VEGA10/VEGA11 prices

[MrTeal]

This would make an interesting poll question, but most people here seem to be assuming V10 and V11 will launch at the same time when coming up with pricing. We know that there's two HBM2 Vega GPUs coming and it seems like this will happen in 2017, but their relative timing is still a mystery. While I wouldn't put any money on it either way, I would not be at all surprised to see Vega 10 and Vega 11 separated by half a year or more.

Summer 2016 - Polaris launches with ~110mm² and 232mm²ides
Jan/1Q 2017 - Vega 11 launches, ~300-350mm², $550 and $400 or $450
Summer 2017 - Vega 10 launches ~450-500mm², $650 and $550, Vega 11 gets slotted in below it.

 

[Despoiler]

The reality is that HBM2 availability is going to drive the Vega release dates. Right now that is Q3 2016 for shipping. To me Q1 2017 is worst case scenario. I suspect AMD is going to move as fast as possible to make Vega available. Before Christmas if they can.

 

[maddie]

This would make an interesting poll question, but most people here seem to be assuming V10 and V11 will launch at the same time when coming up with pricing. We know that there's two HBM2 Vega GPUs coming and it seems like this will happen in 2017, but their relative timing is still a mystery. While I wouldn't put any money on it either way, I would not be at all surprised to see Vega 10 and Vega 11 separated by half a year or more.

Summer 2016 - Polaris launches with ~110mm² and 232mm²ides
Jan/1Q 2017 - Vega 11 launches, ~300-350mm², $550 and $400 or $450
Summer 2017 - Vega 10 launches ~450-500mm², $650 and $550, Vega 11 gets slotted in below it.

What about Navi?

Noting the words scalable used together with nexgen memory implies, to me at least, lower performance to higher performance more easily achieved.

At that point Polaris would be quite old and if you're replacing it with low end Navi, you'll also have a high end Navi.

In other words, the last Vega model would have too short a sales lifetime if released too late as Navi appears to be around 1 year after Vega.

 

[Mopetar]

Depends on the competition.

/thread

It entirely depends on what NVidia currently has or has announced for the immediate future when AMD launches. If they've got something that's a cut above the 1080 (could easily happen if Vega does clock in at ~450 mm^2 and GP104 is the biggest consumer die NVidia has for the time) then expect AMD to price their card as though it were an NV flagship product.

If the market will pay $1000 for the premium (which it's been shown to do) then that's what they'll charge. If NV releases a consumer GP100 (or develops a GP102 to go head to head with AMDs biggest card) then expect more competitive prices.

 

[MrTeal]

What about Navi?

Noting the words scalable used together with nexgen memory implies, to me at least, lower performance to higher performance more easily achieved.

At that point Polaris would be quite old and if you're replacing it with low end Navi, you'll also have a high end Navi.

In other words, the last Vega model would have too short a sales lifetime if released too late as Navi appears to be around 1 year after Vega.

What about Navi? The roadmap presented though far seemed to indicate it might be a 2018 product, though that obviously has a huge amount of uncertainty to it. Even then, it's a product slated for 2+ years out. We might see Navi play out in a similar way to Fiji, with it launching at the end of the 14nm generation prior to the move to a new node.

Polaris won't exactly be that old by the middle of 2018, anyway. GCN1 launched Jan 2012, and GCN2 (Bonaire) didn't launch until March 2013. The first really interesting GCN2 chip (Hawaii) was launched almost 22 months after GCN1, so prior that pretty much everything that AMD sold was a single GCN generation. nVidia had a similar timeframe, with almost two years between the launch of Kepler and (the low power) Maxwell 1, and two and a half years between Kepler and Maxwell 2.

 

[maddie]

What about Navi? The roadmap presented though far seemed to indicate it might be a 2018 product, though that obviously has a huge amount of uncertainty to it. Even then, it's a product slated for 2+ years out. We might see Navi play out in a similar way to Fiji, with it launching at the end of the 14nm generation prior to the move to a new node.

Polaris won't exactly be that old by the middle of 2018, anyway. GCN1 launched Jan 2012, and GCN2 (Bonaire) didn't launch until March 2013. The first really interesting GCN2 chip (Hawaii) was launched almost 22 months after GCN1, so prior that pretty much everything that AMD sold was a single GCN generation. nVidia had a similar timeframe, with almost two years between the launch of Kepler and (the low power) Maxwell 1, and two and a half years between Kepler and Maxwell 2.

Yes, but when last did you see AMD announce the annual release of GPU architectures?

Polaris architecture, Vega architecture and Navi architecture. I'm not sure recent history is an accurate guide for the future. We might be returning to more rapid change in GPUs as used to exist.

Koduri said 4K x 4K x 240Hz with HDR is the long term goal and that using Moore's law is not enough. More rapid architectural and software changes are needed and coming.

 

[MrTeal]

Yes, but when last did you see AMD announce the annual release of GPU architectures?

Polaris architecture, Vega architecture and Navi architecture. I'm not sure recent history is an accurate guide for the future. We might be returning to more rapid change in GPUs as used to exist.

Koduri said 4K x 4K x 240Hz with HDR is the long term goal and that using Moore's law is not enough. More rapid architectural and software changes are needed and coming.

I'm confused by your position then. If you are suggesting that there will be a return to more rapid GPU uArch cadence, why would Vega 10 being released summer '17 with Navi being released sometime in 2018 be unexpected. Yes the lower end of the scale with Polaris will be older at that point, but as long as its efficiency is good that's not really an issue. AMD was still selling VLIW5 GPUs last year as the R5 230 half a decade after Terascale 2 was introduced. They're still a current SKU at Newegg. If Navi does bring a new memory technology, I wouldn't be surprised to see P11 and GDDR5 still in AMD's product stack beyond 2018.

 

[maddie]

I'm confused by your position then. If you are suggesting that there will be a return to more rapid GPU uArch cadence, why would Vega 10 being released summer '17 with Navi being released sometime in 2018 be unexpected. Yes the lower end of the scale with Polaris will be older at that point, but as long as its efficiency is good that's not really an issue. AMD was still selling VLIW5 GPUs last year as the R5 230 half a decade after Terascale 2 was introduced. They're still a current SKU at Newegg. If Navi does bring a new memory technology, I wouldn't be surprised to see P11 and GDDR5 still in AMD's product stack beyond 2018.

I'm also confused. Maybe I'm not explaining my thoughts well enough.

Taking AMD slides at face value, we have Navi as scalable, meaning they will design small and scale to higher levels. I see this as saying Navi can have a top to bottom release fairly close together. The large perf/watt increase alone will relegate previous cards to a 2nd class status even if still available on Newegg. Coming 1 year after 1st Vega card does not give enough time for a middle year Vega release. Only 6-8 months before being overshadowed by Navi.

Any clearer?

 

[MrTeal]

I'm also confused. Maybe I'm not explaining my thoughts well enough.

Taking AMD slides at face value, we have Navi as scalable, meaning they will design small and scale to higher levels. I see this as saying Navi can have a top to bottom release fairly close together. The large perf/watt increase alone will relegate previous cards to a 2nd class status even if still available on Newegg. Coming 1 year after 1st Vega card does not give enough time for a middle year Vega release. Only 6-8 months before being overshadowed by Navi.

Any clearer?

Well yes, if you assume that Navi launches in Q1 2018, vs say fall 2018. If Navi is a H2 or Q4 product, even a summer Vega 10 launch would give a year on market prior to Navi arriving.
We still know nothing about Navi other than the couple descriptors on those roadmaps. Next Gen memory for instance could mean something even higher BW than HBM2, or it could mean Navi will use a type of next gen memory, perhaps GDDR5X or HBM2. Same deal with scalable. If it truly is a CoWoS only product, I don't think we'll be seeing it in the true low end for awhile, so P11 will likely still be in the mix.

 

[maddie]

Well yes, if you assume that Navi launches in Q1 2018, vs say fall 2018. If Navi is a H2 or Q4 product, even a summer Vega 10 launch would give a year on market prior to Navi arriving.
We still know nothing about Navi other than the couple descriptors on those roadmaps. Next Gen memory for instance could mean something even higher BW than HBM2, or it could mean Navi will use a type of next gen memory, perhaps GDDR5X or HBM2. Same deal with scalable. If it truly is a CoWoS only product, I don't think we'll be seeing it in the true low end for awhile, so P11 will likely still be in the mix.

All true. Many assumptions were made.

Still, I'll bet even odds that with increased APU performance levels, P11 might be surpassed by then.

 

[MrTeal]

All true. Many assumptions were made.

Still, I'll bet even odds that with increased APU performance levels, P11 might be surpassed by then.

Maybe. I don't see it getting there any time soon though. P11 should be faster than a 960, but even with dual channel DDR4 overclocked to ludicrous speeds Skylake has half the bandwidth of a 960.

AMD might eventually do some custom APU with HBM onboard that deals with the massive bandwidth issues an APU has, but that will probably be a pretty specialized and price option much like Iris Pro is now.

 

[alcoholbob]

It depends entirely on Titan performance since Nvidia will almost certainly launch a Geforce Titan before AMD will launch a flagship card. It will also have to be at a steep discount performance to price wise. It's the cost of always being second to market.

 

[maddie]

Maybe. I don't see it getting there any time soon though. P11 should be faster than a 960, but even with dual channel DDR4 overclocked to ludicrous speeds Skylake has half the bandwidth of a 960.

AMD might eventually do some custom APU with HBM onboard that deals with the massive bandwidth issues an APU has, but that will probably be a pretty specialized and price option much like Iris Pro is now.

Something has to give as there is already bandwidth problems with APUs. Do they stop improving and returning the market to the lowest end GPUs, lower than Polaris 11?

A 1 stack HBM2 solution? Tiny interposer not much bigger than the sum of the 2 or 3 components?

Something for me to think about.

 

[Slaughterem]

Well yes, if you assume that Navi launches in Q1 2018, vs say fall 2018. If Navi is a H2 or Q4 product, even a summer Vega 10 launch would give a year on market prior to Navi arriving.
We still know nothing about Navi other than the couple descriptors on those roadmaps. Next Gen memory for instance could mean something even higher BW than HBM2, or it could mean Navi will use a type of next gen memory, perhaps GDDR5X or HBM2. Same deal with scalable. If it truly is a CoWoS only product, I don't think we'll be seeing it in the true low end for awhile, so P11 will likely still be in the mix.

I think I found the answer to next gen HBM2 memory for NAVI. This is a very interesting idea for cooling of stacked DRAM chips. It appears that HBM2 and 1 are probably limited due to cooling of a stack where the top chip cools better than the ones below. Pretty logical if you think of it and if AMD has now a patent for this they may be on verge of cornering the market later on with this. Stacked semiconductor chip device with phase change material
http://patents.justia.com/patent/9331053

 

[airfathaaaaa]

I'm confused by your position then. If you are suggesting that there will be a return to more rapid GPU uArch cadence, why would Vega 10 being released summer '17 with Navi being released sometime in 2018 be unexpected. Yes the lower end of the scale with Polaris will be older at that point, but as long as its efficiency is good that's not really an issue. AMD was still selling VLIW5 GPUs last year as the R5 230 half a decade after Terascale 2 was introduced. They're still a current SKU at Newegg. If Navi does bring a new memory technology, I wouldn't be surprised to see P11 and GDDR5 still in AMD's product stack beyond 2018.

vega will be the enthusiast level gcn 4.0 i thought everyone agreed on this also i dont think neither amd or nvidia want to repeat the same mistake amd did with hbm1 and not having enough cards to sell because of it
also i dont think the main feature of navi would be the next gen memory but the scalability of it given that the next consoles will be native CF i would assume they will more likely start producing dual gpu cards for the masses

 

[Despoiler]

I think I found the answer to next gen HBM2 memory for NAVI. This is a very interesting idea for cooling of stacked DRAM chips. It appears that HBM2 and 1 are probably limited due to cooling of a stack where the top chip cools better than the ones below. Pretty logical if you think of it and if AMD has now a patent for this they may be on verge of cornering the market later on with this. Stacked semiconductor chip device with phase change material
http://patents.justia.com/patent/9331053

You are on the right path, but still thinking too small. Remember that HBM is 2.5D solution. With the node jumps bringing down the power requirements, and thus heat, a cooling solution like phase change could feasibly allow the next logical step to happen. AMD could go 3D packaging. The memory sits on top of the GPU or GPUs. They would transition their interposer from GPU to memory communication to GPU to GPU communication. Mind blown?

 

[Slaughterem]

You are on the right path, but still thinking too small. Remember that HBM is 2.5D solution. With the node jumps bringing down the power requirements, and thus heat, a cooling solution like phase change could feasibly allow the next logical step to happen. AMD could go 3D packaging. The memory sits on top of the GPU or GPUs. They would transition their interposer from GPU to memory communication to GPU to GPU communication. Mind blown?

Or remove the interposer to stack dram on APU.

 

[Despoiler]

Or remove the interposer to stack dram on APU.

Right. I qualified my statement as GPU to GPU because that requires an interposer. A single APU would not.