AMD: Speculation, Rumors, and Discussion (Archive)

[CarstenS]

I don't think so. And HBM wouldn't bring enough power savings to bring that large of a bump in perf/watt compared to Polaris.

Apply the same math AMD did with Fiji and you might get there (didn't actually do it, just a wild guess).

 

[Silent_Buddha]

Apply the same math AMD did with Fiji and you might get there (didn't actually do it, just a wild guess).

The implication would be that HBM allowed Fiji based products to be ~2x higher perf/watt than the base GCN GPU used in that slide.

I don't think AMD claimed anything close to that, did they?

Regards,
SB

 

[no-X]

AMD's roadmap showed Polaris 2,5× above Hawaii and Vega 3,9× above Hawaii. That means Vega (HBM) is 1,5× above Polaris (GDDR5) just like Fiji (HBM) was 1,5× above Hawaii (GDDR5).

- Hawai/Tonga: 1,0× (base)
- Fiji (Fury X): 1,5×
- Fiji (Nano): 2,0×
- Polaris: 2,5×
- Vega: 3,9×
- Navi: 5,1×

All these numbers are taken (or extrapolated) from official slides and roadmaps.

 

[firstminion]

That's a train of thoughts I just cannot follow. Why would I as a consumer, interested in purchasing an enthusiast-level card (Vega), buy a mid-range card now for 200 bucks and 6-8 months later the card I originally desired? Being interested in an enthusiast-grade card, chances are, I went for that target as well during my last upgrade cycle, which might put me in the R9 290X/390 range - and coming from there RX480 is - after all that is known today - merely a sidegrade.

Christmas bonus? Desire for HBM2? Give yourself an upgrade and the 480 to your little brother? I don't know, what the heck... It's just more logical that some buyers of Polaris could want to upgrade to Vega. While a buyer of GP104 just won't buy a GP106 when it launches.

When talking about process maturity, a delay of a couple of months is of much lower importance than the fact that they're completely different processes from different fabs...

I'm comparing with GP104, Vega should be close to 2 quarters later (needs to wait for HBM2). To my knowledge that would be relevant.

So do you expect Big Vega by august?

As others have mentioned there will be a GP102 (IMO designed to span Tesla/Quadro/Titan-ti model in similar trend as the GK110 GPU) - fingers crossed this one has a GPC/SM/Cuda structure more similar to P100 than Maxwell.
Also more an assumption but I would expect Volta to release initially as a large die Tesla, similar with what we saw with Pascal; this is due to many aspects being integral to both HPC and scientific research supercomputers, where they have contract commitments for 2017 such as with the US government DOE.

Cheers

So GP102 then, I wasn't aware GP100 has HBM2.

 

[xEx]

According to AMD's slides, Vega has much higher perf/watt than Polaris.

Regards,
SB

Maybe but why would they use a more much expensive Vega chip with HBM just to cut down its power to make it into the Polaris territory instead of just use the more cheaper Polaris chip that already produce 5.6(?)TFPs I also dont think Xbox is going to have a mass production vega in December while consumers wont even have an announcement date for their release

 

[xEx]

That's a train of thoughts I just cannot follow. Why would I as a consumer, interested in purchasing an enthusiast-level card (Vega), buy a mid-range card now for 200 bucks and 6-8 months later the card I originally desired? Being interested in an enthusiast-grade card, chances are, I went for that target as well during my last upgrade cycle, which might put me in the R9 290X/390 range - and coming from there RX480 is - after all that is known today - merely a sidegrade.

If the explicit multi adapter(was that its name?) gets very popular between developers then you could have a Vega + a 5TFPs card helping with post-processing and or filters.

I dont know how popular that feature will be but damn I hope they make a good use of it. it just open so much possibilities.

 

[BRiT]

ArsTechnical Article: http://arstechnica.com/gaming/2016/06/amd-announces-two-more-polaris-video-cards-rx-470-rx-460/

Specs, prices, and release dates for these two Polaris cards were not formally announced at the event, but given the RX 480 is going to retail for $199, they're going to be light on the wallet. Su also brought on a representative to walk around the stage wearing a backpack complete with an Alienware PC installed inside—backpack PCs are something of a trend right now — which could actually be light and comfortable if AMD pulls off a small form factor with the "disruptive" $199 RX 480 card packed inside.

 

[CarstenS]

Christmas bonus? Desire for HBM2? Give yourself an upgrade and the 480 to your little brother? I don't know, what the heck...

I still don't see the point. Given that I'm not coming from an integrated graphics, there's nothing apart from a broken graphics card that would lead me to upgrade twice a year. But maybe I'm the exception to the rule.

It's just more logical that some buyers of Polaris could want to upgrade to Vega. While a buyer of GP104 just won't buy a GP106 when it launches.

Wait - you are not saying AMD is screwing over their customers by leading them to upgrade twice in a row?

 

[xEx]

I dont understand, didn't AMD said that Polaris was aimed for the market of 100 to 300 dollars? then where is the 300 dollars polaris?

 

[xEx]

I still don't see the point. Given that I'm not coming from an integrated graphics, there's nothing apart from a broken graphics card that would lead me to upgrade twice a year. But maybe I'm the exception to the rule.


Wait - you are not saying AMD is screwing over their customers by leading them to upgrade twice in a row?

Only if your plating / ***hoping*** to use the explicit multi adapter of DX12 or having more than 1 PC you could possibly think in buying 2 different segments video cards. People who will buy Polaris wont Buy vega.

 

[CarstenS]

Hm, yeah. DX12 Multi Adapter may present a reason. If there's software taking sufficient advantage of it.

 

[Silent_Buddha]

Maybe but why would they use a more much expensive Vega chip with HBM just to cut down its power to make it into the Polaris territory instead of just use the more cheaper Polaris chip that already produce 5.6(?)TFPs I also dont think Xbox is going to have a mass production vega in December while consumers wont even have an announcement date for their release

It isn't using HBM. It's using GDDR5 (appears to be going for 12 GB going by the render at the presentation). And otherwise I don't see how it's even possible for them to make Scorpio, it'd be approaching 200 watts or more for the entire console. For comparison launch X360 used slightly less than 180 watts when gaming and that was considered . This would put Scorpio at a higher power consumption than that. I guess not impossible that they would go for that, but I find it highly unlikely.

Regards,
SB

 

[ImSpartacus]

Does the 470 have 8 memory packages? Would that signify a 256-bit bus (i.e. Polaris 10)?

 

[CarstenS]

That's the 8 GiB version though.

 

[CarstenS]

Yes and yes.

 

[xEx]

It isn't using HBM. It's using GDDR5 (appears to be going for 12 GB going by the render at the presentation). And otherwise I don't see how it's even possible for them to make Scorpio, it'd be approaching 200 watts or more for the entire console. For comparison launch X360 used slightly less than 180 watts when gaming and that was considered . This would put Scorpio at a higher power consumption than that. I guess not impossible that they would go for that, but I find it highly unlikely.

Regards,
SB

In my opinion it would be easier and better to use P10 with HBM than Vega with GDDR. don't forget that a huge part of Vega pref/watt is the HBM use and not just the architecture.

I also dont see how much profit AMD could get to sell a new big chip aim for 400+ cards for a 300 console.

 

[Silent_Buddha]

AMD's roadmap showed Polaris 2,5× above Hawaii and Vega 3,9× above Hawaii. That means Vega (HBM) is 1,5× above Polaris (GDDR5) just like Fiji (HBM) was 1,5× above Hawaii (GDDR5).

- Hawai/Tonga: 1,0× (base)
- Fiji (Fury X): 1,5×
- Fiji (Nano): 2,0×
- Polaris: 2,5×
- Vega: 3,9×
- Navi: 5,1×

All these numbers are taken (or extrapolated) from official slides and roadmaps.

I don't know what slide you are looking at but the slide here (http://wccftech.com/amd-powerful-c99-polaris-gpu-vega-10-test-vehicle/ ) has it at ~4.524x the base GPU using Polaris at 2.5x as a reference point. Or ~1.8095x Polaris. My using ~4.5x is close enough to use for easy calculations.

Regards,
SB

 

[Silent_Buddha]

In my opinion it would be easier and better to use P10 with HBM than Vega with GDDR. don't forget that a huge part of Vega pref/watt is the HBM use and not just the architecture.

I also dont see how much profit AMD could get to sell a new big chip aim for 400+ cards for a 300 console.

You don't get it. They aren't selling a big chip. Vega and Polaris are both modular architectures that can be used in semi-custom designs (like PlayStation and Xbox). PS4 Neo, for instance, isn't coming with either Polaris 10 or Polaris 11. It's coming with a custom configuration for PS4 Neo that uses the Polaris architecture.

If XBO-T uses Vega it wouldn't be using either Vega 10 or Vega 11 but a custom configuring using the Vega architecture.

Regards,
SB

 

[Esrever]

Would Vega only have HBM or will it be with DDR5x or something for the xbox?

 

[Silent_Buddha]

Would Vega only have HBM or will it be with DDR5x or something for the xbox?

I think it was mentioned that it had 320 GB/s. The render of the PCB showed 12 memory chips, so presumably a 384 bit memory interface. So GDDR5 is the most likely candidate.

Regards,
SB