AMD Navi at TSMC 7nm

[raghu78]

https://www.digitimes.com/news/a20171023PD201.html

AMD's Vega series GPUs are fabricated by GlobalFoundries on 14nm process, but Taiwan Semiconductor Manufacturing Company (TSMC) has won the order from AMD to fabricate its NAVI GPUs using 7nm process technology. As TSMC is also keen on making deployments in advanced packaging technologies, it will continue to maintain coopetition relationships with local OSATs.

AMD should be able to get Navi out in H1 2019 with good yields and performance as TSMC 7nm builds on the successful 10nm ramp. Hopefully AMD can move to a multi die design like EPYC to maximize yields and improve cost.

 

[el etro]

TSMC 7nm may be a better HP process, with less agressive power curves than 7HPC can offer. A big monolithic Navi at 7SoC would be another disaster.

 

[Mopetar]

Wasn't the rumor that Navi was supposed to use a similar approach to Ryzen where it's multiple smaller die connected via an interposer?

I'm curious if the move to TSMC is a result of the process being better for GPUs or GF not having enough capacity to supply AMD with CPUs and GPUs on their 7nm process. Also, not having Navi out until H1 2019 seems like it's going to be especially rough for them on the GPU side once Volta comes out in early to mid 2018.

 

[Saylick]

If AMD is going to use TSMC for Navi, which is rumored to an MCM, I wonder what Nvidia has in store for their MCM implementation, as that seems like the way forward for getting more performance per package. If TSMC's 7nm process can successfully deliver an MCM via Navi, I don't see why Nvidia would not be able to do the same.

 

[tviceman]

If that is an accurate timeline, life is going to be absolutely brutal for AMD if the crypto currency craze continues to die down. Volta will have been on the market for over a year by the time Navi ships.

 

[el etro]

Navi being monolithic and fabbed at TSMC makes more chance that Navi is the true GCN2, not another GCN 1.x .

 

[ksec]

Wasn't the rumor that Navi was supposed to use a similar approach to Ryzen where it's multiple smaller die connected via an interposer?

I'm curious if the move to TSMC is a result of the process being better for GPUs or GF not having enough capacity to supply AMD with CPUs and GPUs on their 7nm process. Also, not having Navi out until H1 2019 seems like it's going to be especially rough for them on the GPU side once Volta comes out in early to mid 2018.

Because AMD only has so much capacity at GF, and they have to decide how they allocate their stocks, if CPU is selling so well, and GF cant give them extra capacity then GPU suffers. And as of right now, for reasons I dont understand, AMD has not been able to get enough chips out from GF, and GF isn't exactly great at ramping things up.

I also wonder if the relationship with GF would end some day. Moving to TSMC seems like a much better choice.

 

[Yotsugi]

Navi being monolithic and fabbed at TSMC makes more chance that Navi is the true GCN2, not another GCN 1.x .

What's the difference, the ISA stays the same.
Besides, Vega already drastically differs from everything else made before assuming it works as intended (and right now it doesn't).

I also wonder if the relationship with GF would end some day.

Why though?

Moving to TSMC seems like a much better choice.

You're suggesting AMD chaining themselves to pet foundry of SoC vendors?
Did you forget planar 20nm shitfest?
Or cancellation of 32nm node?

 

[el etro]

What's the difference, the ISA stays the same.
Besides, Vega already drastically differs from everything else made before assuming it works as intended (and right now it doesn't).

True GCN2 means more performance per TFLOPS in gaming. A 24Tflop GCN1.x card will be ridiculous compared to future Nvidia 7nm offerings. We will see 300W GPUs from AMD competing with 150W Nvidia GPUs again...

 

[Yotsugi]

True GCN2 means more performance per TFLOPS in gaming.

Thats Vega once it works.
Are 17 tris a clock not enough for your majesty?
And let's not talk of fetch once/shade once.
You're talking about GCN without even the slightest idea what GCN is (or is not).

 

[el etro]

But in DX11 results, the card still underperform according to its capacity.

Only usimg games that rely on all of its capacities like the new WF2 can show Vega's true power.


Ins't a too risky strategy to depend on the driver team and game devs to unleash Max performance? AMD needs to pull a R300/R700 out of its hat with Navi, a card that's future proof enough yet have unleashed potential from day 1, launch day.

 

[Yotsugi]

But in DX11 results, the card still underperform according to its capacity.

Because NGG Fast Path is disabled! Unbelieable!

Only usimg games that rely on all of its capacities like the new WF2 can show Vega's true power.

But it chokes in triangle heavy scenes there.

Ins't a too risky strategy to depend on the driver team and game devs to unleash Max performance?

The other way is to bloat the die with FF units and that's going backwards considering where GPUs has been going for quite a while.

AMD needs to pull a R300/R700 out of its hat with Navi, a card that's future proof enough yet have unleashed potential from day 1, launch day.

Vega is your R300, just a little bit more complex.

a card that's future proof enough yet have unleashed potential from day 1, launch day.

They could've launched it finished by sacrificing Radeon Pro along with their SIGGRAPH keynote.
They didn't.
Enjoy the wait!

And we don't have even the slightest idea of what Navi is.
At all.

 

[Jackie60]

Navi will be late, under perform and run hot while being expensive and noisy. It’s what AMD GPUs do these days.

Flamebaiting is not allowed
Markfw
Ababdtec Moderator

 

[Yotsugi]

Navi will be late, under perform and run hot while being expensive and noisy. It’s what AMD GPUs do these days.

And you will <redacted>
It's what you do these days.

Profanity is not allowed in the technical forums
Markfw
Anandtech Moderator

 

[el etro]

Personally i think Navi would be a blast, even on DX11 games. Competitive with Volta GEN2(7nm) in all metrics including efficiency.

 

[Yotsugi]

Personally i think Navi would be a blast, even on DX11 games. Competitive with Volta GEN2(7nm) in all metrics including efficiency.

How can you think it would be competitive when absolutely no one knows what the hell it is, even.

Heck, they've yet to make Vega work for consumers!
Absolutely silly stuff.

 

[el etro]

How can you think it would be competitive when absolutely no one knows what the hell it is, even.

...

I'm saying this based on AMD key persons spokings. "Keep the eye on the ball..." "We were distant from GPUs".

That's why i personally think that Navi will be a blast. All other GCN updates was scheduled/"floorplanned" before Raja takes his step as the GPU leader at AMD. Navi may be his first developed from the scratch uarch at the company. That's why i believe in Navi.

 

[Yotsugi]

That's why i personally think that Navi will be a blast. All other GCN updates was scheduled/"floorplanned" before Raja takes his step as the GPU leader at AMD. Navi may be his first developed from the scratch uarch at the company. That's why i believe in Navi.

Vega most definetly has Raja's input.
Their new geometry pipe is pretty unconventional, reeks of Art-X legacy.

Anyway, no one has even the slightest idea what is Navi all about.

Besides, THEY NEED TO MAKE VEGA WORK FIRST.
ITS THAT SIMPLE!

 

[el etro]

Besides, THEY NEED TO MAKE VEGA WORK FIRST.
ITS THAT SIMPLE!

Vega is too ahead of its time.

 

[Yotsugi]

Vega is too ahead of its time.

No, they simply need to make it work.
Themselves.
Its that simple!

 

[el etro]

Vulkan shows how powerful is Vega, but there's not any other dev team courageous enough to explore Vega strength.

Simple: Nvidia partial DX12 supported cards have 70% of the market, and the DX11.x only Windows 7 haves 50% of steam gaming market share!

 

[IllogicalGlory]

Vulkan shows how powerful is Vega, but there's not any other dev team courageous enough to explore Vega strength.

Simple: Nvidia partial DX12 supported cards have 70% of the market, and the DX11.x only Windows 7 haves 50% of steam gaming market share!

If Vega were destroying the 1080 Ti in Wolfenstein, you might have a point, but it's matching it at best, and generally a bit slower. If it takes a courageous dev team to get Vega performing at the level it should have performed in all games, there's a problem.

 

[geoxile]

It's probably not possible to get the "max" out of Vega at the moment with the front-end it has. You can see GCN's performance per Teraflop change as you go down the models, with the RX 460 having among the highest of the recent GCN cards. In comparison, Vega's perf/TF is like half that of the RX460. Given, scaling up is never perfect but it seems particularly bad to me.

 

[el etro]

Vega suffers the same frontend bottleneck as Fiji, equal fiji. That's why i think that Vega is not the rework from scratch(As it was from R800 to GCN i.e.) that we are looking for.

 

[geoxile]

Vega suffers the same frontend bottleneck as Fiji, equal fiji. That's why i think that Vega is not the rework from scratch(As it was from R800 to GCN i.e.) that we are looking for.

Technically they claim it is with the new NGG fast path but last we got ANY information about it it's still not active though it's detailed and measured in the whitepaper.