Polaris 10 size / performance estimation

[Nerdmaster]

Please guys no thread derailing dont want this thread closed before polaris gets released.

[NTMBK]

I think they had the concept down and when they were going to start designing it or shortly after they started they basically were told "20nm is garbage for GPUs and Apple is going to take most of the volume anyways".

Yeah, I have no idea how far along they were when they changed plans. Look at what NVidia promised with the original Maxwell roadmap back in March 2013: http://images.anandtech.com/doci/7900/OldRoadmap.jpg

Note the logarithmic scale on that- they promised double the DP FLOPS/W over Kepler. The only way you can get that kind of improvement is with a die shrink, and instead what we got was a reduction in DP performance. That roadmap also promises unified virtual memory, which is not present in the final Maxwell. The part on that roadmap bears no resemblance to the Maxwell that got released.

Then a year later, in March 2014, we saw the completely reshuffled roadmap. This had a 28nm Maxwell which matched up with what we finally got. So the change happened some time between 2013 and 2014 (assuming any honesty from NVidia.) If they intended to release the 20/14nm Maxwell with doubled DP performance in 2014, they surely must have been pretty far into the design before it got canned?

Also, don't forget about the 20nm Maxwell which turned up in the Tegra X1. This is a different design (compute capability 5.3) which has features like FP16 SIMD which weren't present in the 28nm part. I can only presume that this also came out of the cancelled "original Maxwell", but I can't claim to know how they're related.

[Stuckey]

My guess is the notebook version will be clocked at 900 Mhz. I am expecting a 1024 sp chip at 100-110 sq mm with a 128 bit GDDR5 bus (7 Gbps = 112 GB/s ) or 128 bit GDDR5X bus (5 Gbps = 160 GB/s). 900x2x1024 = 1.8 TFLOPs (matches PS4 performance). For desktop I think the clocks will be 1.1-1.2 Ghz.

I agree. I think the various versions, mobile, <70W compact desktop, and performance desktop will be exactly the same GPU but using different core/memory clocks. Essentially one small, cheap 14nm GPU to cover everything with a 1080p screen. With the highest performance version (highest clocked) being ~= to a 960. And like the 960, available in both 2GB and 4GB versions.

All AMD need Polaris 10 to be is the best VFM 1080p card for ~$200 or less. Because that's what the vast majority of 1080p gamers actually buy. As witnessed by the popularity of the 950 and 960. And I am 100% certain they will aim to do that with the smallest/cheapest die possible. Because 14nm is expensive. And even more so on bigger dies with inherently lower yields.

[integrated]

There's a lot of good info here on Maxwell that ought to be in a different thread.

I think AMD will configure Polaris 10 (P10) across a range of TDPs.

This part should be able to kick ass in the <75W DT and mobile categories, and get upclocked and stacked with 4/6GB for more serious desktop 1080 work at 75W+

[raghu78]

I agree. I think the various versions, mobile, <70W compact desktop, and performance desktop will be exactly the same GPU but using different core/memory clocks. Essentially one small, cheap 14nm GPU to cover everything with a 1080p screen. With the highest performance version (highest clocked) being ~= to a 960. And like the 960, available in both 2GB and 4GB versions.

All AMD need Polaris 10 to be is the best VFM 1080p card for ~$200 or less. Because that's what the vast majority of 1080p gamers actually buy. As witnessed by the popularity of the 950 and 960. And I am 100% certain they will aim to do that with the smallest/cheapest die possible. Because 14nm is expensive. And even more so on bigger dies with inherently lower yields.

My guess is AMD will have a 3 month lead in time to market with Polaris 10 and should go for gaining market share and price the 4GB versions at USD 200. the 2GB versions can come in at USD 170. Then once Nvidia responds AMD can cut another USD 20- USD 30 and keep the card competitive and still attractive. I think the yields should be good for such a small chip (100 -110 sq mm). AMD has a wafer commitment to meet with GF and thus should push for maximizing sales (thus wafer volume) and gaining market share at the low end. AMD's goal should be to get to a 35% GPU market share asap. Thats when their GPU business becomes a self sustaining business (atleast breakeven to minor profits). My guess is AMD will have Polaris 10 out by late Q2 2016 and Polaris 11 (flagship GPU with HBM2) out by late Q3 2016. In both cases I expect they will have a minimum 3 month time to market lead.

There's a lot of good info here on Maxwell that ought to be in a different thread.

I think AMD will configure Polaris 10 (P10) across a range of TDPs.

This part should be able to kick ass in the <75W DT and mobile categories, and get upclocked and stacked with 4/6GB for more serious desktop 1080 work at 75W+

I think AMD realizes by now how badly they have executed in the low power space in the last 2 years as Nvidia made a killing with GM107. AMD's market share was hurt the most because they had no answer to Nvidia's GM107 in thin and light notebooks and low power SFF gaming desktops. My expectation is GM107 must have taught AMD a lesson on how to design their GPU stack. First launch with a small low power GPU which can run off the PCI-E bus and does not require an external PIC-E 6 pin connector. Then release bigger GPUs as yields improve and wafer allocation also improves. With Raja Koduri claiming that Polaris 10 will deliver console quality gaming in a thin and light notebook I expect a PS4 class (1.8 TFLOPS) GPU at 25-30W. The desktop version will have higher clocks and come in at 50-60w. I think Polaris 10 is aimed at hitting OEM's back to school refresh cycle. I do not expect even overclocked desktop AIBs to require a 6 pin PCI-E power connector.

[tifosi]

AMD really run a tight ship with almost no leaks, but is there any news on when is something mid-to-high end going to hit the streets?

[GeorgiD]

AMD really run a tight ship with almost no leaks, but is there any news on when is something mid-to-high end going to hit the streets?

I guess you missed the leaks around CES 2016. AMD even showcased a running Polaris small card, and measured / showed its power consumption, quite considerably improved against a Maxwell 950 card.

This summer / before the back-to-school season.

[sirroman]

Those arent exactly "leaks", GeorgiD

[cduchesne]

We may not see price drops as inventory could self-clear with another upcoming coin mining ramp-up. Ethereum has gone from about $0.75 per coin to $2.50 just this month and took the 2nd highest market cap of all crypto coins http://coinmarketcap.com/. I believe this is one of the only coins that can / is still mined via GPUs profitably. Used 280x / 280 seem to be most popular by the mining community, however someone posted Fury results showing very efficient results (100w & > 280x performance).

The good news, is Polaris will most likely be coming out right as the the market for mining this coin becomes unprofitable. Unlike the last coin mining boom/bust that caused a bunch of inventory with noncompetitive cards.

Edit: just to make clear, the mining for ethereum should peak anywhere from 100k-1000k GPUs. This isn't such a large amount compared to the whole market, however it will certainly aid in removing the still-seen 280x inventory: http://www.newegg.com/Product/Produc...82E16814131533.

Just posting an update. Ethereum is now at $3.70/coin as of this morning. With the current network hashrate of 700GH, an R9 280x will net $100/month in coins. Clearly such a high payout won't be maintained. The current value supports 100,000 R9 280x if the hashrate quadruples, still netting $25/month in coins per GPU.

Anyhow, I know this is peanuts and possibly annoying, however it's possibly worth keeping a distant eye on if Ethereum continues to gain momentum.

[DCO]

Just posting an update. Ethereum is now at $3.70/coin as of this morning. With the current network hashrate of 700GH, an R9 280x will net $100/month in coins. Clearly such a high payout won't be maintained. The current value supports 100,000 R9 280x if the hashrate quadruples, still netting $25/month in coins per GPU.

Anyhow, I know this is peanuts and possibly annoying, however it's possibly worth keeping a distant eye on if Ethereum continues to gain momentum.

7970, perhaps the best gpu purchase someone could have done.