Exclusive: The AMD Inside Story, Navi GPU Roadmap And The Cost Of Zen To Gamers

[4K8KW10]

Across a wider range of titles its performance is very hit and miss - while it can keep up with the 1070 in some stuff it really falls down in others and it is propped up quite a bit in that review by not falling down quite as badly as some of the other cards in 4K in some titles such as https://www.techpowerup.com/reviews/MSI/RX_580_Mech_2/12.html

It should have been possible to make a much more consistently fast card from it even with the problem of being stuck on 28nm.

Being stuck to 4GB? Has anyone said what it would have been to have Fiji with 8GB/12GB and more bandwidth?

I think it's quite explainable considering AMD's choice to focus on Zen. GCN was at the end of the road, development became a lot harder and they simply didn't have the budget to make everything work.

Zen real development and costs ended with Keller's departure. That is many moons ago and doesn't explain why 2-3 years later, the situation is even worse.

 

[KompuKare]

Zen real development and costs ended with Keller's departure. That is many moons ago and doesn't explain why 2-3 years later, the situation is even worse.

I somehow doubt that.

While I haven't researched it, I would imagine that while the initial development is expensive and is something where you'd have to offer big salaries to get good talent, the next stages (testing, testing, validation, validation, mask making, chipset development, packaging, software drivers etc.) will cost a lot more than that. Even if it is far more mundane / 'boring' work.

A company with a bigger R&D budget would have hired extra people earlier than AMD. Or instead in such a headcount constrained company like AMD, anyone who could do either job (validating etc. ASICs whether CPU or GPU) for would have been concentrating on Zen.

 

[Freddie1980]

Being stuck to 4GB? Has anyone said what it would have been to have Fiji with 8GB/12GB and more bandwidth?

What it would have been? A lot more expensive, if I recall the amount of HBM 1 you could have was directly linked to the size of the bus. To get 8Gb if would have needed an 8,192 bit bus which would have been impractical and prohibitively expensive (that's a mute point because Fury must have prohibitively expensive as it was because it had a rather short production run).

 

[4K8KW10]

I somehow doubt that.

While I haven't researched it, I would imagine that while the initial development is expensive and is something where you'd have to offer big salaries to get good talent, the next stages (testing, testing, validation, validation, mask making, chipset development, packaging, software drivers etc.) will cost a lot more than that. Even if it is far more mundane / 'boring' work.

A company with a bigger R&D budget would have hired extra people earlier than AMD. Or instead in such a headcount constrained company like AMD, anyone who could do either job (validating etc. ASICs whether CPU or GPU) for would have been concentrating on Zen.

AMD always have those expenses, starting from Llano, all the way to Carrizo, Bristol Ridge, etc. Every single chip of theirs goes through these stages.
Zen is no different. It has the same production costs.

What it would have been? A lot more expensive, if I recall the amount of HBM 1 you could have was directly linked to the size of the bus. To get 8Gb if would have needed an 8,192 bit bus which would have been impractical and prohibitively expensive (that's a mute point because Fury must have prohibitively expensive as it was because it had a rather short production run).

Err?! Just say that with HBM, it was impossible.
I am asking what if Fiji had 12GB or 8GB?

 

[EsaT]

AMD always have those expenses, starting from Llano, all the way to Carrizo, Bristol Ridge, etc. Every single chip of theirs goes through these stages.
Zen is no different. It has the same production costs.

Developing new architecture is lot more expensive and resource consuming than something which is upgrade of old.

 

[4K8KW10]

Developing new architecture is lot more expensive and resource consuming than something which is upgrade of old.

Yes, and the initial development phases ended in 2015. Now it's 2018 and we are looking at Ryzen 3, which means Zen is already old.

 

[EsaT]

Yes, and the initial development phases ended in 2015. Now it's 2018 and we are looking at Ryzen 3, which means Zen is already old.

That's when first phase of development ended.
Hardware level/physical design and then turning that to actual working piece of silicon are entirely different things.
And of course they needed to start developing Zen2 and Zen3.

 

[4K8KW10]

Zen2 and Zen3 are polishings of the already old original Zen architecture.

What they need is to launch something graphics right now to replace Polaris and Vega. !

 

[beany_bot]

Zen2 and Zen3 are polishings of the already old original Zen architecture.

What they need is to launch something graphics right now to replace Polaris and Vega. !

Zen is "old" now!?

Wow. Things sure do move faster in tech than I thought...

 

[Gerard]

What they need is to launch something graphics right now to replace Polaris and Vega. !

Yeah i'm sure they have a brand new architecture sitting around right now they just forgot to launch to fill that gap nicely...They'll just forgo the years of design and development and plop one out of thin air.

 

[4K8KW10]

Yeah i'm sure they have a brand new architecture sitting around right now they just forgot to launch to fill that gap nicely...They'll just forgo the years of design and development and plop one out of thin air.

But why brand new architecture? Can't they fix the bottlenecks and increase the clocks of Polaris?

 

[Gerard]

But why brand new architecture? Can't they fix the bottlenecks and increase the clocks of Polaris?

Why would they want to re-release polaris?

 

[4K8KW10]

Why would they want to re-release polaris?

I think at least 20% more performance can be reached either via adding shaders/increasing the clocks/changing the memory to G5X or G6

 

[4K8KW10]

Double post.

 

[Gerard]

I think at least 20% more performance can be reached either via adding shaders/increasing the clocks/changing the memory to G5X or G6

Polaris is done, it was an architecture that was meant to be better than it was then amd came out with this "disruptive price point card" spiel. They did a similar thing with radeon 29xx series when it proved to be a bust and its whole marketing campaign was around "value".

 

[LePhuronn]

I don't believe for a minute that AMD aren't releasing anything to the gaming sector until Navi in 18 months time.

7nm respin for Vega and Hynix HBM2 could bring us an improved and more sensibly priced RX Vega. Samsung HBM2 will stay on the Instinct card.

 

[Panos]

Yes, and the initial development phases ended in 2015. Now it's 2018 and we are looking at Ryzen 3, which means Zen is already old.

Zen 2, Zen 3, Zen 4, Zen 5. Are not just die shrinks, iterations and "more cores".
Most include some new stuff that cannot be done without cutting edge manufacturing like EUV, which will see being used with Zen 3.
Already Zen+ has a lot of new things, like better XFR, IF and PB. Next iteration we have no idea what would bring except 16 core dies.
And AMD is at the start of the road. Above 16x4 (64 core), might need new topology with substrates and "double donuts" instead of IF.

 

[Panos]

I don't believe for a minute that AMD aren't releasing anything to the gaming sector until Navi in 18 months time.

7nm respin for Vega and Hynix HBM2 could bring us an improved and more sensibly priced RX Vega. Samsung HBM2 will stay on the Instinct card.

True. And is something AdoredTV hints at his RTX video, on why NV is rushing the whole lineup at this time. 7nm Vega exists since May, and it's drivers got into the Linux kernel 3 weeks ago.

 

[LePhuronn]

something AdoredTV hints at his RTX video, on why NV is rushing the whole lineup at this time

There isn't a hope in hell that AMD will topple Nvidia any time soon, but the prospect of a refined Vega consistently trampling over Nvidia's midrange segment for reasonable money is probably enough for Nvidia to push the new RTX generation and reinforce the image that AMD are still at least a generation behind and are the inferior product.

And before the usual suspects here start bleating "but AMD ARE a generation behind and inferior", who really cares if you can get your 1440p gameplay at unnecessary Hz for a respectable price? Suddenly competition comes back.

 

[4K8KW10]

Polaris is done, it was an architecture that was meant to be better than it was then amd came out with this "disruptive price point card" spiel. They did a similar thing with radeon 29xx series when it proved to be a bust and its whole marketing campaign was around "value".

Polaris 10/20 aka Ellesmere, part of the Arctic Islands family with GCN 4 is a rather small chip.
We haven't seen a upwards scaled to 4096 shaders Polaris.

Vega is Vega, and belongs to different iteration of the GCN generations, and quite possibly is weaker than a potential big Polaris.