AMD might announce their new CPU lineup on Tuesday

[leadeater]

10 hours ago, Taf the Ghost said:

Well, there is still a penalty for moving from CCX to CCX, so the Domain would act as 3 other "regional" CCXs. Same with the Epyc or Threadripper now. It doesn't matter too much in the massive parallel environments of servers, but it will on things like Threadripper. Though I did originally think they were going to act as a 4-way NUMA node than a fully independent memory controller.

I don't think it's going to matter too much because a lot of the time that hand off between cores ends up being done through system memory, though any parallel compute tasks that have inter core/cache dependencies should suffer a bit (or a lot) when going above the number of threads per chiplet/cluster. Want some more info on how those chiplet pairs are wired 

[leadeater]

3 minutes ago, 2Buck said:

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.

Yes I have seen that video and I can almost guarantee that is an application optimization issue not Windows itself. Quad socket servers running MSSQL have been around for a long time and that is highly optimized and happens to also be one of the fastest DB engines around. Then you have other applications that run faster in Windows than on Linux in that video too.

 

Just because you see an application running poorly under an OS doesn't mean it's the OS that is causing it.

[Tribalinius]

3 hours 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.

that neon green PATA port though

[leadeater]

10 hours ago, Taf the Ghost said:

@leadeater

 

Now it gets really interesting for Desktop, as there almost has to be a controller Die for it as well. Are we going to see the entire Zen 2 product stack with GPUs on them? That would explain some of the size of the Controller die. The full GPU system on the Raven Ridge APUs is about 110mm², which would fit pretty nicely on that central die. 

Once you start stripping out all those memory controllers an I/O ports for that many chiplets I bet that I/O die is going to be very small, putting a GPU in it should be easy. You could even go high end and do 2 chiplets (upper left and right right) and 2 HBM stacks (lower left and right side), that to me sounds better than Kaby-G. 

[JoeCoke]

13 minutes ago, leadeater said:

Yes I have seen that video and I can almost guarantee that is an application optimization issue not Windows itself. Quad socket servers running MSSQL have been around for a long time and that is highly optimized and happens to also be one of the fastest DB engines around. Then you have other applications that run faster in Windows than on Linux in that video too.

 

Just because you see an application running poorly under an OS doesn't mean it's the OS that is causing it.

Yep, it's just a huge chain of coincidences and all these applications tested just happen to be poorly optimized for a widely adopted OS and well optimized for a niche one. And one test performed better on Windows, so the situation evens out anyway.

[leadeater]

2 minutes ago, 2Buck said:

Yep, it's just a huge chain of coincidences and all these applications tested just happen to be poorly optimized for a widely adopted OS and well optimized for a niche one. And one test performed better on Windows, so the situation evens out anyway.

Most of the applications chosen in the test are all Linux native/developed for Linux first, the applications were chosen originally by Phoronix after all. It's not like I'm saying either one isn't potentially a bit better than the other but 'a lot better' isn't something like 3%-5% on average better for example. Drops in performance on orders of 20%+ is extremely rare to be the OS.

[porina]

4 hours 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?

Xeon Phi is a great example of using the right tool for the job. For my interests, what I'd like is as much FPU power as possible, with fast cache. Everything else doesn't really matter. This is essentially a Xeon Phi. It has the potential to chew through a ton of FP instructions. Atom was used for "other stuff" presumably as it was small and low power. It was never meant to do any heavy lifting. If you just run Windows on it, oh dear, you got a bunch of slow atom cores... and performance sucked. It's like buying a drag racer and complaining it doesn't do corners well.

[Taf the Ghost]

7 minutes ago, leadeater said:

Most of the applications chosen in the test are all Linux native/developed for Linux first, the applications were chosen originally by Phoronix after all. It's not like I'm saying either one isn't potentially a bit better than the other but 'a lot better' isn't something like 3%-5% on average better for example. Drops in performance on orders of 20%+ is extremely rare to be the OS.

Might be the libraries for compiling on Windows.

[leadeater]

1 minute ago, Taf the Ghost said:

Might be the libraries for compiling on Windows.

I know a lot of the time in the past open source/linux devs who would write stuff for windows would do manual core scheduling in C/C++ because 'they knew better' than Windows which would piss Microsoft off to no end. Why do you think Game Mode exists for TR .

[TechyBen]

2 hours ago, NunoLava1998 said:

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

Ok.

[Hold my beer]

via Imgflip Meme Generator

(Inserts meme from 2001!!!)

 

No, but really. I'm loving what AMD are doing. Pity everything I do is core speed dependent, not multi core dependent. So for now, I'm still using NVidia/Intel.

[Taf the Ghost]

20 minutes ago, leadeater said:

Once you start stripping out all those memory controllers an I/O ports for that many chiplets I bet that I/O die is going to be very small, putting a GPU in it should be easy. You could even go high end and do 2 chiplets (upper left and right right) and 2 HBM stacks (lower left and right side), that to me sounds better than Kaby-G. 

The APU on 7nm isn't until late 2019/early 2020, so my assumption is AMD is going to need to roll with the +1 die for Desktop. Which means there's little reason not to put at least the Media & 2D engine on it. Maybe even a cut-down baseline set of CUs (say 7 for yield reasons, but 6 active). That ~100mm² is worth a massive amount of TAM being available for OEMs.

[leadeater]

6 hours ago, porina said:

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.

I am wondering, not expecting it though, that since AMD can combine those FMA units to do larger ops if they do have or could later release a microcode update that adds support for AVX-512. If you have 2 256bit FMA units that 1 512bit unit, much like they are doing in Zen/Zen+ now. That would only give them 1 512bit unit though, Xeon Gold/Plat has 2 so you'd be back at AMD doing half throughput.

[TechyBen]

11 minutes ago, Taf the Ghost said:

The APU on 7nm isn't until late 2019/early 2020, so my assumption is AMD is going to need to roll with the +1 die for Desktop. Which means there's little reason not to put at least the Media & 2D engine on it. Maybe even a cut-down baseline set of CUs (say 7 for yield reasons, but 6 active). That ~100mm² is worth a massive amount of TAM being available for OEMs.

So they are going APU route with this design?

 

It could get very flexible on chiplets/hbm/gpu dies I guess. But gonna be strange figuring out which is the best for balance performance/heat/io etc. I mean, who is going to need a vega APU with 64 cores?

 

4 minutes ago, leadeater said:

I am wondering, not expecting it though, that since AMD can combine those FMA units to do larger ops if they do have or could later release a microcode update that adds support for AVX-512. If you have 2 256bit FMA units that 1 512bit unit, much like they are doing in Zen/Zen+ now. That would only give them 1 512bit unit though, Xeon Gold/Plat has 2 so you'd be back at AMD doing half throughput.

Interesting. Intel, at trying to be the fastest at everything, seems to have painted themselves into a corner at instead being the fastest in specialist workloads (by adding silicone for them?).

 

AMD instead tried to be just average at everything, then ended up being fastest at the average workload.

[GoldenLag]

1 minute ago, TechyBen said:

I mean, who is going to need a vega APU with 64 cores?

someone who really wants a really small Blenderbox?

[Taf the Ghost]

Just now, leadeater said:

I am wondering, not expecting it though, that since AMD can combine those FMA units to do larger ops if they do have or could later release a microcode update that adds support for AVX-512. If you have 2 256bit FMA units that 1 512bit unit, much like they are doing in Zen/Zen+ now. That would only give them 1 512bit unit though, Xeon Gold/Plat has 2 so you'd be back at AMD doing half throughput.

One of the leaks for AIDA64's DB has AVX512 in Zen3. Which makes some sense, as the entire instruction set isn't even fully active.

[Taf the Ghost]

7 minutes ago, TechyBen said:

So they are going APU route with this design?

 

It could get very flexible on chiplets/hbm/gpu dies I guess. But gonna be strange figuring out which is the best for balance performance/heat/io etc. I mean, who is going to need a vega APU with 64 cores?

 

Interesting. Intel, at trying to be the fastest at everything, seems to have painted themselves into a corner at instead being the fastest in specialist workloads (by adding silicone for them?).

 

AMD instead tried to be just average at everything, then ended up being fastest at the average workload.

The addition of a GPU on all Desktop SKUs matters in the OEM space.

[Mira Yurizaki]

31 minutes ago, leadeater said:

I know a lot of the time in the past open source/linux devs who would write stuff for windows would do manual core scheduling in C/C++ because 'they knew better' than Windows which would piss Microsoft off to no end. Why do you think Game Mode exists for TR .

Isn't Game Mode to prevent the apps that aren't NUMA aware to force the app on one die or the other so it won't take a performance hit?

[GoldenLag]

2 minutes ago, M.Yurizaki said:

Isn't Game Mode to prevent the apps that aren't NUMA aware to force the app on one die or the other so it won't take a performance hit?

we do have adaptive mode now on threadripper so i assume AMD has some put some effort to make it check this, though i dunno how well it works

[suits]

Would they really annouse new Ryzen chips calling it "New Horizon"? No, it would be "New Horyzen". Everyone knows that.

[porina]

48 minutes ago, leadeater said:

I am wondering, not expecting it though, that since AMD can combine those FMA units to do larger ops if they do have or could later release a microcode update that adds support for AVX-512. If you have 2 256bit FMA units that 1 512bit unit, much like they are doing in Zen/Zen+ now. That would only give them 1 512bit unit though, Xeon Gold/Plat has 2 so you'd be back at AMD doing half throughput.

I'm not familiar with the practical usage details of AVX-512 vs AVX2, but at a high level AVX-512 only becomes interesting with two units, for double the throughput over AVX2. Otherwise it would be no better.

 

The reasons I can think of for Intel to implement single unit AVX-512 are to expand the perceived availability of hardware support so software developers may be more interested in implementing its use.

[leadeater]

1 hour ago, M.Yurizaki said:

Isn't Game Mode to prevent the apps that aren't NUMA aware to force the app on one die or the other so it won't take a performance hit?

Both, there are applications that can't handle so many cores and won't actually run.

[Mira Yurizaki]

14 minutes ago, leadeater said:

Both, there are applications that can't handle so many cores and won't actually run.

That seems a little odd to me given how Windows' scheduling works.

 

Unless this is the product of a programmer trying to be smart

[leadeater]

46 minutes ago, M.Yurizaki said:

Unless this is the product of a programmer trying to be smart

It's this, some libraries are coded for a fixed maximum number of cores because they are "those people that know better".

[mxk]

lol

[MandoPanda]

On 11/5/2018 at 3:25 AM, RorzNZ said:

Next HoriZen

Serious missed opportunity here.