AMD Adds A Seemingly New Polaris ID To Their Linux Driver

[mrostu]

Some intriguing news here. But I think it's about a custom Polaris chip or a special version (for Chinese market e.g.) of current Polaris 20.

Would be nice to see Vega-based mid- and low-range chips on 12nm lithography. In this case I could get RX 680 8GB and don't wait for the Navi GPUs. But I believe it will not happen.

[ms178]

A Polaris 30 on 12 nm and a 7 nm Vega 20 for Gamers would make sense if Navi is more of a late 2019 launch and if the yields were acceptable on both processes. It would enable AMD to shift all of its Vega GPU production to TSMC (Vega 12 might turn out to be a die shrink of Vega 10 without the ML/HPC specific parts but that is pure speculation). As GloFo would still get to manufacture the 12 nm Polaris chips, they could shift more production lines from 14nm to 12nm.

What is in for the consumer? A 15% performance boost would be good enough to be well ahead of the GTX 1060 in many games. And there was a rumor out that the next mainstream chips from Nvidia will get postponed to 2019. If that turns out to be true AMD could take the lead in the volume market (hopefully as soon as Q4 2018). If it turns out to be false, then it should be a stop gap measure to buy some time until Navi arrives but it needs competitive pricing then. Concerning Vega 20 and Vega 12, the 7nm process should help to make it more competitive against the 1080. If pricing and availability were slightly better than Vega 10 right now, that could help them to be the better value proposition in the end. A lot of speculation on my part, I know. I can't wait to hear some hard facts of their plans! The near silence up until now makes me nervous that there is nothing new to expect until 2019 which would be a huge disappointment for me.

[ms178]

Maybe there will be RX580x2 or even quad/hexa crossfire?

Long term, I think AMD confirmed on working on multi chip modules for GPU's. But their GPU chief architect seemed to be not that enthusiastic about using this approach in Gaming products, at least in the short term. Long time ago, around 2011 or 2012, they promised to make better use of all the ressources inside a computer, and I am still waiting for that to happen. I can imagine high end Ryzen APU's + dGPU combinations working seemlessly together using unified memory (HBM2 / HBM3 / GDDR 6?) and a cache coherent interconnect (Gen-Z). I guess these new technologies are the ground work for a new push into heterogenious computing soon.

[Nille_kungen]

It could be anything like an laptop chip, embedded or even a new intel combined chip.
But i don't think it's anything i will ever notice or anything exiting since i would think if that was the case then it would be vega based instead and i can't see AMD releasing new interesting polaris cards anymore.

[riklaunim]

Long term, I think AMD confirmed on working on multi chip modules for GPU's.

R9 295X2 or Radeon Pro Duo weren't bad products and right now they likely need something that would benchmark above Vega 64 while having much much better availability and price/performance ratio. Dual-GPU card would be quite "normal" product for AMD - Crossfire or DX12/Vulkan multi-GPU. As Polaris doesn't look expensive or hard to make (and they have Glofo stuck on higher nodes) they could try with more than two GPUs. That could also be a early demo tool for chiplet designs - offering experimental interfaces to mimic behaviors of a chiplet GPU so that developers can play with it. Quad or Hexa RX580 would be so insane and nerdy...

[tuxd3v]

I think Amd is trying to clean the mess, that Rx4xx or 5xx are..
I mean creating a Bios Firmware that does not require pcie atomics...

All Nvidia cards don't require,
So this is to break Nvidia advantage in compatibility issues..

[ms178]

R9 295X2 or Radeon Pro Duo weren't bad products and right now they likely need something that would benchmark above Vega 64 while having much much better availability and price/performance ratio. Dual-GPU card would be quite "normal" product for AMD - Crossfire or DX12/Vulkan multi-GPU. As Polaris doesn't look expensive or hard to make (and they have Glofo stuck on higher nodes) they could try with more than two GPUs. That could also be a early demo tool for chiplet designs - offering experimental interfaces to mimic behaviors of a chiplet GPU so that developers can play with it. Quad or Hexa RX580 would be so insane and nerdy...

From an engineering perspective they were impressive, yes. They also pushed the limits in terms of GPU cooling. But I would be surprised if they did make financial sense for AMD. They were more of a HALO / niche product for enthusiasts to begin with, also to get some headlines and show what they are capable of doing to the world. From a practical point of view, even back then Crossfire had its issues and was dependent on the support of the game developer. This major achilles heel is still with us today and needs to be solved first to make it a viable solution in the future again for the desktop market.

If you want a dual-Vega card today, you can get the Radeon Pro V340. Its nowhere near cheap tough, but with today's limitations it makes more financial sense for AMD and more practical sense for the users to place this product for the professional market.

[bridgman]

I think Amd is trying to clean the mess, that Rx4xx or 5xx are..
I mean creating a Bios Firmware that does not require pcie atomics...

Just FYI that wasn't BIOS firmware, it was MEC microcode (what marketing calls ACE).

[riklaunim]

If you want a dual-Vega card today, you can get the Radeon Pro V340. Its nowhere near cheap tough, but with today's limitations it makes more financial sense for AMD and more practical sense for the users to place this product for the professional market.

dGPU Vega is dead. Barely available and expensive. Some shops have Vega 64 above RTX 2080 price. Polaris is old project on old node and likely already saturated it market, plus Globalfoundies is on a loosing side in the AMD agreements. They could spare some Polaris chips for experimental product to steal some press from Nvidia right now. AMD isn't expecting to compete with Nvidia on the marketshare anytime soon (like 2+ years), so at least they could just make some commotion - especially in the light of cold reception of new RTX carts. Navi will have to be brilliant and cheap to have market chances, but that isn't coming somewhere 2019 from what we can estimate now. 7nm was supposed to launch hard and quickly but it seems it's not that easy (and expensive). Big navi was mentioned somewhere 2020... And new arch after that.

AMD was first with Mantle, was first with x86_64, and would be cool if they would be first with some MCM GPU even if it would be overgrown crossfire combo. This will be bigger jump than DX12/Vulkan, but as we love tech and cool new stuff - AMD pls YOLO

[Marc Driftmeyer]

Anyone who thinks releasing a card with better performance than the 1080Ti is going to grab market share knows nothing about the market.