AMD: Speculation, Rumors, and Discussion (Archive)

[silent_guy]

Polaris is certainly more than just a shrink if only by virtue of the fact that there is no such thing anymore (since 55 nm at least).

You're talking optical shrink. A port of a design from one process to smaller one is still very much called a shrink.

 

[CSI PC]

If Polaris is just a current GPU shrunk down to 14nm then AMD will be dead to me. So I'm hoping - well, expecting really - that that isn't the case, because we need some decent competition in the dGPU marketplace. Otherwise NV will be free to do whatever they like, and we know what a bunch of shitty people they are over there. They'd abuse the hell out of their market position, you know this is true.

TBH I think NVIDIA is going to be just as disappointing to the enthusiasts/performance early adopters; just like we are waiting yet again on Intel to release the "true" CPU successor (ok maybe ignoring mobile where Intel has now started to improve a lot).
I think enthusiasts may need to wait awhile on both manufacturers' GPUs and Intel releasing a worthy successor CPU.
Cheers

 

[Triskaine]

You're talking optical shrink. A port of a design from one process to smaller one is still very much called a shrink.

Semantics. My Employer prefers the term redesign, since it is quite rare nowadays that a design is ported to a smaller process without any design changes at all.

 

[3dilettante]

You're talking optical shrink. A port of a design from one process to smaller one is still very much called a shrink.

Come to think of it, isn't Fiji something along the lines of ~30% more power-efficient than Tonga?
Port Fiji to a reduced CU count, tweak a few options, and then avoid pushing GCN as far past its sweet spot as AMD did outside of Nano, and then base your marketing graph on the unmemorable launch version of Tonga.

 

[Grall]

If Polaris is just a shrink but it manages to perform well with improved perf/W, with just low level fixes, then why not?

Because after years of 28nm GPUs, I expect a more radical departure than a die shrink of existing IP. Feature-complete DX12 support for one. H.265 codec in hardware, and so on.

 

[silent_guy]

Semantics. My Employer prefers the term redesign, since it is quite rare nowadays that a design is ported to a smaller process without any design changes at all.

My employer calls it a shrink. Are we even now?
And it's not semantics when grall meant exactly that: a straight port of GCN v8 from 28 to 14nm.

 

[silent_guy]

Because after years of 28nm GPUs, I expect a more radical departure than a die shrink of existing IP. Feature-complete DX12 support for one. H.265 codec in hardware, and so on.

When looking at the IP levels, there seem to be increases for quite a bit of blocks.

So it really depends what the gfx IP level covers. If it's just the shader core and not things like ROPs and other essential blocks that are not part of the shaders, then they could still add those feature levels while keeping the shaders identical.

There definitely are changes in their codecs. So that concern is probably covered as well.

 

[Grall]

When looking at the IP levels, there seem to be increases for quite a bit of blocks.

Yes, it would be very strange if it wasn't - like what the heck have they been doing for all this time? But seriously - feature-complete DX12 is like a gold standard to me. MS should never have allowed optional features, it's just holding adoption back.

 

[Alexko]

Come to think of it, isn't Fiji something along the lines of ~30% more power-efficient than Tonga?
Port Fiji to a reduced CU count, tweak a few options, and then avoid pushing GCN as far past its sweet spot as AMD did outside of Nano, and then base your marketing graph on the unmemorable launch version of Tonga.

But Fiji's 30% advantage over Tonga is mostly HBM-related, isn't it?

 

[3dilettante]

But Fiji's 30% advantage over Tonga is mostly HBM-related, isn't it?

It would be complicated to tease it out. Power-wise, the interface saved 20-30W over Hawaii, which makes power reduction for memory ~10% of the budget. On the other hand, Tonga's bus is half that of Hawaii.
How well Fiji used its bandwidth versus Tonga is not clear. It wasn't clear versus Hawaii.
It was claimed that Fiji implemented better power management when Fury X was launched. That could be new to Fiji, although it's also the case that AMD has marketed things as new that were there all along.

In terms of the performance component, Tonga was not fully enabled at launch, and there was a significant time gap between it and Fiji.
AMD didn't choose the 3xx generation with years of process learning and possibly other tweaks such as improved firmware and Powertune heuristics as a starting point.

 

[silent_guy]

But Fiji's 30% advantage over Tonga is mostly HBM-related, isn't it?

HBM alone can't explain most of it. They must have done more than that.

OTOH, using Tonga as reference makes it easier to come up with 2.5x claims.

 

[fellix]

AMD’s Polaris 10 GPU May Contain 2304 Shader cores and 8 GB of GDDR5 VRAM with a 256-bit Memory bus

 

[gamervivek]

So based on the all the leaks AMD's Nextgen Finfet Lineup will (roughly) look like this:

P11/Baffin: 20 CUs 128-bit GDDR5 ~ 115 mm²
P10/Ellesmere: 40 CUs 256-bit GDDR5(X) ~ 220 mm²
V10/Greenland: 64 CUs (Gen 9 Graphics IP) (1/2 rate DP) 2048-bit HBM2 ~ 360 mm²

If Nvidia can get GP100 (or a derivative of it) into the Desktop they have this one in the bag.

There are two vega chips and greenland is something other than a dGPU. From the linkedin link,

Project ‘Greenland’; Leading project of Graphic IP v9.0
As leading chip of first graphic IP v9.0 generation, it has full capacity of 4096 shader processor, along with whole new SOC v15 architecture.

So the Vega chips might have the gfx ip as 9.0(8.0 currently in Tonga/Fiji/Polaris) too and I'm guessing Tahiti + Hawaii sized chips. Outside chance of something on the level of Fiji.

The hitch in estimating Polaris cards' performance is that the Baffin XT samples are shipping to India at 60% of the price of Fiji XT. Ellsmere is going to be pricier still so its XT incarnation is going to end up around 500$ and higher price easily. Very likely that it ends up closer to Fiji XT than Grenada XT in that case.

 

[silent_guy]

There are two vega chips and greenland is something other than a dGPU. From the linkedin link,

Amazing. You'd think that a few of these cases would result in a strongly worded email sent to All? Too bad he didn't get the memo about listing the die size as well...

What is it about AMD that they attract a seemingly never ending stream of tattling idiots?

So the Vega chips might have the gfx ip as 9.0(8.0 currently in Tonga/Fiji/Polaris) too and I'm guessing Tahiti + Hawaii sized chips. Outside chance of something on the level of Fiji.

At least we now see an explanation for the perf/W jump between Polaris and Vega.

 

[Jawed]

A SOC isn't a GPU ASIC, is it? This 4096 ALU rumour perhaps correlates with the HPC Zen + HBM GPU SOC. EDIT: whoops, spoke too soon: his description of R290 is as a SOC. So everything is a SOC in his world view.

Also, 9.0 versus 8.x pretty much confirms that Polaris is a tweak, much like Tonga/Fiji is a tweak.

EDIT 2: Has the console forum noticed Project G and Project K?... PS4.5 and XB150?

 

[yuri]

his description of R290 is as a SOC. So everything is a SOC in his world view.

Well, public, marketing and journalist's definition of SoC is "it got everything packed on a single die". CPU, GPU, IO, IMC, ... all together
However, it seems most of engineers view a SoC as a chip with multiple processing units. So GPUs kinda tend to be SoCs...

Here is an article on this topic:

if an ASIC contains one or more processor cores then it's an SoC.

http://www.eetimes.com/author.asp?section_id=36&doc_id=1322856

 

[pTmdfx]

Perhaps for those things being promised for Polaris, they are possible not to alter the external of the architecture, thus causing no increment in the driver-facing GFXIP level? Say for pixel discard accelerator.

P.S. It seems there is GFXIP 8.1.

 

[Nemo]

Hawaii -- Graphics IP7, GCN Gen2
Tonga -- Graphics IP8, GCN Gen3
Fiji -- Graphics IP8, GCN Gen3
Polaris -- same Graphics IP8, but new GCN Gen4?

 

[3dilettante]

The correlation between GCN ISA revision and GFX IP is not perfect. There are version differences between the CI implementations, but at least from an ISA standpoint it isn't clear that there is a difference.

There are certain implementation differences like Hawaii's DP capability that could point to some relationship between CU architecture and GFX, but at the same time Carrizo's GCN3 has certain architectural features active that the other 8.0 devices do not.

AMD is free to point to whatever implementation or specification tweak and call it Gen4, and the changes might not impact the graphics domain.

 

[gamervivek]

Amazing. You'd think that a few of these cases would result in a strongly worded email sent to All? Too bad he didn't get the memo about listing the die size as well...

What is it about AMD that they attract a seemingly never ending stream of tattling idiots?


At least we now see an explanation for the perf/W jump between Polaris and Vega.

Folks like John Bridgman are more discreet though.

https://semiaccurate.com/forums/showpost.php?p=258461&postcount=514

As for the gfx ip and the revised polaris architecture slide that someone posted on SA as well, he had this to say,

Yeah, the second picture doesn't look accurate at all. I was initially surprised how few software-visible changes there were in Polaris but AFAICS the original slides seem to be about right.