RX 7900XT, 15,360 cores, MCM, Tapeout Q4

[Harlequin]

AMD are going chiplet so `600nm monster` wont happen.

 

[KompuKare]

AMD are going chiplet so `600nm monster` wont happen.

Okay, I did wonder about that.
Wonder what the size they'd go for then?
Yields and cost change massively for 300mm² or even a 400mm² chiplet, so much so that it's almost three 300mm² chiplets for the cost of one 600mm² one.
However, against that 2x300mm² won't really get them much farther than what a monolith could do.
Maybe the flagship will be 4 chiplets?

 

[Harlequin]

The biggest cost will be 5nm, a significant factor over 7nm+

 

[Grim5]

The biggest cost will be 5nm, a significant factor over 7nm+

Yes it's few thousand more expensive per wafer than 7nm. TSMC will be posting more record profits next year as between Apple, AMD, Intel and Nvidia they'll be smashed spitting out 5nm wafers at crazy profits

 

[Hotwired]

It doesn't seem like good business sense to release a 7900XT that's twice as powerful as a 6900XT. All they need to do is make a card thats around 50% faster and can play 4K games at over 80fps with full RT and it will sell. If the 7900XT is really double the performance of a 6900XT then I suspect whoever buys it will not need to upgrade for quite a while.

I bet that card should be cheap too?

Consumers haven't respected AMD cards for a long time and being acceptable just cements the second best image.

No reason not to take some Ryzen money and bash Nvidia over the head with it. Can't make things any worse.

 

[Harlequin]

Yes it's few thousand more expensive per wafer than 7nm. TSMC will be posting more record profits next year as between Apple, AMD, Intel and Nvidia they'll be smashed spitting out 5nm wafers at crazy profits

Ypu do know how much the fabs are costing to build? TSMC are carrying around $35 billion in debt and thats growing. They might have a quaterly profit of $4.9 billion but are using that to leverage more debt so new fabs can be built; the new USA one will be at least $17 billion and the next Taiwan will be $20 billion.

 

[Grim5]

Ypu do know how much the fabs are costing to build? TSMC are carrying around $35 billion in debt and thats growing. They might have a quaterly profit of $4.9 billion but are using that to leverage more debt so new fabs can be built; the new USA one will be at least $17 billion and the next Taiwan will be $20 billion.

Nothing wrong with debt, it cuts your tax obligations because interest costs are deducted

 

[humbug]

I bet that card should be cheap too?

Consumers haven't respected AMD cards for a long time and being acceptable just cements the second best image.

No reason not to take some Ryzen money and bash Nvidia over the head with it. Can't make things any worse.

Exactly, also being first with an MCM design they can knock out a massive unattainable 15K + core halo GPU as a giant middle finger to Nvidia, just as they did and still are doing to Intel.

If you have the chart topping GPU its worth more in mindshare than pretty much anything, Nvidia understand this, its why they keep pushing the size of GPU's to new heights. With MCM AMD have an opportunity to crush Nvidia. Effectively 3 6900XT's.
Pair that later with 3X chiplet 24 core 48 thread mainstream Zen 4 CPU and AMD look like the top dog out of the 3.
That's what brings people to your brand.

Hey guys listen, we do all the consoles, including the new Steam Deck, do you have a Samsung phone? our chips in that, drive a Tesla? yup the computer in that is ours too, now look at our GPU's and CPU's, they are the fastest on earth.

 

[Phixsator]

Lovelace shipping Q2 2022, 60 to 100% faster than RTX3090.

RDNA3 100% faster than 6900XT.

Internally AMD expects to beat Lovelace.

Internally, Nvidia has a backup plan in case they lose the performance crown - and that is to drop their prices and ship huge numbers of cards to hold overwhelming market share.

AMD intends to own the position of top dog performance, with execs floating the idea of pricing RDNA3 cards higher up than RDNA2 - an example is give of $2000 RRP for the 7900xt

I really hope that isn't the case but I wouldn't be surprised if AMD started to see dollar signs everywhere if they did manage to hold the performance crown. I can't blame them yet it frustrates me, unless its only the top card.. But lets be real for a second the price will follow down the stack sadly

 

[humbug]

Navi 31, RDNA3. RX 7900XT.
7,680 ALU's per GDC (Graphics Core Die) X2 for a total of 15,360 ALU's, the 512MB MCD in the middle is like a giant L3 Cache bridge connecting the two dies together. Think Zen 1 only in 3 separate pieces of silicon.
A 6900XT has 5120 ALU's.

https://twitter.com/Olrak29_/status/1419702500945371150

 

[KompuKare]

Navi 31, RDNA3. RX 7900XT.
7,680 ALU's per GDC (Graphics Core Die) X2 for a total of 15,360 ALU's, the 512MB MCD in the middle is like a giant L3 Cache bridge connecting the two dies together. Think Zen 1 only in 3 separate pieces of silicon.
A 6900XT has 5120 ALU's.

https://twitter.com/Olrak29_/status/1419702500945371150

I assume this is based on various speculation from all those authors.
Good news, though: 512MB cache should still be too little for Ethereum's DAG file which needs 4GB or more!

 

[timorous]

So a 15k shader 3 chip monster card.

As this is going to be the 7xxx series AMD might as well call it the 7990X2 3Ghz edition for nostalgia.

 

[Harlequin]

I assume this is based on various speculation from all those authors.
Good news, though: 512MB cache should still be too little for Ethereum's DAG file which needs 4GB or more!

erm DAG file goes into VRAM?

 

[KompuKare]

erm DAG file goes into VRAM?

Yes, well but note that Navi 21 performs poorly at mining as the main VRAM is slow and the cache trick doesn't work for Ethereum mining.

 

[Harlequin]

Yes, well but note that Navi 21 performs poorly at mining as the main VRAM is slow and the cache trick doesn't work for Ethereum mining.

you dont mine do you? 3090>3080>Radeon 7 > entire navi 21 range > 3070<>3060ti `poor for mining` is sujective.

 

[KompuKare]

you dont mine do you? 3090>3080>Radeon 7 > entire navi 21 range > 3070<>3060ti `poor for mining` is sujective.

No, I just have IGP and a old R5 260X ATM, so this is just looking at WhatToMine.com mainly occasionally looking when it might all come crashing down again.
No sign of that, yet.
Don't know how accurate their figures are:

So, yes quite subjective.
Had assumed that GA102 would be far better, but while they hash well they also use a lot of power.
A bit suspicious of WhatToMine's figures as they have the 3060Ti and the 3070 exactly the same.

 

[Harlequin]

which is as i said above - the 3060ti and 3070 are literally equal for mining, but the 3060ti is cheaper however the 3070 is more available. ETH is about memory throughput, RVN and ERGO are about core power (and thus your chart would be different).

 

[RBImGuy]

All the die hard core nvidia fans going to change their tune and go amd.
Its that good.
A whole new way to change the game of graphics cards
rdna2 currently is faster than nvidias where it actually counts.
Bloodbath is coming.

 

[Calin Banc]

There is of coure another way to approach pricing:

1. How much would it cost to make
2. And, if wafers are still in short supply (which they will be...) how much profit could they make producing CPUs instead (this is mainly revelant for AMD)?

Let's say AMD go all out and produce a 600mm² monster for the flagship.
We can plug that into a wafer yield calculator (taking the 0.09 defect per cm which is something TSMC mentioned for 7nm - although that was a few years ago now):
https://caly-technologies.com/die-yield-calculator/
Now 600mm² is 24.5mm x 24.5mm, which gets us this:

For the die (ignoring testing, packaging, VRAM, board etc.), the only other consideration is how much wafers cost.
TSMC 7nm was meant to be about $10,000 per wafer but prices are rumourd to actually have gone up a bit.
6nm might be very similar rather than costing more as EVU might be super expensive but throughput and amount of masks is reduced.
5nm will probably be more, maybe $15,000 per wafer.
Anyway, $10k get us to $190 per die, $15k to around $280.
(This ignores that with redudancy not all dies with defects will be wasted, although against that not every good die is able to hit clocks/power targets.)
Those numbers aren't that bad actually: in theory that means a dual chiplet card might be doable for around $1,200 or so even with the $15k figure.

However, AMD at least have more profitable things they can do with wafers and that would be Zen3 or Zen4 CCDs or even APUs.
The current Zen CCD is 83.7mm² which means the wafer costs between $15 and $23 ($10k / $15k). Zen 4 might be a bit bigger, and of course the IO dies come on top of that.
In other words, CPUs have far higher margins.
We see that with Zen3 vs Navi21. A 5600X @ $300 versus a RX 6800 @ $579 using the $10k figure gives margins of 95% and 72% (with $15k per wafer it is 92% and 58%).

Hard to see how things will play out.
AMD seem to madly focused on margins ATM even at the cost of marketshare and mindshare. The design and mask costs are fixed, so once wafer are available it makes no sense to price too high, but what do we know?

Well, can always make GPUs with a chiplet of 83.7mm² as well?
Sure, you have the vRAM and other "stuff", but if doable for smaller sizes, then normally AMD should be able to play around with them in low / mddle end as well. Theoretically, perhaps, at lower prices than current gen.

 

[Grim5]

All the die hard core nvidia fans going to change their tune and go amd.
Its that good.
A whole new way to change the game of graphics cards
rdna2 currently is faster than nvidias where it actually counts.
Bloodbath is coming.

None of my screens support Freesync so what amd sells for a graphics card does not work for me unless they support gsync.

If AMD ever supports VRR then it can work with 1 of my screens but not the others, though AMD has been promising VRR over HDMI for years and stil nothing