Vega/Navi Rumors (Updated)

[Sweepr]

Vega Die Shot

 

[raghu78]

Vega Die Shot

thanks for the die shot.Can anybody come up with rough die size estimates for Vega based on HBM2 package size .

HBM2 package size = 7.75 mm x 11.87 mm

http://www.anandtech.com/show/10527/sk-hynix-adds-hbm2-4-gb-memory-q3

All HBM2 memory products from the company will have 1.2 V VDD/VDDQ and will come in new packages featuring 7.75 mm × 11.87 mm (91.99 mm2) dimensions.

Edit : Looks like its between 530-560 sq mm . Right in between GP102 and GP100

http://videocardz.com/65477/amd-vega-gpu-pictured-features-two-hbm2-stacks

Vega is a direct competitor to GP100 in the HPC space and has support for 1/2 precision FP64 rate so that explains such a large die size. I hope AMD has made significant improvements to perf/watt and perf/sqmm. Thats been their glaring problem in the last few years against Nvidia GPUs.

 

[laamanaator]

Edit : Looks like its between 530-560 sq mm . Right in between GP102 and GP100

Yep, I got 553mm^2 based on that picture

 

[Qwertilot]

Think they're likely to struggle for strict perf/watt etc when gaming - look how much extra die space/heat etc GP100 uses up with its extra compute stuff.

NV can afford to run two distinct product lines for gaming/compute but AMD definitely can't and where they had to make a decision for compute vs gaming, they'll likely have given priority to compute for these massive chips.

 

[ShintaiDK]

So this is big Vega. ~550mm2 250W or so and only 2 stacks, so 8GB.

 

[AtenRa]

Think they're likely to struggle for strict perf/watt etc when gaming - look how much extra die space/heat etc GP100 uses up with its extra compute stuff.

NV can afford to run two distinct product lines for gaming/compute but AMD definitely can't and where they had to make a decision for compute vs gaming, they'll likely have given priority to compute for these massive chips.

They are using HBM2 so i believe a new Vega Nano at 175W TDP will be faster than GTX 1080 and could reach close to 1080Ti performance in DX-12 games. Fiji Nano is already neck and neck with GTX 980Ti in the majority of latest 2016 games.

 

[Krteq]

So this is big Vega. ~550mm2 250W or so and only 2 stacks, so 8GB.

4, 8 and even 16GBs are possible

 

[ShintaiDK]

They are using HBM2 so i believe a new Vega Nano at 175W TDP will be faster than GTX 1080 and could reach close to 1080Ti performance in DX-12 games. Fiji Nano is already neck and neck with GTX 980Ti in the majority of latest 2016 games.

That's as likely as your Polaris claim.

 

[ShintaiDK]

4, 8 and even 16GBs are possible

There are no 8Hi stacks yet, just as they failed with HBM1. Also just because its possible doesn't mean you will see it. If lucky you may see 16GB as HPC one day. It looks like AMD repeated the Fiji memory mistake. GP100 uses 4 stacks for this exact reason.

 

[AtenRa]

That's as likely as your Polaris claim.

This has turned out to be very very close in late 2016

 

[ShintaiDK]

http://techreport.com/review/31224/the-curtain-comes-up-on-amd-vega-architecture/4

Techreport estimate performance to be 1070-1080. Vega is a GP104 competitor from the looks of it.

 

[Sven_eng]

http://techreport.com/review/31224/the-curtain-comes-up-on-amd-vega-architecture/4

Techreport estimate performance to be 1070-1080. Vega is a GP104 competitor from the looks of it.

Quoting for easy access to future signature.

 

[AtenRa]

http://techreport.com/review/31224/the-curtain-comes-up-on-amd-vega-architecture/4

Techreport estimate performance to be 1070-1080. Vega is a GP104 competitor from the looks of it.

Its funny they believe a 500-550mm2 VEGA die with an improved architecture over Polaris will be a competitor to 300mm2 GP104 when a 232mm2 Polaris 10 is faster than a 200mm2 GP106 in many of the latest games.

edit: A 330-350mm2 Vega die could be the competitor to GP104 we are waiting for.

 

[ShintaiDK]

Its funny they believe a 500-550mm2 VEGA die with an improved architecture over Polaris will be a competitor to 300mm2 GP104 when a 232mm2 Polaris 10 is faster than a 200mm2 GP106 in many of the latest games.

How was all the Polaris claims? Fiji performance, GTX980TI killer and what not. 75-100W and so on?

Vega looks to be like GP100. A HPC oriented GPU.

Polaris 10 uses 33% more bandwidth, 30% more transistors and 37% more power to compete with GP106.

Quoting for easy access to future signature.

Just save the Techreport link since its their estimation. Much easier.

 

[AtenRa]

How was all the Polaris claims? Fiji performance, GTX980TI killer and what not.

Polaris 10 uses 33% more bandwidth, 30% more transistors and 37% more power to compete with GP106.



Just save the Techreport link since its their estimation. Much easier.

Well one exception, VEGA is not Polaris. Even a 350mm2 Polaris die could easily compete against 300mm2 GP104 Pascal, im sure you know deep down that a 500-550mm2 HBM2 VEGA is GP102/100 competitor but you dont like to acknowledge it

 

[Det0x]

Looks very promising

Noticed its not 2x throughput per clock like the cloud image but the slide says"over" 2x

Geometry throughput slide : Data base on AMD engineering design of Vega.Radeon R9 Fury X has 4 geometry engines and a peak of 4 polygons per clock.Vega is designed to handle up to 11 polygons per clock with 4 geometry engines.This represents an increase of 2.6x

 

[ShintaiDK]

Well one exception, VEGA is not Polaris. Even a 350mm2 Polaris 10 die could easily compete against 300mm2 GP104 Pascal, im sure you know deep down that a 500-550mm2 HBM2 VEGA is GP102/100 competitor but you dont like to acknowledge it

Feel free to prove it.

 

[badb0y]

How was all the Polaris claims? Fiji performance, GTX980TI killer and what not. 75-100W and so on?

Vega looks to be like GP100. A HPC oriented GPU.

Polaris 10 uses 33% more bandwidth, 30% more transistors and 37% more power to compete with GP106.



Just save the Techreport link since its their estimation. Much easier.

Front page article mentions that the tile rasterization that nVidia used to get those power efficiency gains are implemented in Vega so using Polaris as a reference point is pointless.

 

[AtenRa]

There are no 8Hi stacks yet, just as they failed with HBM1. Also just because its possible doesn't mean you will see it. If lucky you may see 16GB as HPC one day. It looks like AMD repeated the Fiji memory mistake. GP100 uses 4 stacks for this exact reason.

GP100 is a HPC only, GP102 doesnt even use HBM2

And 8GB is fine for 4K, im sure it will be fine in 2017-2018.

 

[AtenRa]

Feel free to prove it.

If

232mm2 Polaris = Equal/faster than 200mm2 Pascal

Then

350mm2 Polaris = Equal/faster than 300mm2 Pascal

 

[ShintaiDK]

If

232mm2 Polaris = Equal/faster than 200mm2 Pascal

Then

350mm2 Polaris = Equal/faster than 300mm2 Pascal

We already established your home math doesn't work out. That's why I asked you to prove it

And you need 2x232mm2 5.7 billion transistors to match a GP104 dont you. Or 11.4 billion transistors to match 7.2 billion. And do I have to mention the power usage?

 

[zinfamous]

Would have liked more specifics, but it didn't stop the shameless FUD spreading from the other camp, despite the equal lack of details from their perspective. It's going to be a looooooong ~3 months+

 

[AtenRa]

We already established your home math doesn't work out. That's why I asked you to prove it

You have established that, that doesnt mean its correct though

 

[zinfamous]

We already established your home math doesn't work out. That's why I asked you to prove it

He gave you a proof. It's called humor. Try to have fun.

 

[Qwertilot]

Well one exception, VEGA is not Polaris. Even a 350mm2 Polaris 10 die could easily compete against 300mm2 GP104 Pascal, im sure you know deep down that a 500-550mm2 HBM2 VEGA is GP102/100 competitor but you dont like to acknowledge it

Hummm. Even GP100 isn't a very good GP102 competitor for gaming purposes

Near identical gaming performance I think? GP100 is significantly larger (610mm2 vs 471mm2), more power draw - so I guess slower once you really clock GP102 - no doubt a lot more expensive to make etc.

It really is a non trivial built in handicap for AMD to try and overcome this, on top of everything else. Hopefully vega is least a success in the compute market as that could get them some much needed income.