The latest LLVM patch confirms the increased register size for N31/N32. And it also says that the allocation granularity went up by 50% (24/12 vGPRs in Wave32/64 mode instead of 16/8 vGPRs or 8/4 vGPRs with RDNA1). This implies also 50% more register banks (instead of larger banks) and therefore more register bandwidth to satisfy the VOPD instructions (and potential co-issue of two vector instructions from two wavefronts).
- Status
No they haven't and I have no clue why some sites think they have. The slide very clearly states the prices are Neweggs cheapest of each on September 15th, nothing more and nothing less.
Now the question is how much stock of some of these parts are there. At $700 for the 6900xt there isn't much room between that and the 4080 runt edition to take advantage of nvidia's pricing. But maybe stock is so low they think it will all be cleared out by the RDNA 3 launch?
If they can get 4080 runt performance in raytracing and rasterization with 16gigs of ram at 3080 launch prices I would be happy to go back
Whatever the stock situation is, it is bound to be less severe in absolute numbers for AMD just due to volume. The problem I can see though is a conflict of interest for many of their board partners who also make Nvidia boards. How happy are you going to be to sell a new great value chip, when its success takes away sales from the mountain of old inventory you are sitting on?
Flappy Pannus
Veteran
Ah damn missed that.
Flappy Pannus
Veteran
I guess we'll see, but AMD's drops have been pretty consistent over the past few weeks. No indication stock is an issue so far, once we see prices trend upwards a tad that would be an indication but it's been a steady slope and virtually all models week to week are in stock.
That may be a bit optimistic. Undercutting it by $200 on a new process with more ram? They could not have the concern of Nvidia with the same level of inventory sure and that would affect that - it's certainly what I hope, but I think Nvidia by now has an idea of what AMD is about to deliver and how much it costs. Fingers crossed but I'm skeptical.
I dunno, so close to a new graphics cards releases and the glut of product on 2nd hand markets might not really allow that to be an indication of anything.
Isn't RDNA 3 a hybrid 7/5nm product with some of the chiplets being on 7nm and others on 5nm ? Also with infinity cache AMD should be able to get away with slower memory.
I would think that ray tracing would be the biggest issue here. I remember the 3080 mopped the floor with the 6900xt.
I don't think Nvidia really cares about AMD.
I guess we wont know the target AMD has to hit really until the 4090 hits
Yeah, I'm dubious about the chiplet design... but I guess we will find out soon enough. Honestly, I don't think they even really need to compete with the 4000 series RT perf. As long as they can offer something better than the 3000 series at a reasonable price, they will have a winner. Hopefully, it is much improved from RDNA2.
Nvidia certainly knows far more about AMD's upcoming products than we do, so I highly doubt that. They most likely don't have much of a reason to worry about it anyway.
Subtlesnake
Regular
If they just doubled 6900 XT performance across the board then the ray tracing performance would be around the 3090 Ti level in the worst case. But AMD have said they have improved their raytracing capabilities specifically, so we can hope for better than that.
One could only hope. The more competition the better
IMO neither more banks nor larger banks are implied in the patch.
If Wave64 were to get single-cycle VALU execution all across the board as some earlier patches hinted, having 50% more banks does rhyme with the theme, in order to be able to supply max 6 VGPR operands per clock for Wave64 FMA.
(for VOPD and co-issuing from 2 waves, one can still argue they are designed to maximise usage of existing resources; Wave64 single-cycle EX not so much, unless FMA/IMA/3-input opcodes are excluded deliberately)
Last edited:
Well, with a +50% increase of total register space, something should go up by 50%. How many choices do we have?
I think it's that they can argue the 4080 12gb can outperform whatever ends up in that range when combined with DLSS 3.0, even if it doesn't end up competitive in most scenarios outside of that.
Nvidia marketing really like to point out the 2 - 4x edge cases that 99% of games will never hit to justify the pricing. Just like they compare DLSS 2.0 vs DLSS 3.0 to obfuscate the numbers.
It might end up being a valid comparison if DLSS 3.0 is widely adopted and provides good quality results to be fair, just like DLSS 2.0 was.
What I meant is — those are the two choices, but the patch did not suggest which.
If anything, an earlier VOPD patch on the bundling algorithm suggested it might be sticking to 4 VGPR banks still, though it could still be updated at any time.
gamervivek
Regular
If the Angstronomics numbers are correct, then Navi31 should at least double the 6900XT performance. 2.4x shaders, the clockspeed would be about a GHz higher. The ROPs should be only 50% more, however the clockspeed jump should also increase the throughput. But then I thought the same for AD102, and the numbers haven't really lined up that way. 
If AMD end up on top in raster numbers, I just don't see them pricing it lower than the $1200 4080, and definitely closer to the 4090 if not the same. So the crazy prices are here to stay.
If AMD end up on top in raster numbers, I just don't see them pricing it lower than the $1200 4080, and definitely closer to the 4090 if not the same. So the crazy prices are here to stay.
If the Angstronomics numbers are correct, then Navi31 should at least double the 6900XT performance. 2.4x shaders, the clockspeed would be about a GHz higher. The ROPs should be only 50% more, however the clockspeed jump should also increase the throughput. But then I thought the same for AD102, and the numbers haven't really lined up that way.
If AMD end up on top in raster numbers, I just don't see them pricing it lower than the $1200 4080, and definitely closer to the 4090 if not the same. So the crazy prices are here to stay.
Then what would be the reason anyone would go with an AMD gpu? Same price, same raster performance but weaker RT/ML performance. Besides the brand loyality then. If i'd be AMD i would be trying to catch some marketshare back, now is their chance to do so. Atleast if competing with RTX4000. To me it seems like RTX4000 isnt really ready for the mass market yet and RTX3000 is left in stocks.
If high prices like that are here to stay remains to be seen. Might very well be possible, but we cant predict how the market will react to that. I personally dont think that companies can forever up the prices of gpus, phones and other such devices like that. Some are dropping close to two grand on a mobile phone these days. Its just crazy.
- Status
