AMD might announce their new CPU lineup on Tuesday

[TechyBen]

18 hours ago, Mihle said:

Also judging from the photo of the Zen 2 thing, it's 8 squares that is marked  with infinity, so if it's 64 core Max it's most likely 8 chiplets with 8 cores each.

Yep. Intel complained AMD is just "gluing" things together.

AMD: Hold my beer.

 

PS, entire system on a chip for PC soon?

[porina]

6 hours ago, leadeater said:

Yep, though with good optimization you can get a ton more performance by making sure operations are in optimal sizes for the data caches meaning you won't thrash the system memory as much. You can get up to 10x AVX performance even on Intel platforms when doing that.

At times it is unavoidable though. Prime95 at large FFTs is generally ram bandwidth limited on consumer parts, as are parts of Y-cruncher. For Prime95-like loads there is a small potential get out of jail free card for high core counts. Total L3 cache may be sufficient to negate the need to hit ram hard. This doesn't work on lower core counts, but at the number of cores we're talking about here, that's potentially a lot of L3.

6 hours ago, leadeater said:

Not sold on that 8 channel memory being correct though, double the number of cores with no more memory lanes. That's putting a ton of faith in the restructure of the CCXs being good enough.

8 channels at 64 cores? 8 cores per channel... that's going to hurt. To be "practically unlimited" I'd aim for 2 cores per channel.

5 hours ago, leadeater said:

The FMA units have been increased from 128bit to 256bit, oh yea full AVX2 throughput (relative to Intel).

This is the biggest thing for me, assuming it performs as expected. Not like FMA4 before Ryzen (it did twice the work, but took twice the time, so no throughput benefit). Arguably Intel still has AVX-512 but general support for it is still going to take some time, and we probably wont see it in mainstream until they get their 10nm sorted. Still, it is present in server parts already.

[cj09beira]

23 minutes ago, Taf the Ghost said:

@leadeater

 

Apparently, there won't be Gen Z on Rome? Or at least not standard. AMD might have had it in the original design/some version of the Controller Die but then removed it for launch. Might actually be the Socket is the issue. Will have to keep an eye on it.

amd didn't release everything, there is still too much space in that io die left to be explained for example, so they might be keeping that info for the real release

[TechyBen]

Just now, VegetableStu said:

CPU, GPU, PCH, heck even make the CPU take in direct-die HBM O_O

 

imagine Ryzen complete system on a Threadripper-sized substrate O_O

Oh it could get better... use "infinity glue", and just socket in GPU die, CPU die and a HBM die (ok, so not all on 1 die, but you get the idea) so we can have an upgrade cycle... and drop slots all together. XD Would be about nano itx board sized!

[Tribalinius]

Good job AMD, that was a great presentation. Nothing against Milan but I can't wait to see what's after in their roadmap. At the rate they are going, I would not be surprised to see next-gen 256C/512T, DDR5, PCIe 5.0.

[Mihle]

1 hour ago, TechyBen said:

Yep. Intel complained AMD is just "gluing" things together.

AMD: Hold my beer.

 

PS, entire system on a chip for PC soon?

Just a note that the post you quoted was written before they shared that it actually would have 8 chiplets. They had just posted the photo of the IO die with 2 chiplets.

[Mihle]

1 hour ago, porina said:

 

8 channels at 64 cores? 8 cores per channel... that's going to hurt. To be "practically unlimited" I'd aim for 2 cores per channel.

 

At least all cores now have direct access to all memory channels without going through another die than the IO die. Will still be memory starved at memory heavy stuff.

43 minutes ago, Tribalinius said:

Good job AMD, that was a great presentation. Nothing against Milan but I can't wait to see what's after in their roadmap. At the rate they are going, I would not be surprised to see next-gen 256C/512T, DDR5, PCIe 5.0.

For even a 128c on one socket without making a bigger one, they would have to make the IO die smaller to. Especially for 256, that I don't think will happen, there is just not enough space for enough chiplets and even bigger IO die. 

To add chiplets you need more infinity fabric stuff on the IO die. Alternative would be to make CCXes bigger probably, but without making especially the IO die smaller it wouldn't be enough space in that sized socket.

[TechyBen]

Three Words... Double Sided PCB. XD

Hole in the Motherboard, Put connector pins on the edges... and cool both sides with a fan/heatsink.

 

Chiplets everywhere:

https://imgflip.com/i/2lxqu5

via Imgflip Meme Generator

 

PS, PM me for where to send the cheque in the post AMD...

[TigerHawk]

5 minutes ago, TechyBen said:

Three Words... Double Sided PCB. XD

Hole in the Motherboard, Put connector pins on the edges... and cool both sides with a fan/heatsink.

 

Chiplets everywhere:

https://imgflip.com/i/2lxqu5

via Imgflip Meme Generator

 

PS, PM me for where to send the cheque in the post AMD...

You joke but graphics cards already put memory on the back of the PCB.

It'll be the return of slot CPUs like the Pentium II and early IIIs and first gen Athlons.

It could be a thing again. They would need one hell of a fast interface though. Would PCIE 4.0(5.0?) be enough?

[Taf the Ghost]

Vertical Slot CPUs to make a comeback, haha.

[Mira Yurizaki]

1 hour ago, TechyBen said:

Yep. Intel complained AMD is just "gluing" things together.

AMD: Hold my beer.

Funny how 12 years ago, AMD was likely poking fun at Intel for "gluing" two dies together to make a quad core when they were make a "true quad core" chip.

 

Just now, TigerHawk said:

It could be a thing again. They would need one hell of a fast interface though. Would PCIE 4.0(5.0?) be enough?

That's all it would need since that's the only interface CPUs seem to be talking out of these days (except memory).

 

But design sort of limits your cooling options too unless you start building heat sinks with pillars that mount on the motherboard or something.

[Taf the Ghost]

2 minutes ago, TigerHawk said:

You joke but graphics cards already put memory on the back of the PCB.

It'll be the return of slot CPUs like the Pentium II and early IIIs and first gen Athlons.

It could be a thing again. They would need one hell of a fast interface though. Would PCIE 4.0(5.0?) be enough?

Nice synergy, haha, though I don't think they could use Slots again. Pin-outs would be an issue.

[TechyBen]

1 minute ago, M.Yurizaki said:

Funny how 12 years ago, AMD was likely poking fun at Intel for "gluing" two dies together to make a quad core when they were make a "true quad core" chip.

Yeah.. but performance and results. AMD has doubled down on improving. It has backtracked where it saw no returns (older multi core failed to get performance from them). Intel has doubled down on things that don't work, and has not made progress (may be through no fault of their own).

 

AMDs CCX did have some memory/process allocation problems. If it's not fixed, then it will be as bat as Intels (and others) HT spectre/etc security risks. So changes need to be made, and it seems in the short term, AMD is smashing it. Intel needs to either do an AMD, and comeback in the long therm, or have a running start now with something... but their "more cores" seems to be failing.

[TigerHawk]

2 minutes ago, Taf the Ghost said:

Nice synergy, haha, though I don't think they could use Slots again. Pin-outs would be an issue.

Well the whole module is unlikely to be very small given the thermal density of having a double sided CPU, so the heatsink would need to be quite large and beefy. Therefor they would have the space underneath to put multiple slots on the bottom to make up for the lack of area for 2000 pins or however many they need? Maybe? Like a module that takes up the space equivalent of two or three DIMM slots? lol?

Maybe CPU sockets as a whole will just vanish and you will slap your CPUs into a PCIEx16 slot next to your graphics card.

[GoldenLag]

3 minutes ago, TigerHawk said:

Maybe CPU sockets as a whole will just vanish and you will slap your CPUs into a PCIEx16 slot next to your graphics card.

Ever heard of project larebee? Intel GPU using CPU cores.

 

This developed into Xeon Phi

[TigerHawk]

Just now, GoldenLag said:

Ever heard of project larebee? Intel GPU using CPU cores.

 

This developed into Xeon Phi

Yeah. I believe Linus used it in several videos, some 64 core 256 thread thing. It couldn't even run windows very well though IIRC?

[GoldenLag]

16 minutes ago, TigerHawk said:

Yeah. I believe Linus used it in several videos, some 64 core 256 thread thing. It couldn't even run windows very well though IIRC?

Well, Windows is not very multithreaded and doesnt like handling cores very well. But yes. That Xeon Phi. Its purpose is to provide GPU like performance and PCIe lanes for GPUs. Its essentially to increase density

[Mira Yurizaki]

24 minutes ago, TigerHawk said:

Maybe CPU sockets as a whole will just vanish and you will slap your CPUs into a PCIEx16 slot next to your graphics card.

ECS kind of did this:

 

2 minutes ago, GoldenLag said:

Well, Windows is not very multithreaded and doesnt like handling cores very well. But yes. That Xeon Phi. Its purpose is to provide GPU like performance and PCIe lanes for GPUs. Its essentially to increase density

Windows' problem is mostly how it schedules work, which can be fixed with a kernel patch if your processor really needs special considerations.

[DeepFriendLettuce]

On 11/4/2018 at 10:49 AM, Humbug said:

about the 7nm Vega data centers parts?

I was assuming it was going to be able the 7nm Vega GPU as well. 

[DeepFriendLettuce]

17 minutes ago, M.Yurizaki said:

ECS kind of did this:

 

Windows' problem is mostly how it schedules work, which can be fixed with a kernel patch if your processor really needs special considerations.

I litterally had an issue earlier this week that was resolved by adjusting Scheduled tasks that Windows had setup on the system.

[Ravager911]

A short summary:

[Mira Yurizaki]

33 minutes ago, Phenix51 said:

I litterally had an issue earlier this week that was resolved by adjusting Scheduled tasks that Windows had setup on the system.

What I was describing is different.

[NunoLava1998]

can some of you stop with the shitty imgflip memes from 2011

[leadeater]

3 hours ago, GoldenLag said:

Well, Windows is not very multithreaded and doesnt like handling cores very well.

Windows handles many cores just fine, 2990WX is not representative of that type of issue at all. It's just a bad design that breaks the NUMA standard and didn't get special kernel consideration like in Linux because of it's nicheness, and realistically because of how complex the Windows kernel is in comparison would not be worth the risk.

[JoeCoke]

3 minutes ago, leadeater said:

Windows handles many cores just fine, 2990WX is not representative of that type of issue at all. It's just a bad design that breaks the NUMA standard and didn't get special kernel consideration like in Linux because of it's nicheness, and realistically because of how complex the Windows kernel is in comparison would not be worth the risk.

It's not just the 2990WX though, Linux also handles the 7980XE a lot better. This applies to any high core count solutions AFAIK. Yes, the 2990WX makes the situation more exaggerated due to it's unusual design, but it's definitely not just the 2990WX.