[Rumor, Tweaktown] AMD to launch next-gen Navi graphics cards at E3

[Glo.]

This really makes you wonder if yields really are that good? Why cripple the card with a smaller bus and vram? dumb market segmentation. Better to give it 12gb but lower clocks and or less CUs if one thinks it's too close to a 5700.

They can simply reuse Mobile chips on desktops.

 

[lifeblood]

This really makes you wonder if yields really are that good? Why cripple the card with a smaller bus and vram? dumb market segmentation. Better to give it 12gb but lower clocks and or less CUs if one thinks it's too close to a 5700.

I agree it would have made more sense to reduce the CU's & clocks rather than the bus. Leaving it at 8GB vram would have given AMD a nice market differentiator compared to the 1660 Super & Ti. Making it 12GB would also serve the purpose of making it stand out but would probably increase the cost (or reduce the profits) so much that it eliminated any potential benefit.

 

[VirtualLarry]

Just slap 12gb and make some progress in this category of gpu, ryzen 1600 had 6C/12T but games don't really use that many cores but amd still went for it so why not 12gb for $300. Could use more than 8gb in future at 1440p if someone is ok with locking to 30fps.

That would make a sweet card for mining one of the GrinCookoo coins, the one designed for larger VRAM cards, like the 1080ti's 11GB of VRAM. Coupled with the fact that it's GDDR6... let 'er rip! I think that there is a market for a 12GB RX 5600XT, albeit a niche one.

Edit: The coin that I'm talking about, is the SAME ONE that got Sapphire to release a 16GB VRAM RX 570 special-edition mining card.

Edit: Though, I don't know if the mining software that mines that coin, has been adapted to the RDNA architecture yet. I seem to only be able to mine using ClaymoreDual on my RX 5700.

 

[lobz]

I think it's fair since Nvidia has jacked prices up so far this gen.

Still don't get why the cards are complete garbage because of that.

 

[DeathReborn]

You can't have 8GB of VRAM with a 192bit bus. And 12GB is a total joke for this level of card. The card is not fast enough to run high enough settings or resolutions that would require 8+ GB.

I take it you haven't heard of Mixed Density VRAM before? Nvidia did it on the 550 Ti 1B 192bit: https://www.anandtech.com/show/4221/nvidias-gtx-550-ti-coming-up-short-at-150/2

The 5600XT looks quite decent vs the competition. Both vs the 1660ti and the 2060, which also have 6GB. 6GB is still quite bad at this price point.

The 5500XT is a turd, and I've already stated this in the 5500xt thread (or maybe it was on this thread? Not sure) . It doesnt have one redeeming quality vs the 1650S.

@SteveGrabowski compared the 960 to the preceding 760 with it's barely moved performance so it's only fair to do the same to the 5500XT and the 470/570/480/580. The 960 did perform close to the R9 285/280, but it was a poor product that still sold well. I do feel that the 5700 & 1660S make the 5600XT a tough sell, if the 2060/5700 dropped to $299 & the 1660 Ti to $249 it'd be an even tougher market for the 5600XT.

 

[Ranulf]

Why not?

Because 12 GB's of GDDR6 VRAM would cost 90$ on a 280$ GPU.

Its that bloody simple.

2 GB GDDR6 modules cost 50% more than 1 GB modules. And each 1 GB module costs 10$ with potential of going up by another 10-20% during 2020.

This is also the reason why RTX 2060 Supers costs 50$ more than RTX 2060, BOM is 20-25$ higher, than 2060, and this is the reason why RX 5500 XT 8 GB has 30$ MSRP higher than RX 5500 XT 4 GB with at least 20$ higher BOM.

Sure and the same argument can be used to show how silly the 5600XT's prices are compared to a 5700 or a 2060. The 1660ti's prices are silly already due to the 1660Super. Why would you buy a 5500XT 8GB at $199 (most seem to be going for $220-30 atm for Pulse and Gaming X type models) over a 1660Super at $230-250? The only reasons I can think of are you only want AMD cards or you're going to use that card for 3+ years and the 8GB of ram will help with that, probably.

 

[Stuka87]

I take it you haven't heard of Mixed Density VRAM before? Nvidia did it on the 550 Ti 1B 192bit: https://www.anandtech.com/show/4221/nvidias-gtx-550-ti-coming-up-short-at-150/2

GDDR6 standard has specs for memory chips with in-between capacities (such as 1.5Gb chips). But no memory manufacturer makes it. And if nobody makes it, nobody can use it.

The mixed capacity that the 550Ti had came with a major performance hit. It was NOT a good way to handle things.

 

[Glo.]

Sure and the same argument can be used to show how silly the 5600XT's prices are compared to a 5700 or a 2060. The 1660ti's prices are silly already due to the 1660Super. Why would you buy a 5500XT 8GB at $199 (most seem to be going for $220-30 atm for Pulse and Gaming X type models) over a 1660Super at $230-250? The only reasons I can think of are you only want AMD cards or you're going to use that card for 3+ years and the 8GB of ram will help with that, probably.

By silly you mean how much less AMD charges a premium compared to Nvidia, for the BOM costs?

 

[DeathReborn]

GDDR6 standard has specs for memory chips with in-between capacities (such as 1.5Gb chips). But no memory manufacturer makes it. And if nobody makes it, nobody can use it.

The mixed capacity that the 550Ti had came with a major performance hit. It was NOT a good way to handle things.

The performance was only badly impacted in CUDA, in general gaming it was faster than the 450 it replaced & the 5770, although it wa funnily enough overpriced... funny how nobody has learned from this oft repeated mistake. I was simply pointing out that 8GB on 192bit is entirely possible, 6 chips 4x 1GB (8Gbit) & 2x 2GB (16Gbit) both chips that are in production, unlike the 1.5GB etc chips.

 

[GodisanAtheist]

I take it you haven't heard of Mixed Density VRAM before? Nvidia did it on the 550 Ti 1B 192bit: https://www.anandtech.com/show/4221/nvidias-gtx-550-ti-coming-up-short-at-150/2

- OR AMD can always try BUS segmentation a la the GTX 970

 

[insertcarehere]

By silly you mean how much less AMD charges a premium compared to Nvidia, for the BOM costs?

Silly would be giving a damn about a card's BOM as a factor when deciding to purchase it.

 

[EXCellR8]

nvm

 

[Mopetar]

Silly would be giving a damn about a card's BOM as a factor when deciding to purchase it.

Indeed. Either it's worth what it costs or it isn't. Unless you could build it yourself for a lesser price you don't have much of a choice.

 

[DisEnchantment]

I think the bigger issue for both AMD and NVIDIA moving forward is not whether they can design high end cards, aka big dies, but whether the process has good enough yields for such large dies.
Additionally with Samsung 7nm yields not being economical outside of small Smartphone SoCs, then there is the capacity problem with NVIDIA also trying to move to TSMC for their high end at least.

Navi10 is only 251mm2 for a reason not because AMD cannot design a 2080Ti competitor but because in 2019 there not enough wafers to go around (at least for AMD) and the process is not mature enough to make 450+ mm2 dies with a high enough yield. Lisa confirmed of tight wafer availability, which in part is a fault of AMD for not having ordered enough in advance.
Since the chip has a conservative amount of CUs they needed to clock it high to reach the performance target but also losing some efficiency in the process.
I don't believe AMD could not have designed and launched a 2080Ti competitor last Computex if it were not for process/wafer constraints. <450 mm2 Navi10 could possibly reach 2080Ti performance under 300W if they could clock it modestly.

With AMD's console chip also fabbing in H2 2020, there is gonna be a real dearth of Wafers at TSMC. Add to the fact that the consoles are going to be around the 400mm2 die area.
But on the other hand, I have been seeing reports that the yields for bigger dies is improving a lot for 7+. It is presumably cheaper, higher yield and improved characteristics.

There is a reason nobody in HPC is talking about TSMC's 5nm except Smartphone guys like Apple. Ian mentioned some weeks ago yields even for tiny <100 mm dies are horrid for 5nm but at least it can alleviate the capacity problem for 7nm when Apple moves to 5nm.
Additionally this new interest in doing MCM and mGPU.
I think NVIDIA will fab the smaller GPUs at Samsung, probably AMD as well at some point in the future.
With RAM prices and Wafer costs ballooning in 2020 be prepared to pay top dollars for your next NVIDIA/AMD high end. There is a planned increase of WPM in H2 2020 at TSMC but remains to be seen if those are not going to be gobbled up by MS and Sony Orders.
For AMD they would rather use those limited wafers to produce EPYC rather than make some 2080Ti competitor.

 

[SpaceBeer]

One big problem for AMD is volume. High-end dGPU market is ~5% (volume, not earinings). And AMD might have 20% of those. So they are probably selling 0,5M units a year, which might not be enough to justify high production cost. At the same time, nVidia might be selling 2M units or more (gaming + HPC), and therefore can afford large dies production

 

[DisEnchantment]

One big problem for AMD is volume. High-end dGPU market is ~5% (volume, not earinings). And AMD might have 20% of those. So they are probably selling 0,5M units a year, which might not be enough to justify high production cost. At the same time, nVidia might be selling 2M units or more (gaming + HPC), and therefore can afford large dies production

For 12nm sure.
But for 7nm nvidia has no wafer allocation at TSMC since Samsung's 7nm is not feasible for their bigger dies. Even if they can afford they cannot book anything until possibly late H2 2020 and coincidentally MS and Sony are gobbling up WPMs in H2 to stockpile chips for holiday release of consoles.

For AMD they should not even try. Get 50-80 big GPU dies or 750+ chiplets for EPYC per wafer from fixed 20k+ WPMs they have?

AMD is selling everything they make from these 20k+ WPMs they have right now.

 

[Dribble]

I do feel that the 5700 & 1660S make the 5600XT a tough sell, if the 2060/5700 dropped to $299 & the 1660 Ti to $249 it'd be an even tougher market for the 5600XT.

Seems like Nvidia might agree: https://wccftech.com/nvidia-geforce-rtx-2060-price-cut-299-usd-tackle-amd-rx-5600-xt/

 

[lobz]

But that was so obvious. I can't help but think what many of you here are right and AMD are praying that the Radeons don't sell too well and they don't have to waste so many wafers on them. I must say... mission perfectly accomplished.

 

[Glo.]

Silly would be giving a damn about a card's BOM as a factor when deciding to purchase it.

Its not about buying a GPU based on BOM costs, but its completely silly to expect suddenly higher VRAM capacities on 200-250$ GPUs, considering the costs of GDDR6 memory chips.

Let me put this to your minds guys. 128 Bit GDDR6 8 GB memory frame buffer costs the same as 6 GDDR6 memory chips, on a 192 bit memory bus. 12 GB of VRAM would cost 90$. 10$ more than 256 Bit memory frame buffer.

Its completely financially pointless to offer higher memory capcities on lower cost GPUs.

And no, don't expect that companies will suddenly turn into charities. Silicon Wafer prices went up by 20% during 2H 2019.

 

[maddie]

I think the bigger issue for both AMD and NVIDIA moving forward is not whether they can design high end cards, aka big dies, but whether the process has good enough yields for such large dies.
Additionally with Samsung 7nm yields not being economical outside of small Smartphone SoCs, then there is the capacity problem with NVIDIA also trying to move to TSMC for their high end at least.

Navi10 is only 251mm2 for a reason not because AMD cannot design a 2080Ti competitor but because in 2019 there not enough wafers to go around (at least for AMD) and the process is not mature enough to make 450+ mm2 dies with a high enough yield. Lisa confirmed of tight wafer availability, which in part is a fault of AMD for not having ordered enough in advance.
Since the chip has a conservative amount of CUs they needed to clock it high to reach the performance target but also losing some efficiency in the process.
I don't believe AMD could not have designed and launched a 2080Ti competitor last Computex if it were not for process/wafer constraints. <450 mm2 Navi10 could possibly reach 2080Ti performance under 300W if they could clock it modestly.

With AMD's console chip also fabbing in H2 2020, there is gonna be a real dearth of Wafers at TSMC. Add to the fact that the consoles are going to be around the 400mm2 die area.
But on the other hand, I have been seeing reports that the yields for bigger dies is improving a lot for 7+. It is presumably cheaper, higher yield and improved characteristics.

There is a reason nobody in HPC is talking about TSMC's 5nm except Smartphone guys like Apple. Ian mentioned some weeks ago yields even for tiny <100 mm dies are horrid for 5nm but at least it can alleviate the capacity problem for 7nm when Apple moves to 5nm.
Additionally this new interest in doing MCM and mGPU.
I think NVIDIA will fab the smaller GPUs at Samsung, probably AMD as well at some point in the future.
With RAM prices and Wafer costs ballooning in 2020 be prepared to pay top dollars for your next NVIDIA/AMD high end. There is a planned increase of WPM in H2 2020 at TSMC but remains to be seen if those are not going to be gobbled up by MS and Sony Orders.
For AMD they would rather use those limited wafers to produce EPYC rather than make some 2080Ti competitor.

AFAIK, the yields on the 7nm process is better for large (>250 mm^2) HPC dies than mobile, as stated by TSMC. Wafer availability however has been constrained, but is expected to dramatically improve.

 

[DisEnchantment]

Indeed, For TSMC yields did improve for the lesser dense hpc processes, but not for Samsung.

 

[Stuka87]

I think the bigger issue for both AMD and NVIDIA moving forward is not whether they can design high end cards, aka big dies, but whether the process has good enough yields for such large dies.
Additionally with Samsung 7nm yields not being economical outside of small Smartphone SoCs, then there is the capacity problem with NVIDIA also trying to move to TSMC for their high end at least.

Navi10 is only 251mm2 for a reason not because AMD cannot design a 2080Ti competitor but because in 2019 there not enough wafers to go around (at least for AMD) and the process is not mature enough to make 450+ mm2 dies with a high enough yield. Lisa confirmed of tight wafer availability, which in part is a fault of AMD for not having ordered enough in advance.
Since the chip has a conservative amount of CUs they needed to clock it high to reach the performance target but also losing some efficiency in the process.
I don't believe AMD could not have designed and launched a 2080Ti competitor last Computex if it were not for process/wafer constraints. <450 mm2 Navi10 could possibly reach 2080Ti performance under 300W if they could clock it modestly.

With AMD's console chip also fabbing in H2 2020, there is gonna be a real dearth of Wafers at TSMC. Add to the fact that the consoles are going to be around the 400mm2 die area.
But on the other hand, I have been seeing reports that the yields for bigger dies is improving a lot for 7+. It is presumably cheaper, higher yield and improved characteristics.

There is a reason nobody in HPC is talking about TSMC's 5nm except Smartphone guys like Apple. Ian mentioned some weeks ago yields even for tiny <100 mm dies are horrid for 5nm but at least it can alleviate the capacity problem for 7nm when Apple moves to 5nm.
Additionally this new interest in doing MCM and mGPU.
I think NVIDIA will fab the smaller GPUs at Samsung, probably AMD as well at some point in the future.
With RAM prices and Wafer costs ballooning in 2020 be prepared to pay top dollars for your next NVIDIA/AMD high end. There is a planned increase of WPM in H2 2020 at TSMC but remains to be seen if those are not going to be gobbled up by MS and Sony Orders.
For AMD they would rather use those limited wafers to produce EPYC rather than make some 2080Ti competitor.

Navi 10 was never intended to be high end. Its the size it is because thats what AMD chose to fit that performance bracket. There is no way for them to know what the wafer availability will be years before it ships when they start the design process.

And I get the feeling AMD ordered every available 7nm wafer they could get their hands on. There is a shortage because of overall industry demand, not because AMD didn't "order enough".

 

[maddie]

Indeed, For TSMC yields did improve for the lesser dense hpc processes, but not for Samsung.

Navi 10 was never intended to be high end. Its the size it is because thats what AMD chose to fit that performance bracket. There is no way for them to know what the wafer availability will be years before it ships when they start the design process.

And I get the feeling AMD ordered every available 7nm wafer they could get their hands on. There is a shortage because of overall industry demand, not because AMD didn't "order enough".

Where does this leave Nvidia and competitive pricing in future GPUs. It's not like they can just wish 7nm wafers into existence. If AMD gets most of the increased 7nm fab capacity at TSMC this year, as rumored, then what happens to Nvidia.

 

[Glo.]

GTX 1660 Super, with any AMD CPU in Overwatch 1080p Epic settings avearges 135 FPS. 10% above this, means around 150 FPS for RX 5600 XT with the same CPU.

Which is exactly what you should get from this GPU.

GTX 1660 Ti with Intel CPU averages around 150 FPS in the same game, and settings. This means that RX 5600 XT should average around 160-165 FPS wit hthe same CPU, which is exactly what you should get from this GPU.

Its exactly between GTX 1660 Ti, and RTX 2060 in terms of performance. 1660 Ti averages 150 FPS, RTX 2060 averages 180 FPS in 1080p, epic settings.

160-165 is exactly between them.

 

[Stuka87]

Where does this leave Nvidia and competitive pricing in future GPUs. It's not like they can just wish 7nm wafers into existence. If AMD gets most of the increased 7nm fab capacity at TSMC this year, as rumored, then what happens to Nvidia.

That's a good question. It seems like Samsung isn't going to be an option for a while, unless they fix their 7nm HPC issues. For the time being nVidia is competitive with their 12nm stuff, but that won't always be the case.