than 24 core Xeon?

Haha well I put my foot right in my mouth. Completely forgot about that chip *facepalm

[TVwazhere]

12 minutes ago, goodtofufriday said:

The only computers staying on intel going forward are my CAD computers, as they are still largely single core applications, for some reason.

Because companies like Autodesk dont want to reinvest rebuilding their code from the ground up to include multiple cores in the part modeling, assembly, construction (revit), etc, which from what I've been told, is what it would take for programs like Inventor (and solidworks I think) to use multi cores when doing routine functions.

[goodtofufriday]

5 minutes ago, mynameisjuan said:

Where did I say OC? I am talking pure stock speeds on Intel vs AMD, both top tier, intel wins hands down. And apparently cores only matter up to 18 or are we purposely forgetting intel's 24 core xeon?

I'm not forgetting xeon, which is why I mentioned the server space and epyc, and the cost per core on server licensing. For per core perf, without OC, intel and amd are more or less clock for clock from what I can tell when counting the boost/core ratio for both. 

Those claiming 4.8ghz boost on all cores is not a realistic figure for the enterprise space. 

[WereCat]

2 minutes ago, DildorTheDecent said:

Actual custom loop or shitty aluminium AIO? 

 

240mm rad or 1000mm+ rad ?

 

What do you mean exactly? 

It is unknown whether it was on AIO or custom. The peak voltage was 1.25V though

[DildorTheDecent]

Just now, WereCat said:

It is unknown whether it was on AIO or custom. The peak voltage was 1.25V though

Impressive volts for the clock. Just like Skylake and Kaby. 

 

Probably was an AIO though. Some old rattly H100i. Big boy processor needs big boy cooling.

[WereCat]

5 minutes ago, DildorTheDecent said:

Impressive volts for the clock. Just like Skylake and Kaby. 

 

Probably was an AIO though. Some old rattly H100i. Big boy processor needs big boy cooling.

The report also claims that it seems the IHS used solder because the temperatures are quite low considering the extremely high power draw. (BTW, you can find this on WCCF, IDK who tested it or how reliable the info is... some guy).

 

EDIT:

Considering that the CPU is 165W TDP at base 2.6GHz... I am quite sure that just the CPU was using at least 350W-400W at 4.8GHz.

[TOMPPIX]

5 minutes ago, WereCat said:

The report also claims that it seems the IHS used solder because the temperatures are quite low considering the extremely high power draw. (BTW, you can find this on WCCF, IDK who tested it or how reliable the info is... some guy).

http://www.guru3d.com/news-story/der8auer-delids-intel-12-andor-18-core-skylake-x-cpus.html

solder, nope.avi

[NumLock21]

No surprise there with AMD Threaripper

 

Mainstream. Intel Z270 vs AMD Ryzen

Enthusiast: Intel X299 vs AMD Threadripper

Server: Intel Xeon vs AMD Eypc

[DildorTheDecent]

Just now, WereCat said:

The report also claims that it seems the IHS used solder because the temperatures are quite low considering the extremely high power draw. (BTW, you can find this on WCCF, IDK who tested it... some guy).

Some random Korean done the testing. Navigating their forum is causing a new type of pain and suffering. 

 

Rig looks decent enough though. No word on actual rad size. 

Spoiler

 

[Tech_Dreamer]

Rumor has it they tied Raja kudori to as chair & pumped him full of steroids & LSD to make him come up with a great new CPU architecture .

[Nizkus]

2 minutes ago, Tech_Dreamer said:

Rumor has it they tied Raja kudori to as chair & pumped him full of steroids & LSD to make him come up with a great new CPU architecture .

B-but Raja Koduri is GPU guy

[Tech_Dreamer]

Just now, Nizkus said:

B-but Raja Koduri is GPU guy

exactly!

[Mira Yurizaki]

24 minutes ago, TVwazhere said:

Because companies like Autodesk dont want to reinvest rebuilding their code from the ground up to include multiple cores in the part modeling, assembly, construction (revit), etc, which from what I've been told, is what it would take for programs like Inventor (and solidworks I think) to use multi cores when doing routine functions.

Which I can see their point for two reasons:

• Multithreading is hard to get right.
• They're widely trusted across the industry, which probably means much of their core logic is very old and has a lot of confidence. So it's a case of "if it ain't broke, don't fix it"

[TVwazhere]

1 minute ago, M.Yurizaki said:

Which I can see their point for two reasons:

• Multithreading is hard to get right.
• They're widely trusted across the industry, which probably means much of their core logic is very old and has a lot of confidence. So it's a case of "if it ain't broke, don't fix it"

I definitely understand the reasoning and logic; it's not worth the resources, bug fixes and reputation as a stable software for the added performance of cutting a circle in a cube.

Still..... It'd be nice to cut that circle three tenths of a second faster (On a serious note, larger assemblies of over 500 parts)

[Energycore]

1 hour ago, Dylanc1500 said:

Intel isn't freaking out when they have 18 cores running at 4.8 ghz.

Do they though? Show me the available product that can do that right now.