ThreadRipper 1950X 16-core? 32-core?

[arisythila]

Hello,

Been awhile since I've posted here. There honestly hasn't been a bit of stuff going on in the industry that has been OMG amazing. Then Threadripper was released.

I've been doing some research and stuff, and ran across and article on WCCFTech. Thread Ripper comes with 4 die packages, each with 8 cores. AMD said that 2 of these packages were dummy cores. So the guys over there went and paid 1k for a retail package to find out if it was just their engineering sample that was mysterious.

http://wccftech.com/amd-threadripper...-epyc-32-core/

What they ended up finding out in all die packages were legitimate. The only problem is 2/4 are disabled.

Does anybody have any insight in this at all? Are we going to have some interesting CPU hacks to enable all 32 cores of the 1950? What are your thoughts?

Thanks,

Michael

[Charles Wirth]

WCCF went based on the data that from a source they seem to have not given credit to or ignored. A company paid for the CPU or AMD provided.

Marketing viral + special = make revenue off the video, no one was out $1000 drama bull or the CPU may have been dead already.

I am sure he has a FLIR camera, they are under $250 and a very useful tool. I would have taken thermal images and video of the post to see activity.

I would not be surprised that AMD has not learned from the past and still soft locks stuff in BIOS.

[arisythila]

Very well could just be a ploy to make revenue...

I think the bigger question is DID AMD learn not to lock stuff in BIOS, or what is going on here. Are the 4 die packages just quad core processors, or are they using 8 core die packages, and just disabling 1/2 the cores in each package. a FLIR would tell us quite a bit on this... If one part of the CPU is hotter than the other, or if its uniform across the 4 die packages.

I am thinking hard about picking one up here just to check this out.

Thanks,

Michael

[Charles Wirth]

Just noticed you are from Las Vegas,

I have a FLIRone camera if you want to do some images.

[chew*]

If you look at the difference from epyc to TR there are quite a few differences.

The first glaring difference is epyc is B2 stepping.

TR is B1 stepping.

So TR is epyc is bunk.

The second difference can be found comparing motherboards.

The dram channels.

Epyc cores each have access to its own channel as does tr...tr has access to less however.

Even if somehow AMD did enable 2 more cores it should limit or break the ability to use various memory modes due to lack of channels.

NUMA VS UMA.

The only way around this in the same space limitation of current TR boards is to limit each die to 2 slots ch A/B per die....this however would require a rewire of trace routing imo...

[Charles Wirth]

Or an Epyc mobo? Would a TR even fit?

For this test I would just be looking to see if there was any activity in the bum cores.

[arisythila]

Just noticed you are from Las Vegas,

I have a FLIRone camera if you want to do some images.

I just moved here last month, Building a house up north near Tule Springs, Actually Tule Springs is literally my back yard.

If I get one, I may take you up on that offer. Would be interesting to see if all 4 packages are getting warm, versus just 2 of them.

If you look at the difference from epyc to TR there are quite a few differences.

The first glaring difference is epyc is B2 stepping.

TR is B1 stepping.

So TR is epyc is bunk.

The second difference can be found comparing motherboards.

The dram channels.

Epyc cores each have access to its own channel as does tr...tr has access to less however.

Even if somehow AMD did enable 2 more cores it should limit or break the ability to use various memory modes due to lack of channels.

NUMA VS UMA.

The only way around this in the same space limitation of current TR boards is to limit each die to 2 slots ch A/B per die....this however would require a rewire of trace routing imo...

Yeah, That makes sense. It will be interesting to see how things pan out there. If this is essentially what EPYC will be like minus some memory channels or something. I don't know.

I just found the article to be interesting, and opened up some memories from the early Athlon days when you could unlock multipliers, and other things with this or that.

But if there are 4 8-Core packages on chip, and 2 of those are disabled. One could wonder why did they do that, and what would be the benefit of that. My thought is they could have taken 4 8-Core Ryzen chips, and disabled 1/2 of the cores, maybe they are doing some sort of binning or something. None the less a FLIR will show us this information, if heat is coming from all 4 packages versus just 2 of the packages. What are these "dummy" cores doing? Are they there just to look cool?

[Charles Wirth]

My guess these are cpus binned too low for EPYC TDP standards and resold and TR, cores are clipped to get under that 180W limit.

Like the cores, I am sure they can clip memory channels.

So far no one has mentioned laser cuts on the CPU as they could be invisible to the naked eye.

[arisythila]

My guess these are cpus binned too low for EPYC TDP standards and resold and TR, cores are clipped to get under that 180W limit.

Like the cores, I am sure they can clip memory channels.

So far no one has mentioned laser cuts on the CPU as they could be invisible to the naked eye.

One would wonder ...

Are they 8 Core chips, and if they are, why put 4 dies on them on a chip only designed to use 2 of them?

It doesn't really make sense to me, to do something just "because". I've learned that the simplest explanation is always the true explanation.

It will be interesting to see what happens in the future. Maybe this is EYPC in disguise??

Here is a picture of the EPYC 7k 32-Core Processor.


Interesting.

[chew*]

Sometimes the simplest of explanations make sense.

Going to B2 on epyc meant there was a B1.

Why toss them in the trash when they are good enough for consumer/prosumer.

[arisythila]

Sometimes the simplest of explanations make sense.

Going to B2 on epyc meant there was a B1.

Why toss them in the trash when they are good enough for consumer/prosumer.

Boom... That could be EXACTLY what it is... But could also mean we may see 32-cores on the consumer side.

[zanzabar]

Boom... That could be EXACTLY what it is... But could also mean we may see 32-cores on the consumer side.

no direct memory access would kind of suck, it would also require cooling well beyond air or the asetek crap they want to use with TR. they would also need a pwm and MB that would work with a 500W load for it to use useful for consumer enthusiasts. two dies seems good IMO, if they get the thing to reasonable clock speeds next refresh it will be really good.

if you wanted a low clock for server chip the 1p epyc is not that much more per core.

[arisythila]

Depends on the market. I do a lot of video encoding, have a MASSIVE plex library, Right now I use a bunch of machines to encode, I would however like to get that down to less machines. Encoding in HEVC 10-bit doing a 2-pass for the highest quality takes awhile, even with my main Encoding Cloud. Upgrading to 2011-3 would be pretty expensive, but if I could get better performance with 32-cores versus 48 Cores. That would be awesome.

[bluestang]

What are you using for encoding? I thought there was a core limit with at which it adversely affected video quality if you use to many cores on 1 video. But of course if you plan on splitting up cores for simultaneous encoding then it won't be affected.

[zanzabar]

What are you using for encoding? I thought there was a core limit with at which it adversely affected video quality if you use to many cores on 1 video. But of course if you plan on splitting up cores for simultaneous encoding then it won't be affected.

it uses 2 cores 4 threads per stream, and you can have as many streams as cores -1.

i also dont get why you would want to bulk transcode for stream and not just get devices that work with h264/x265 and just encode everything with that outside of plex.

[bluestang]

Ah, I was thinking HandBrake or RipBot.

[arisythila]

Right now I use a mix of DVDFab, AnyDVD, and RipBot264.



This is my last crunching job using ripbot, but even these X5690's are not as efficient as I'd like them to be. I encode everything to be H265 8/10Bit currently, the shows that can benefit from the higher color spectrum I do HEVC 10Bit, but most of the other shows I just leave at 8-bit, as its quicker, and less to worry about.

Thanks,

Michael