Question Speculation: RDNA2 + CDNA Architectures thread

[fleshconsumed]

I know everyone is excited about big NAVI, but does anyone know if there is going to be RDNA2 replacement for 5700XT this fall?

 

[dr1337]

Nope, we know both consoles clocks/CUs, no RDNA1 gpu could reach those numbers... even if it were ported to 5nm, it doesn't add up.

I mean the slide clearly mentions increasing clockspeed and mentions reduced logic complexity and switching power as well. Unless you're still talking about the fake 2.7ghz number, RDNA2 is 100% going to see better clock speeds.

 

[beginner99]

I know everyone is excited about big NAVI, but does anyone know if there is going to be RDNA2 replacement for 5700XT this fall?

Some people in AMD know but what we have is rumors at best. I suspect rdna1 will simply move down the stack, others have said it will be replaced completely due to the given code name of chips we have for rdna2 meanign there would be no logical place in the price structure for navi10.

 

[Glo.]

I mean the slide clearly mentions increasing clockspeed and mentions reduced logic complexity and switching power as well. Unless you're still talking about the fake 2.7ghz number, RDNA2 is 100% going to see better clock speeds.View attachment 23170

15% IPC increase, 25% clock speed increase, 10% lower power draw, including all of those improvements.

So 40 CU GPU: 2.3 GHz clock speed(25% over 1.85 GHz), 200W TDP, 15% IPC increase over RDNA1.

Well, its pretty darn good on paper I'd say...

Or... 36 CU GPU, 2.175 GHz clock speed(+25% over 1.75 GHz), 145W TBP, +15% IPC increase over RX 5600 XT.

 

[Glo.]

Some people in AMD know but what we have is rumors at best. I suspect rdna1 will simply move down the stack, others have said it will be replaced completely due to the given code name of chips we have for rdna2 meanign there would be no logical place in the price structure for navi10.

AMD has the incredible flexibility to react to what Nvidia offers. And they can use Navi 10 as even RX 6500 and 6500 XT, and put the Navi 23 dies in RX 6600 and 6700 Series SKUs.

 

[maddie]

Why would they do that

AMD has the incredible flexibility to react to what Nvidia offers. And they can use Navi 10 as even RX 6500 and 6500 XT, and put the Navi 23 dies in RX 6600 and 6700 Series SKUs.

Why would they do that? Die size still matters.

 

[Glo.]

Why would they do that

Why would they do that? Die size still matters.

Because they don't have smaller GPU in RDNA2 lineup?

N23 appears to be the smallest one, with 240 mm2.

 

[maddie]

Because they don't have smaller GPU in RDNA2 lineup?

N23 appears to be the smallest one, with 240 mm2.

Appears is correct. Selling a 250 die at a much lower cost than a 240 die??? Uncharacteristic.

 

[Glo.]

Appears is correct. Selling a 250 die at a much lower cost than a 240 die??? Uncharacteristic.

Even if that 240 mm2 die has higher performance than that 250 mm2 die?

Is it really that uncharacteristic? Or illogical?

 

[maddie]

Even if that 240 mm2 die has higher performance than that 250 mm2 die?

Is it really that uncharacteristic? Or illogical?

A higher cost product in a lower price higher volume segment? If you say so.

 

[Glo.]

A higher cost product in a lower price higher volume segment? If you say so.

Polaris series?

232 mm2 dies at the cost of 130$.

 

[beginner99]

Polaris series?

232 mm2 dies at the cost of 130$.

Polaris isn't on the same process. If the 240mm die is cheaper to make which we can assume, then it's logical to use a more cutdown version of that die instead of a older uarch bigger die. And since they are on the same 7nm process unless I missed something, then I have to agree with maddie that econimically it makes no sense to keep selling navi10.

 

[Glo.]

Polaris isn't on the same process. If the 240mm die is cheaper to make which we can assume, then it's logical to use a more cutdown version of that die instead of a older uarch bigger die. And since they are on the same 7nm process unless I missed something, then I have to agree with maddie that econimically it makes no sense to keep selling navi10.

I used P10 example to show that even if the die is large, it still can be used for entry level products, with entry level prices.

 

[Stuka87]

Navi 10 is all payed for as far as R&D goes. I don't see there being any reason for selling it for cheaper than it is currently being sold for as a lower end 6500 XT or something.

 

[Glo.]

Navi 10 is all payed for as far as R&D goes. I don't see there being any reason for selling it for cheaper than it is currently being sold for as a lower end 6500 XT or something.

If That GPU can compete with Nvidia's RTX 3050 Ti, and with RTX 3060(to some degree), why AMD would lower margin on higher performing Navi 23 GPU, and put it to compete with those GPUs?

They would undercut themselves, in terms of business decision.

Lets say that Navi 23 has RTX 2080 Super performance. Do you release it as RX 6600 XT, with 299$ price tag, knowing that RTX 3070 with 399$ price tag will have the same performance level, but worse efficiency?

If AMD has the competitive advantage from business perspective, because their technology is so good, why hold back?

If Navi 10 is enough to compete with low-end products from Nvidia, why would you release GPU that can be used for HIGHER MARGIN products to compete with them? There is absolutely no reason for AMD to do so, from business point of view.

 

[maddie]

I used P10 example to show that even if the die is large, it still can be used for entry level products, with entry level prices.

Poor comparison. You're not accounting for the identical process lines. Is 7nm now a buyers market? If there is any strong demand for 7nm wafers then what you propose is not really a good idea.

 

[Glo.]

Poor comparison. You're not accounting for the identical process lines. Is 7nm now a buyers market? If there is any strong demand for 7nm wafers then what you propose is not really a good idea.

Its all about the margin. You still will be able to get margin on the N10 GPU as a X500 SKU, especially if it will come with 6 GB GDDR6, instead of all 8 memory chips.

And As I have said, in the upper post, Its all about Margin. If Navi 23 can be sold for higher margin than N10, why not go for it, from AMD's perspective in competitive landscape?

 

[beginner99]

why not go for it, from AMD's perspective in competitive landscape?

If the Navi23 die is cheaper to make than Navi10 die which given the die sizes of 240 vs 250mm probably means yes this is the case, then selling a more expensive die at a lower price makes 0 economical sense. Better to cut-down the Navi23 die to navi10 performance level and sell that cheaper. Your margin will still be higher because the die is cheaper. As added bonus the cheaper SKU also gets all the newest video decode/encode features.

Now I can admit that if the die sizes are that close and due to the fact that OEMs could simply update the model name and leave the manufacturing the same in a total cost for the card, yes selling a 5700(xt) as a lower 6000-series model could make sense.

 

[DisEnchantment]

At this point, it is a matter of whether N23 has taped out and if the silicon is ready. If not AMD would rebrand Navi10 to cover the lower performance tier.
Otherwise no point in fabbing a 251mm2 die which can be outperformed by a 240mm2 die on two largely similar processes with simliar costs and using almost 90% common equipment.
P10 existed for a while due to being cheaper to manufcature and WSA to some extent.

Since AMD already ported RDNA2 on N7(P) for the consoles it wouldn't be much of a challenge to repeat for N23. Much of the work is done.

• Designing around the same libraries and tool kits
• Porting design to N7(P) PDK
• Synthesis, timing analysis and simulations largely the same.
• For similar Navi2x design, it would be a matter of configuration of the blocks, redo the floor plan and place and route and other small backend steps

Still there are costs and effort involved but nowhere near porting to a new PDK and new libraries and the functional blocks have already been designed, so it is about ROI whether it makes business sense. If there is money to be gained they will do it.

I prepared a sample workflow using older design flow, but still generically OK.

 

[Glo.]

If the Navi23 die is cheaper to make than Navi10 die which given the die sizes of 240 vs 250mm probably means yes this is the case, then selling a more expensive die at a lower price makes 0 economical sense. Better to cut-down the Navi23 die to navi10 performance level and sell that cheaper. Your margin will still be higher because the die is cheaper. As added bonus the cheaper SKU also gets all the newest video decode/encode features.

Now I can admit that if the die sizes are that close and due to the fact that OEMs could simply update the model name and leave the manufacturing the same in a total cost for the card, yes selling a 5700(xt) as a lower 6000-series model could make sense.

It makes zero sense to cut down cheaper die(N23) to Navi 10 performance levels, because that UNDERCUTS the price margin.

If the manufacturing cost of N10 die is 100$ for RX 5700 XT performance levels and manufacturing cost for N23 die is 98$ for RTX 2080 Super levels, by cutting down N23 to RTX 5700 XT levels you effectively LOSE MORE MONEY, that you could have earned if you would not do this.

Price margin on N23 will be higher because of its performance, than it will be for N10 when Nvidia will release their entry level products.

Remember, the die sizes are marignally different. So the manufacturing costs will be marginally different as well. Yes N23 MAY BE cheaper to make. But will it be meaningful difference?

When it goes for calculating price margin - yes it will be meaningful, because that way selling cheaper die for more will earn you more money than if you would sell more of the cheaper die, at lower price margins!

 

[maddie]

It makes zero sense to cut down cheaper die(N23) to Navi 10 performance levels, because that UNDERCUTS the price margin.

If the manufacturing cost of N10 die is 100$ for RX 5700 XT performance levels and manufacturing cost for N23 die is 98$ for RTX 2080 Super levels, by cutting down N23 to RTX 5700 XT levels you effectively LOSE MORE MONEY, that you could have earned if you would not do this.

Price margin on N23 will be higher because of its performance, than it will be for N10 when Nvidia will release their entry level products.

Remember, the die sizes are marignally different. So the manufacturing costs will be marginally different as well. Yes N23 MAY BE cheaper to make. But will it be meaningful difference?

When it goes for calculating price margin - yes it will be meaningful, because that way selling cheaper die for more will earn you more money than if you would sell more of the cheaper die, at lower price margins!

Nice picking prices to suit your argument. One rule of what ifs, GIGO.

 

[Glo.]

Nice picking prices to suit your argument. One rule of what ifs, GIGO.

What did you attatched yourself to the specific prices? Its just analogy, and example.

Look beyond the numbers, to see the bigger picture.

 

[Ajay]

At this point, it is a matter of whether N23 has taped out and if the silicon is ready. If not AMD would rebrand Navi10 to cover the lower performance tier.
Otherwise no point in fabbing a 251mm2 die which can be outperformed by a 240mm2 die on two largely similar processes with simliar costs and using almost 90% common equipment.
P10 existed for a while due to being cheaper to manufcature and WSA to some extent.

Since AMD already ported RDNA2 on N7(P) for the consoles it wouldn't be much of a challenge to repeat for N23. Much of the work is done.

• Designing around the same libraries and tool kits
• Porting design to N7(P) PDK
• Synthesis, timing analysis and simulations largely the same.
• For similar Navi2x design, it would be a matter of configuration of the blocks, redo the floor plan and place and route and other small backend steps

Still there are costs and effort involved but nowhere near porting to a new PDK and new libraries and the functional blocks have already been designed, so it is about ROI whether it makes business sense. If there is money to be gained they will do it.

I prepared a sample workflow using older design flow, but still generically OK.


View attachment 23192

Where'd you get that chart? I've been looking for something like that for a while.

 

[maddie]

What did you attatched yourself to the specific prices? Its just analogy, and example.

Look beyond the numbers, to see the bigger picture.

One can always choose an imaginary number to achieve a desired outcome. You're constructing a strawman like scenario to support your view, that's all. The bigger picture deals with facts not assumptions.

 

[Glo.]

One can always choose an imaginary number to achieve a desired outcome. You're constructing a strawman like scenario to support your view, that's all. The bigger picture deals with facts not assumptions.

And on what did you based your assumption that smaller die size is cheaper to make, then, in the first place, hmmm?