AMD Discussion

[koyoot]

Infinity Fabric is quite interesting feature.

Let me give you an example. There will be possible "combining" the power of GPU from APU with the discreet GPU.

For example. You can have 95W APU with GPU that has 16 CU's, and you can have discreet GPU with 48 CU's. Combined, they will have 64 CU's, that are visible to your application.

Its first step to scalability of GPUs, that AMD touted in their roadmap(With Navi GPU architecture slated for 2018+ years). Of course not in every situation you will get that particular effect. But what AMD achieved appears to be very close, to this scenario.

 

[Joe The Dragon]

IF you remember my post from some time ago, about 2017 being for AMD brand reset, well I got some information that can validate this.

AMD will build their brand on one simple idea. Innovation. This is what they are after. Let me give you an example. From AMD perspective: lets offer 8 core CPU with lowest price possible, that can shake up the market. Lets offer 6 core CPU with lowest possible price, that can shake up the market. Lets offer 4 core CPU with a price that can shake up the market. This is innovation in a market that has not seen any innovation for last... 10 years. Lets make high-end CPUs a mainstream solution, and lets be the brand who everybody will associate with this shake up.

GPUs? Lets offer future proof idea for GPUs, extremely powerful from compute perspective, with lots of graphical capabilities, and features, priced extremely competitively, with... very good efficiency. Competitive efficiency. Where is the innovation? In hardware features, and ideas, for execution of the tasks.

Biggest innovation however comes with APUs. For those especially was designed Infinity Fabric. And they will get the highest benefit from using it.

Don't forget also more pci-e in sub $250-$300 cups as well vs intel's 16+4(DMI).

16 for video is good but 4 dmi is to slow for pci-e storage, networking, usb all over the X4 link. And that is with out even thinking of 10G-E.

 

[koyoot]

 

[ManuelGomes]

CPUs seem nice. Balanced, perfectly differentiated.
Mobo however... same old assortment of goods from Taiwan I guess. Vivid LED DJ? Smart Ear? FLY.NET? Wow, this a Disco or a mobo? At least the name is not something like Super Duper FTW Ultra GT whatever (well, it is indeed GT, go figure!).
[doublepost=1485384701][/doublepost]SR5 or SR7 inside Scorpio?

 

[AidenShaw]

CPUs seem nice. Balanced, perfectly differentiated.

I was actually a bit puzzled by the CPU specs. SMT is basically next to useless for a desktop for most apps, and even for the apps where it helps the benefit is typically in the 10% to 25% range. The 8C/16T would have to be priced within 10% to 20% of the 8C/8T. The jump from 4C/8T to 8C/8T could be quite a bit more, since you're getting twice as many real cores.

(This guess is based on statements around these parts the Zen's SMT is as good as hyper-threading - which is something that I usually disable on Intel processors.)

 

[koyoot]

Bristol Ridge APUs will be priced under 100$ mark. That gives you an idea, where SR3 CPUs will fall.

Currently the most reliable rumor is saying that Top-end SR3 CPU with 4C/8T will cost 199$. And will offer i7 4790K level of performance.

SR5 or SR7 inside Scorpio?

Im afraid that neither .

 

[AidenShaw]

Bristol Ridge APUs will be priced under 100$ mark. That gives you an idea, where SR3 CPUs will fall.

Currently the most reliable rumor is saying that Top-end SR3 CPU with 4C/8T will cost 199$. And will offer i7 4790K level of performance.
Im afraid that neither .

The broken record repeats - links?

 

[ManuelGomes]

koyoot, it's supposed to be an 8 core CPU. Won't they be using Ryzen?
That would be a bit weird since it's also rumored to be Vega inside (or only Polaris), or a derivative of it.
Maybe it's all crap. But for such a new powerful console going with old tech seems odd.
The 6TFlops must come from somewhere.

 

[koyoot]

koyoot, it's supposed to be an 8 core CPU. Won't they be using Ryzen?
That would be a bit weird since it's also rumored to be Vega inside (or only Polaris), or a derivative of it.
Maybe it's all crap. But for such a new powerful console going with old tech seems odd.
The 6TFlops must come from somewhere.

The GPU has 384 bit memory controller. Vega has only HBM2. And funniest part: Project Scorpio lacks ESRAM, which iwas inherent part of two previous versions of Xbox console.

The essence: Project Scorpio APU is completely custom.

 

[koyoot]

It appears that 8 core Ryzen die is under 200mm2 in size.

 

[Stacc]

It appears that 8 core Ryzen die is under 200mm2 in size.

Sounds about right, maybe a little small. 4 core skylake is 122 mm2 and 10 core Broadwell-E is 246 mm2.

 

[koyoot]

It is actually right in the middle of the die sizes, you brought up .

 

[Stacc]

There is no integrated graphics on the ryzen die, right? That comes from a separate die on the cpu package.

 

[koyoot]

Ryzen is just the CPU. Raven Ridge is a combination of 4C/8T setup with 16 CU GPU and 2 stacks of HBM2.

The thing is, that Raven Ridge GPU is slightly bigger, compared to Polaris 11.

 

[ManuelGomes]

But there are rumors of a GDDR5X Vega coming up as well, maybe Vega 11. I guess I read it somewhere, don't know where now. No link for Aiden I'm afraid :-(

 

[Stacc]

Ryzen is just the CPU. Raven Ridge is a combination of 4C/8T setup with 16 CU GPU and 2 stacks of HBM2.

The thing is, that Raven Ridge GPU is slightly bigger, compared to Polaris 11.

If the raven ridge GPU is based on Vega then it being bigger makes sense. It seems Vega compute units take up more die space.

2 Stacks of HBM2 seems like overkill for a 16 CU GPU. Its the same amount of bandwidth Vega 10 is going to get. Are you sure its not 1? The only other thing I could think of is if they are going to use it for some crazy CPU caching scheme.

But there are rumors of a GDDR5X Vega coming up as well, maybe Vega 11. I guess I read it somewhere, don't know where now. No link for Aiden I'm afraid :-(

That makes sense. If Vega 10 needs ~500 GB/s with 2 stacks of HBM2, then a smaller vega GPU could get by with plain old GDDR5(X).

 

[koyoot]

If the raven ridge GPU is based on Vega then it being bigger makes sense. It seems Vega compute units take up more die space.

2 Stacks of HBM2 seems like overkill for a 16 CU GPU. Its the same amount of bandwidth Vega 10 is going to get. Are you sure its not 1? The only other thing I could think of is if they are going to use it for some crazy CPU caching scheme.



That makes sense. If Vega 10 needs ~500 GB/s with 2 stacks of HBM2, then a smaller vega GPU could get by with plain old GDDR5(X).

Yes, I know it sounds like overkill, but trust me it isn't. Its part of the increase of performance I have written before. 1024 GCN core chip in APU will be between 25 and 75% faster than 1024 GCN core RX 460.

Yes, we are in the ranges of RX 470D/WX 5100.

HBM2 is used not only by GPU, but also CPU. They are connected together through interposer, and Infinity Fabric, to increase the bandwidth of data, and throughput, itself.

I know that people will say that for CPU latency is more crucial than bandwidth. Yes it is, but this is the exact case for which Infinity Fabric was designed. To help mitigate the latency of HBM2.

 

[AidenShaw]

Yes, I know it sounds like overkill, but trust me it isn't. Its part of the increase of performance I have written before. 1024 GCN core chip in APU will be between 25 and 75% faster than 1024 GCN core RX 460.

Yes, we are in the ranges of RX 470D/WX 5100.

HBM2 is used not only by GPU, but also CPU. They are connected together through interposer, and Infinity Fabric, to increase the bandwidth of data, and throughput, itself.

I know that people will say that for CPU latency is more crucial than bandwidth. Yes it is, but this is the exact case for which Infinity Fabric was designed. To help mitigate the latency of HBM2.

I hope for the sake of people who own AMD stock that Zen isn't the big yawn that Polaris is.

Lots of hype before it showed up, then "huh?".

And if you own Apple stock, hope for Pascal in the MP7,1.

 

[koyoot]

I hope for the sake of people who own AMD stock that Zen isn't the big yawn that Polaris is.

Lots of hype before it showed up, then "huh?".

And if you own Apple stock, hope for Pascal in the MP7,1.

In case you missed it, AMD did a live demo of Ryzen CPU performance, and its per clock slightly faster than Broadwell-E CPU.

Secondly, I should hope for Pascal GPUs in MP 7.1 even if Vega is better architecture?

Speaking of stock. You have to be soft in the head to not buy the AMD stock right now .

 

[koyoot]

https://twitter.com/BitsAndChipsEng/status/824984734212247554

If we consider that a 8c/16t Ryzen die costs like a i7-7700K die, AMD can price a 8c/16t @ $400 without problem: margins remain very high

If anyone asks what this means: base SKU with 8C/16T and 3.45/3.7 GHz, can cost 399$. Higher end for example 3.6/4.0 GHz can cost 599-649$.

Before you will stop me saying that the price gap is too big. Ask yourself how big the difference in pricing is between i7 6800K and 6850K. The same CPU, but with slightly higher core clock.

 

[koyoot]

http://www.bitsandchips.it/52-engli...obal-foundries-pdf-on-the-14nm-finfet-process

Interesting analysis of GloFo PDF, about 14 nm process. According to them, the APU can have at 95W 4.0 GHz core clock on CPU.

I was more "realistic". 3.4/3.7 GHz and 1350 MHz on core on 16 CU GPU. Plus 4 GB's of HBM2.

 

[AidenShaw]

http://www.bitsandchips.it/52-engli...obal-foundries-pdf-on-the-14nm-finfet-process

Interesting analysis of GloFo PDF, about 14 nm process. According to them, the APU can have at 95W 4.0 GHz core clock on CPU.

I was more "realistic". 3.4/3.7 GHz and 1350 MHz on core on 16 CU GPU. Plus 4 GB's of HBM2.

Is the HBM2 RAM used like an L4 cache - or is that all you get?

4 GiB is typical for a good smartphone - but I can't see how it would be adequate for anything resembling a computer.

 

[koyoot]

Is the HBM2 RAM used like an L4 cache - or is that all you get?

4 GiB is typical for a good smartphone - but I can't see how it would be adequate for anything resembling a computer.

Its cache memory. AMD described HBM2 as cache memory.

Don't worry, its something like EDRAM in Intel CPUs, but much better.

I was more "realistic". 3.4/3.7 GHz and 1350 MHz on core on 16 CU GPU. Plus 4 GB's of HBM2.

Correction here. I was thinking that 95W, desktop APU, can have those specs.

 

[ManuelGomes]

No 6 core Ryzen?
http://wccftech.com/amd-ryzen-processor-6-core-model-700-usd-rumor/

 

[koyoot]

Its WTFTech. They do not even know what is the base clock for all of AMD SKUs at launch, depsite AMD announcing it, WTFTech repeats them incorrectly.

6 cores will be available. Be it at launch or later in the year, we will see.