AMD Radeon RX 6600 XT Launching For 1080p RDNA2 Gaming At ~$379 USD

[drakonas777]

Oh, I see birdie has already memed around 6600XT price. It's funny how mr. value guy was totally silent during 3080Ti launch: +70% MSRP for +15% performance was a good value for him I guess, since it's a "quality product"...

MSRP of 6600XT is not that bad considering current competition and semiconductor economy context. Problem is scalpers.

[drakonas777]

Forking GPU architectures for a different node is unrealistic/borderline unfeasible way to counter market changes. It's too expensive and too much of "inertia" to dynamically follow the demand. The most realistic thing AMD could have done is to extend lifetime of Polaris cards without any product changes. And even then it's not like you can shuffle around 12N resources like you wish - reservations of GloFo wafers exist most likely and all this assuming some general material shortages do not affect older nodes (which they did at least at some point AFAIK).

Another way could be a small and cheap 7N GPU with 4G of VRAM to make it unattractive for miners, but powerful enough for 1080p ~medium for those, who are building new systems (mid range especially) and can't get a GPU at all, at the sane price that is. It would be a temporary solution to wait until market comes down. 5300XT was something like this, but it got nerfed due 7N shortage in general I assume. NAVI24 could be something like this at the and of the year.

I do't think there is an easy solution to this problem. Either way you would have to pay some premium for some inferior product compared to "normal" conditions.

[drakonas777]

By "forking GPU architectures for a different node" I meant backporting to non-compatible ones, basically like 4/5->6/7->8/10->12/14. This involves considerable amount of R&D and general production pipeline time. Mere shrink to compatible one sounds reasonable to me. I would certainly take something like RX690 on 12N+ or whatever assuming sane pricing. On the other hand, it wouldn't change the performance that much, so even standard RX480/580 would work for me just fine, especially if standard one would save months of waiting for shrank one.

[bridgman]

Realistically anything we did with Polaris at this point would be aimed at miners, not gamers. If the goal was increasing gaming GPU supply (leaving gamers still competing with miners for the supply) then starting with newer products would make more sense, and moving to newer fab processes than 7nm would probably work better.

Once you start changing more than a few transistors on a chip you add two major chunks of work - redoing all of the simulation/emulation testing in the affected areas (which is significant) and redoing all of the physical layout for the changed blocks and any larger blocks which contained them. In practice that means redoing pretty much the entire chip for the kind of changes discussed here (FP16, AV1).

One other thing I should mention early is that all this discussion assumes there is capacity at the AIB partner level, in terms of both their own production capacity and availability of all the sub-components that go into making a dGPU board. Many of those parts are also seeing shortages and extremely high prices at the moment.

Probably the easiest way to think about it is levels of effort:

* Low effort:

Restart/increase RX 590 production (perhaps with more mining-friendly clocks) with current physical design, which I believe is the 14nm layout fabbed on 12LP. I'm not sure how much the 14nm design was updated to take advantage of 12nm - the leading rumor is that it's basically 14nm masks on 12LP fab to get better speed/power.

I say restart/increase because I don't know if we are still placing orders for full size Polaris dies or only for the smaller 550/560 dies.

* Small effort:

Move RX 590 to 12LP+ with no change other than clocks/voltages, but without a shrink (or maybe with a linear shrink if feasible, which I doubt).

* Medium effort:

Redo RX 590 physical layout for 12LP+ taking full advantage of reduced feature size

Redo newer chips (RDNA2 or 5500XT) for newer fab process

* Large effort:

Any logic design change (FP16, AV1 were examples) to RX 590 plus move to 12LP+

These are all approximate but I believe they are generally correct.

Anyways, key point is that making a Polaris-based product at this point would have to be aimed at miners because the current design works well in that role. My impression is that it is pretty well balanced between compute and memory throughput in its current form.

Bringing Polaris up to date for gaming ends up requiring as much or more engineering work than starting with a newer product generation.