[UPDATE] AMD hosting "capsaicin" live webcast at GDC on March 14th-could announce Fury X2 and demo Polaris

[Doobeedoo]

Fury X2 feels overdue, Polaris being so close to. Like mentioned in top post comments, I doubt they would make it i n 14nm though. Would've been great if they managed to supply it with more HBM per GPU for sure.

 

[Watermelon Guy]

2 minutes ago, Doobeedoo said:

Fury X2 feels overdue, Polaris being so close to. Like mentioned in top post comments, I doubt they would make it i n 14nm though. Would've been great if they managed to supply it with more HBM per GPU for sure.

 

They probably wont be releasing a flagship until q3/q4. Sk hynix are mass producing HBM then. http://wccftech.com/sk-hynix-hbm2-mass-production-q3-2016/

[Doobeedoo]

28 minutes ago, Watermelon Guy said:

They probably wont be releasing a flagship until q3/q4. Sk hynix are mass producing HBM then. http://wccftech.com/sk-hynix-hbm2-mass-production-q3-2016/

Samsung already haves 4GB stacks which they had mass production from January so Hynix is an alternative.

[Watermelon Guy]

44 minutes ago, Doobeedoo said:

Samsung already haves 4GB stacks which they had mass production from January so Hynix is an alternative.

Amd will probably be using Hynix, considering they helped them develop HBM

[Citadelen]

2 minutes ago, Watermelon Guy said:

Amd will probably be using Hynix, considering they helped them develop HBM

I'm pretty sure Hynix can only supply to AMD, but Samsung can supply to both.

[Watermelon Guy]

5 minutes ago, Citadelen said:

I'm pretty sure Hynix can only supply to AMD, but Samsung can supply to both.

Meh, Im really not bothered by the first leap in a process node, its always underwhelming. They talk about hugge per. Gains, but they get 40% at best.

[Doobeedoo]

13 minutes ago, Watermelon Guy said:

Amd will probably be using Hynix, considering they helped them develop HBM

I'm sure they will, though later on. If everything is on schedule like it seems to be soon then first batch will be with Samsung for sure.

[DocSwag]

1 hour ago, Citadelen said:

I'm pretty sure Hynix can only supply to AMD, but Samsung can supply to both.

I think AMD gets priority access to Hynix's. I'm pretty sure AMD can get it at first and after a while when AMD doesn't need the HBM Nvidia can get it too. I'm not too sure, though. I could be wrong. Not sure about the specifics.

[Citadelen]

Just now, DocSwag said:

I think AMD gets priority access to Hynix's. I'm pretty sure AMD can get it at first and after a while when AMD doesn't need the HBM Nvidia can get it too. I'm not too sure, though. I could be wrong. Not sure about the specifics.

As am I.

[DocSwag]

1 hour ago, Watermelon Guy said:

Meh, Im really not bothered by the first leap in a process node, its always underwhelming. They talk about hugge per. Gains, but they get 40% at best.

*in a marketing voice*

But wait! This time we are transitioning not one, but two nodes at once! We are making a double jump by going from 28nm to 14nm! And even more, we are jumping to FinFET, which bring even bigger efficiency games, which makes YOUR graphics look even better! On top of that, we have redesigned nearly everything on the GPU, and are integrating HBM2 memory on the GPU, which will give you THREE TIMES higher memory bandwidth as well as SMALLER cards! Wait and buy when Polaris launches!

 

I have no idea. I would imagine it would be more than 40% since this is a pretty big update, but then again 28nm has been optimized pretty heavily so the jump may not be so big. Only time will tell.

[DocSwag]

Yo guys just updated this article with new info.

[Master Disaster]

I'm confused. What exactly does chilli have to do with AMD GPUS

[Texhnlyze]

Inb4 "it's an overclockers dream".

[Humbug]

On ‎3‎/‎8‎/‎2016 at 0:44 AM, Watermelon Guy said:

Meh, Im really not bothered by the first leap in a process node, its always underwhelming. They talk about hugge per. Gains, but they get 40% at best.

40% is huge

[DocSwag]

On March 7, 2016 at 11:14 AM, Watermelon Guy said:

Meh, Im really not bothered by the first leap in a process node, its always underwhelming. They talk about hugge per. Gains, but they get 40% at best.

40% IPC in single core is a really big jump, and keep in mind that's against excavator, not vishera, so compared to vishera it'd be more like 60-80%.

Plus multi core should be vastly improved due to AMD finally getting rid of that module system.

[Citadelen]

On ‎08‎/‎03‎/‎2016 at 7:14 PM, Watermelon Guy said:

Meh, Im really not bothered by the first leap in a process node, its always underwhelming. They talk about hugge per. Gains, but they get 40% at best.

40% is only from not having shared resources, other factors come into play, cache latency, architectural improvements, less leakage ect.

 

[Tomsen]

5 minutes ago, DocSwag said:

40% IPC in single core is a really big jump, and keep in mind that's against excavator, not vishera, so compared to vishera it'd be more like 60-80%.

Plus multi core should be vastly improved due to AMD finally getting rid of that module system.

That was one of the "module systems" positive size. A fairly goot MT scaling.

CMT was more a limitation of ST.

 

1 minute ago, Citadelen said:

40% is only from not having shared resources, other factors come into play, cache latency, architectural improvements, less leakage ect.

 

How do you know? You are not going to be gaining anything major, if you have a real bottleneck in the chip.

[Misanthrope]

On 3/7/2016 at 0:33 PM, Thunderjolt said:

Oooooooh boy, If you thought the R9 Fury was a hot card just wait till you see our new "capsaicin" branded ones. You'll be able to cook dinner on it

Except nobody thinks that. Mainly cause it's false: at 275w it's quite reasonable, in fact fairly close to Nvidia not too long ago with the 780ti.

 

 

[Prysin]

2 hours ago, Tomsen said:

That was one of the "module systems" positive size. A fairly goot MT scaling.

CMT was more a limitation of ST.

 

How do you know? You are not going to be gaining anything major, if you have a real bottleneck in the chip.

actually CMT has poor MT scaling too.

 

a Core i7 scales around 4.8 times its single threaded performance... which with cruel napkin math means HT adds 20% performance to each physical core in Cinebench R15...

In contrast, a FX 6300 scales 5.09 times single core performance, a FX 8000 scores 6.67 times its single core performance. Meaning CMT cannot achieve perfect scaling, let alone any reasonable scaling at all. Bear in mind, i tested this with my 4790k and my FX 8320... Even OCd to 4.62GHz, the 8320 couldnt get closer to equal scaling then what i wrote earlier.

 

Thing is, CMT offers ONE benefit, the ability to do float and ALU operations simultaneously. However due to the shared nature of the modules, it cannot actually do it simultaneously. The whole FX line is plagued by queueing and fetch latency making it slow down.

 

The same issue persists in Steamroller, although it is not nearly as bad with a scaling of 3.74x single thread, it is the closest AMDs current CMT architecture has gotten to decent scalability.

 

CMT is a limitation in and of itself, since it is a bastardization of Clustered Integer Core and Simultaneous Multi Threading. Yes, has 50% of the layout of a Hyper Threaded CPU, and 50% the layout of "Clustered Integer Core", mashed together so that it somehow works without crashing.

[Tomsen]

15 minutes ago, Prysin said:

actually CMT has poor MT scaling too.

 

a Core i7 scales around 4.8 times its single threaded performance... which with cruel napkin math means HT adds 20% performance to each physical core in Cinebench R15...

In contrast, a FX 6300 scales 5.09 times single core performance, a FX 8000 scores 6.67 times its single core performance. Meaning CMT cannot achieve perfect scaling, let alone any reasonable scaling at all. Bear in mind, i tested this with my 4790k and my FX 8320... Even OCd to 4.62GHz, the 8320 couldnt get closer to equal scaling then what i wrote earlier.

 

Thing is, CMT offers ONE benefit, the ability to do float and ALU operations simultaneously. However due to the shared nature of the modules, it cannot actually do it simultaneously. The whole FX line is plagued by queueing and fetch latency making it slow down.

 

The same issue persists in Steamroller, although it is not nearly as bad with a scaling of 3.74x single thread, it is the closest AMDs current CMT architecture has gotten to decent scalability.

 

CMT is a limitation in and of itself, since it is a bastardization of Clustered Integer Core and Simultaneous Multi Threading. Yes, has 50% of the layout of a Hyper Threaded CPU, and 50% the layout of "Clustered Integer Core", mashed together so that it somehow works without crashing.

You are not going to be seeing as well of a scaling on CMT design as a regular core design, that is common sense. Bulldozer and piledriver do share the decoders, and can be limited in certain MT workloads. But my point is, CMT can provide high MT performance (even surpassing the 4770k, in certain workloads). Btw, Isn't cinebench also heavily reliable on FP performance?

 

I have also done some testing, using a debian system, and optimized binaries (for both the 4770k, and FX 8350). Some test, the 4770k won, but on some very intensive MT workloads, the 8350 was slightly ahead.

 

That is just wrong. That is the effect of superscalar implementation. Why do you think it cannot issue both a ALU and FPU/SIMD operation in the same cycle?

 

Steamroller added additional decoders, to combat that bottleneck.

 

It is more related to the implementation of CMT. 50% the layout of hyper-threading? That is an interesting estimate, very much doubt it. AMD only needed to update the FPU/SIMD scheduler to support "SMT". Can we even call it "SMT", when that is the nature of SIMD operations?

[Prysin]

1 minute ago, Tomsen said:

You are not going to be seeing as well of a scaling on CMT design as a regular core design, that is common sense. Bulldozer and piledriver do share the decoders, and can be limited in certain MT workloads. But my point is, CMT can provide high MT performance (even surpassing the 4770k, in certain workloads). Btw, Isn't cinebench also heavily reliable on FP performance?

 

I have also done some testing, using a debian system, and optimized binaries (for both the 4770k, and FX 8350). Some test, the 4770k won, but on some very intensive MT workloads, the 8350 was slightly ahead.

 

That is just wrong. That is the effect of superscalar implementation. Why do you think it cannot issue both a ALU and FPU/SIMD operation in the same cycle?

 

Steamroller added additional decoders, to combat that bottleneck.

 

It is more related to the implementation of CMT. 50% the layout of hyper-threading? That is an interesting estimate, very much doubt it. AMD only needed to update the FPU/SIMD scheduler to support "SMT". Can we even call it "SMT", when that is the nature of SIMD operations?

well according to wikipedia and some other articles they did base it out of SMT aswell as Clustered Integer Core which is a concept made by another company (not IBM or Intel).

 

If you look at block diagrams of Intels Core CPUs, alongside CIC you see very close similarities, but not total equalities.

[Tomsen]

1 hour ago, Prysin said:

well according to wikipedia and some other articles they did base it out of SMT aswell as Clustered Integer Core which is a concept made by another company (not IBM or Intel).

 

If you look at block diagrams of Intels Core CPUs, alongside CIC you see very close similarities, but not total equalities.

Wikipedia states: " The floating-point cores are similar to a single core processor that has the SMT ability". Yes, similar "CMT" implementation has been seen previously in customized chips.

 

Block diagrams are basically marketing material. It is with a very high level of abstraction, hence why many block diagrams looks alike.

 

[Citadelen]

6 hours ago, Tomsen said:

-snip-

AMD has stated that a move to SMT alone would yield greater than 40% gains in IPC.

[Tomsen]

1 minute ago, Citadelen said:

AMD has stated that a move to SMT alone would yield greater than 40% gains in IPC.

Are we considering single-thread or multi-thread IPC? Single-threaded, SMT isn't going to increase IPC. Multi-threading, sure, 

[Citadelen]

11 minutes ago, Tomsen said:

Are we considering single-thread or multi-thread IPC? Single-threaded, SMT isn't going to increase IPC. Multi-threading, sure, 

CMT only affects single-threaded performance in a negative way, as the core only has the resources of half a "cluster" to draw from. A switch to SMT will improve both for AMD, as cores will be stronger, and proper cores with hyper-threading is better than four "clusters".