[Rumor, Tweaktown] AMD to launch next-gen Navi graphics cards at E3

[NostaSeronx]

More akin to what nV does, really.
But yes, the SIMDs are completely different.

Nvidia is really weird about it;
- Compiler-based instruction.r__-reuse
- 4 operand reuse caches to each 4 register banks each with 4 operand slots.

AMD way is;
- Scheduler-based, no new instructions
- Each ALU has a source buffer which can hold 6 source operands and a VDST which can hold 8 destination operands.
- SIMD16 thus has 96 source register reuses and 128 destination register reuses.
- Destination registers can be reused as source registers for dependent operations.

Which could explain why a 40 CU(66AF:F1)/36CU(66AF:F0) is as big as the Radeon VII die.

 

[Glo.]

Which could explain why a 40 CU(66AF:F1)/36CU(66AF:F0) is as big as the Radeon VII die.

Nope . The GPU you are quoting has 20 CUs

 

[Yotsugi]

Navi appears to be much more Graphics focused architecture than Compute.

Not until they've touched their FF stuff.
It's just better SIMDs so far.
And yes, it does degrade ALU density in a very not fun way.

Nvidia is really weird about it;
- Compiler-based instruction.r__-reuse
- 4 operand reuse caches to each 4 register banks each with 4 operand slots.

AMD way is;
- Scheduler-based, no new instructions
- Each ALU has a source buffer which can hold 6 source operands and a VDST which can hold 8 destination operands.
- SIMD16 thus has 96 source register reuses and 128 destination register reuses.
- Destination registers can be reused as source registers for dependent operations.

Which could explain why a 40 CU(66AF:F1)/36CU(66AF:F0) is as big as the Radeon VII die.

As in superscalarity, not opreuse.

 

[Glo.]

Not until they've touched their FF stuff.
It's just better SIMDs so far.
And yes, it does degrade ALU density in a very not fun way.

FF? Front end?

 

[Yotsugi]

I called V20 a GK210 of sorts, and now you should be getting why.

FF? Front end?

Fixed-function.
Geometry, setup, ROPs, all that stuff.

 

[Yotsugi]

Bam, Navi to be delivered Q3 per AMD ER, as I've said.

 

[DisEnchantment]

Bam! Q3 it is!!
https://www.anandtech.com/show/14286/amd-7nm-navi-consumer-gpu-to-launch-in-q3

 

[AtenRa]

From the AMD Q1 2019 earnings call,


https://seekingalpha.com/article/42...2019-results-earnings-call-transcript?page=10

Q – Hans Mosesmann

Hey thanks. Lisa, regarding Navi, if you comment on it what is the positioning of that particular product relative to the -- your current 7-nanometer GPU? And regarding Navi, can you tell us if it's going to include rate tracing?

A – Lisa Su

Yeah, so, Hans, we are excited about Navi. Navi is a new architecture for us in gaming. It has a lot of new features, across the Navi architecture. Things are progressing well. We expect it to launch in the third quarter.

From a positioning standpoint, I probably won't go through it in great detail right now other than to say that it is 7-nanometer, Navi, but it will be positioned below where, for example, our Radeon VII's is position today from a price point standpoint.

And then, in terms of Ray tracing, again, we will talk more about our overall Navi roadmap as we get closer to the launch.

 

[Head1985]

So at the end of Q3 will be launch?They always launch GPU at the end of launch date(polaris, vega).Year after turing.I smell another vega fail, but i hope for AMD sake it will be atleast decent.
I just hope it wont be like vega-paper launch at the end of launch date with reviews in next month and cards in shops in yet another month....( so for navi in november/december in shops)

 

[Krteq]

Where have you seen info about the END of the Q3?

 

[Ajay]

From the AMD Q1 2019 earnings call,


https://seekingalpha.com/article/42...2019-results-earnings-call-transcript?page=10

it is interesting that Su talks about a Navi *roadmap*. I hold out more hope now that there will be at least two distinct GPU dies (in addition to cut downs).

 

[Yotsugi]

So at the end of Q3 will be launch?

No.

Year after turing.I smell another vega fail, but i hope for AMD sake it will be atleast decent.

You're smelling the wrong winds.

it is interesting that Su talks about a Navi *roadmap*

3 dies total.

 

[Glo.]

Year after turing.I smell another vega fail, but i hope for AMD sake it will be atleast decent.

Whats so special about Turing, that Navi, new GPU arch, on new node cannot at least compete with it?

 

[Yotsugi]

Whats so special about Turing, that Navi, new GPU arch, on new node cannot at least compete with it?

Nvidia.
Nvidia makes it special.

 

[Head1985]

Whats so special about Turing, that Navi, new GPU arch, on new node cannot at least compete with it?

We dont know anything about navi other than its 7nm.
And turing its pretty good arch.It brings 15-20% IPC vs pascal at same node+initial RT support.Its just overpriced.

 

[ozzy702]

Whats so special about Turing, that Navi, new GPU arch, on new node cannot at least compete with it?

7nm wasn't impressive for Radeon VII, and AMD hasn't exactly executed well over the past several generations so I think there's room for some healthy skepticism when it comes to NAVI. AMD has had "new GPU arch" time after time which amounted to very little so NAVI would need to be a massive improvement even on 7nm just to catch Turing in performance/watt.

Where I do expect to be pleased is in price/performance because that's where AMD has always shined (whether they'd like to or not) and I expect AMD to compete very well in this metric with NVIDIA's bloated prices. I'm hopeful, but I'm not going to get sucked into the AMD hype train that happens before every GPU release only to be disappointed when reality hits.

 

[Yotsugi]

We dont know anything about navi other than its 7nm.

Which in itself is incredible.
Also you do, look up the patches.

7nm wasn't impressive for Radeon VII

Because the implementation isn't impressive either.

 

[Stuka87]

I am glad Lisa stated that Navi would fall below Vega VII in price. She did not mention performance (Which is to be expected, that will come later), but if it matches vega VII for less, thats a win for consumers.

 

[Guru]

7nm wasn't impressive for Radeon VII, and AMD hasn't exactly executed well over the past several generations so I think there's room for some healthy skepticism when it comes to NAVI. AMD has had "new GPU arch" time after time which amounted to very little so NAVI would need to be a massive improvement even on 7nm just to catch Turing in performance/watt.

Where I do expect to be pleased is in price/performance because that's where AMD has always shined (whether they'd like to or not) and I expect AMD to compete very well in this metric with NVIDIA's bloated prices. I'm hopeful, but I'm not going to get sucked into the AMD hype train that happens before every GPU release only to be disappointed when reality hits.

WTF are you talking about? They've essentially taken a cut down Vega 64, essentially a Vega 60 at the same power target and increased performance by 35%. Less CU's, same power consumption, 8GB more HBM memory at 35% performance increase! That is very impressive and that is with MINOR architectural changes, Navi is a new architecture that is developed alongside 7nm process, a more mature process as well than when AMD released their MI professional GPU's last year!

 

[Ajay]

3 dies total.

Well, that would be impressive. I just have doubts that AMD can raise enough cash to spin that many dies. They can’t take money from Zen development because the cadence is, and needs to be, high.

 

[Yotsugi]

Well, that would be impressive

Why?

I just have doubts that AMD can raise enough cash to spin that many dies

They, like, ramped the R&D for 8q's straight.

They can’t take money from Zen development because the cadence is

Zen is now chiplets, way less dies per lineup.

 

[Ajay]

Why?

How many iterations of Polaris were there?

They, like, ramped the R&D for 8q's straight.

Almost all of it on CPU.

Zen is now chiplets, way less dies per lineup.

Naples and Summit shared the same die, so does Rome and Matisse. They do now have separate IO dies, so that’s actually an added expense (including packaging expenses).

 

[Yotsugi]

How many iterations of Polaris were there?

3 dies, too.

Almost all of it on CPU.

No, not even close, my dear.

They do now have separate IO dies, so that’s actually an added expense (including packaging expenses).

AMD also made 2 14nm APUs, being RR and RAVEN2.

 

[Ajay]

3 dies, too.

No, not even close, my dear.

AMD also made 2 14nm APUs, being RR and RAVEN2.

1. Fair enough, I forgot the 590.
2. Seriously, you think the majority of R&D went somewhere other than CPU. Where then?
3. Well, if 3rd gen APU goes with a Matisse chiplet, you will have a point - AFAIK, that’s just speculation right now (and wouldn’t make any sense for ultrabooks, etc. ).

Feel free to post links showing me where I’m wrong.

 

[Yotsugi]

Fair enough, I forgot the 590.

Wait, no, FOUR.
Polaris10/20 (570/580) Polaris11 (560&friends) Polaris12 (550) and Polaris30.

Seriously, you think the majority of R&D went somewhere other than CPU. Where then?

It's spread pretty evenly, maybe weighted a bit more on GPU side this cycle.

Well, if 3rd gen APU goes with a Matisse chiplet

The 45W one, yes (Renoir).
Dali is monolithic SoC.