AMD Radeon RX 6700 XT: All You Need to Know

[windwhirl]

40CU/2560SP up to 2424Mhz on 251 mm2 die for $479??? That's Polaris sized GPU for God sake!
It'll be 12,4 TFlops at best... so 25% faster than 5700XT for 20% higher MSRP and 50% higher AIB prices. 5% added value in best case scenario and 45% price to performance regression in reality. AMD go eat Sh/t!

TSMC stopped giving volume discounts, for starters. And newer nodes are more expensive than older ones.

Where did you get Navi22 being 251mm2 from? That is the size of the 5700XT (Navi 12). Navi 22 is apparently 334mm2 (Navi 21 aka 6800/6900 is 520mm2). It is also on a much more expensive process than Polaris (7nm versus 12nm) so the die cost is likely double that of Polaris.

On that, some analysts believe the wafer cost for 7nm is more than double the cost for 16/12 nm.

Alleged Prices of TSMC Silicon Wafers Appear

TSMC, one of the biggest silicon manufacturers in the world, usually doesn't disclose company pricing of the silicon it manufactures and only shares that with its customers. It appears that RetiredEngineer (@chiakokhua on Twitter) got a hold of the pricing of TSMCs wafers on every manufacturing...

www.techpowerup.com

 

[Relixo]

Title: "AMD Radeon RX 6700 XT: All You Need to Know"
In the reality this is what we all have to know: Miners will buy ALL the cards!!!
So, go to hell miners aaaaaand

 

[RedelZaVedno]

Where did you get Navi22 being 251mm2 from? That is the size of the 5700XT (Navi 12). Navi 22 is apparently 334mm2 (Navi 21 aka 6800/6900 is 520mm2). It is also on a much more expensive process than Polaris (7nm versus 12nm) so the die cost is likely double that of Polaris.

6900XT has 520 mm² die... Half it and you get to 260 minus smaller infinity cache and you get to RNDA1's die size. Other numbers are from AMD given to Steve at Gamers Nexus before presentation. It's overpriced every way you spin it. TSMC's 7nm process is very mature by now and we're talking about really small die size here, so number of non-functional dies per water must be really low, meaning it's cheap to produce. 6700XT could cost 350 bucks and AMD would still make higher profit margin on it as it did on Vegas and even Polaris. It's clearly a rip off not worth paying like everything else on today's dGPU market. I hate people who defend these 2 greedy corporations and don't even own their stocks. Profit margins of both have been skyrocketing this decade and they still want to milk us more. We're voting with our wallets here. There will be no back to normal once mining craze ends IF gamers buy GPUs at these MSRPs or higher. I've been buying GPUs since early 90ies and dGPU market has never been worse than it's today.

 

[sepheronx]

Title: "AMD Radeon RX 6700 XT: All You Need to Know"
In the reality this is what we all have to know: Miners will buy ALL the cards!!!
So, go to hell miners aaaaaand
View attachment 190789

The RX 6800 aren't really a good price/performance for miners. A 6800XT nets you what a 3070 does and the 3070 seems to be roughly about $300+ more expensive in Canada.

 

[Tom Yum]

6900XT has 520 mm² die... Half it and you get to 260 minus smaller infinity cache and you get to RNDA1's die size. Other numbers are from AMD given to Steve at Hardware Nexus before presentation. It's overpriced every way you spin it. TSMC's 7nm process is very mature by now and we're talking about really small die size here, so number of non-functional dies per water must be really low, meaning it's cheap to produce. 6700XT could cost 350 bucks and AMD would still make higher profit margin on it as it did with Vegas and even Polaris. It's clearly a rip off north worth paying like everything else on today's dGPU market. I hate people who defend these 2 greedy corporations and not even owning their stocks. We're voting with our wallets. There will be no back to normal once mining craze ends IF gamers buy GPUs at these MSRPs or higher.

It doesn't work that way, there is much more to a GPU than just shaders, and even the Infinity cache is more than half Navi 21 (96MB vs 128MB). So Navi 22 is definitely more than 260mm2. And the maturity of 7nm is irrelevant, the shortage is not because of low yield, it's because TSMC can only make ~140,000 7nm wafers a month, and that has to satisfy the world's demand for 7nm (PS5, XBSX, 3000/5000 series Ryzen, Epyc, 4000 series Ryzen mobile, various Qualcomm, MediaTek, Broadcom, etc).

The days of yearly die shrinks on similarly priced fabrication processes are over. $/transistor is going up with each node, that is just a reality brought on by physics. Could AMD sell a bit cheaper? Sure, and it wouldn't make a lick of difference to the prices we pay. The only thing that will drive prices down is competition and supply commensurate with demand. We finally have competition, we just need supply from TSMC/Samsung to increase which can take years unfortunately.

 

[kruk]

TSMC's 7nm process is very mature by now and we're talking about really small die size here, so number of non-functional dies per water must be really low, meaning it's cheap to produce. 6700XT could cost 350 bucks and AMD would still make higher profit margin on it as it did on Vegas and even Polaris. It's clearly a rip off not worth paying like everything else on today's dGPU market.

Yeah, but nVidia dictates the MSRP prices and they just follow. Since the wafers are limited, it wouldn't make sense for them to go into a price war, especially since they can get much more CPUs from a wafer than GPUs (Zen 3 dies are only like 80 mm2) and sell them with bigger profits.

 

[TheinsanegamerN]

AMD has surely lost any remnants of conscience. A 192bit wide GPU for a nice price of $500.

OMFG.

This world has gone crazy. Competition? Customers first? Modest margins? Never heard of it!

AMD is the new Apple. It became obvious when they announced the pricing for the Ryzen 5000 series.

Ummm... Quite frankly WGAF? It coud be 32 bit for all I care, if it performs then it performs. I thought that would have been pretty obvious when the geforce 980 was trading blows with the radeon 290 despite the 980 having only 256 bits compared to hawaii's 512.

This is a card that, if it is faster then a 3070, will be twice the speed of a vega 64 with more VRAM for $479. Vega was $650 in 2016, as was the 1080. Doubling performance three tiers down from the high end in two generations is pretty good considering how much trouble AMD has had lately with GPUs, and it is a surprising jump over the 5700xt, which has the same core count.

 

[RedelZaVedno]

It doesn't work that way, there is much more to a GPU than just shaders, and even the Infinity cache is more than half Navi 21 (96MB vs 128MB). So Navi 22 is definitely more than 260mm2. And the maturity of 7nm is irrelevant, the shortage is not because of low yield, it's because TSMC can only make ~140,000 7nm wafers a month, and that has to satisfy the world's demand for 7nm (PS5, XBSX, 3000/5000 series Ryzen, Epyc, 4000 series Ryzen mobile, various Qualcomm, MediaTek, Broadcom, etc).

The days of yearly die shrinks on similarly priced fabrication processes are over. $/transistor is going up with each node, that is just a reality brought on by physics. Could AMD sell a bit cheaper? Sure, and it wouldn't make a lick of difference to the prices we pay. The only thing that will drive prices down is competition and supply commensurate with demand. We finally have competition, we just need supply from TSMC/Samsung to increase which can take years unfortunately.

I'm not buying this argument. Just look at NVidia profit margin graph. They went from 29% (2009) to 63% (2020). Similar is true for AMD after Zen moment: from 22% to 45% in 5 years. TSMC jacking up prices and covid supply chain problems are to blame only partially. Majority of price hikes over the last 6 years can be explained by AMD's & NGreedia's skyrocketing profit margins and people still defend them. I just don't get it why. It's us who're getting milked. It's duopoly bordering on monopoly with the market maker setting up prices and weaker player going along with it in a cartel like arrangement not trying to compete for the market share so they both make more profits and consumers lose.

 

[Taz100420]

If this pulls the same hash rates as a 5700/XT and pulls about same wattage, no miner will dump their 5700s and go to these. Unless the 6700 stay cheaper in price than what they bought their 5700s for, then they'll probably sell their 5700s and buy 6700s lol

I bought my 5700 for $299 in December 2019 myself. *If* the performance of the 6700XT was on par or greater with my 5700, I'd sell it, make twice the profit and get a 6700 XT if they don't sell out in first few seconds lol. I don't see that happening though XD

 

[Mistral]

my guess is these will cost about 700+ CAD and may not reach their MSRP until the launch of the 7000 series

It's not like that wouldn't be 600CAD before taxes even at MSRP...

 

[Chrispy_]

40CU/2560SP up to 2424Mhz on 251 mm2 die for $479??? That's Polaris sized GPU for God sake!
It'll be 12,4 TFlops at best... so 25% faster than 5700XT for 20% higher MSRP and 50% higher AIB prices. 5% added value in best case scenario and 45% price to performance regression in reality. AMD go eat Sh/t you've become worse than Ngreedia!

It's a bigger die than that. VideoCardz and Igor's Lab have both had confirmation that the die area is 335 mm^2
Don't forget, there's a bunch of additional cache that the 5700XT never had, and there's also all of the hardware raytracing units.

Polaris was only 232mm^2 and on top of that GloFo 14nm was an unpopular node at a time when wafer supply was good and GloFo had no other major committments to other customers; AMD managed to get chips out of GloFo at a good price. TMSC's 7FF is a constrained, sought-after, premium node with a lot of bidders pushing the price up and TSMC basically getting whatever they ask for it.

Polaris:
Small die,
cheap node with no competition,
plenty of supply,
cheap PCB with simple VRMs
cheap alu extrusion cooler for 150-180W cards.

Navi 22:
44% larger die, despite a higher density node
the most expensive node currently available,
bidding against many competitors with much deeper pockets than AMD
supply constraints everywhere pushing up prices,
manufacturing difficulties due to COVID and water shortages
higher-end PCB with more layers and more complex power delivery
more expensive vapor-chamber and multi-heatpipe soldered coolers for 230-350W cards.

This is a much closer comparison in terms of transistor count, die area, and board/cooler design to the $699 RadeonVII (7nm Vega64). If you want to make a Polaris comparison, wait for Navi23 which should be much closer to the die size and transistor count of Polaris, TPU seems to have some leaked/rumoured specs here:

AMD Radeon RX 6600 XT Specs

AMD Navi 23, 2589 MHz, 2048 Cores, 128 TMUs, 64 ROPs, 8192 MB GDDR6, 2000 MHz, 128 bit

www.techpowerup.com

 

[B-Real]

Faster than the 3070 (I know they chose many AMD titles, but seeing the Borderlands 3, Gears 5 or WD: Legion results, overall I think it will be a bit better), having +4GB VRAM and being cheaper? Well, this won't last for long even if miners wouldn't buy it.

40CU/2560SP up to 2424Mhz on 251 mm2 die for $479??? That's Polaris sized GPU for God sake!
It'll be 12,4 TFlops at best... so 25% faster than 5700XT for 20% higher MSRP and 50% higher AIB prices. 5% added value in best case scenario and 45% price to performance regression in reality. AMD go eat Sh/t you've become worse than Ngreedia!

What are you talking about? This will be a bit faster than the 3070 (around +35% compared to the 5700 XT) given the benchmarks + having +4GB VRAM + have lower MSRP.

 

[TheinsanegamerN]

Faster than the 3070 (I know they chose many AMD titles, but seeing the Borderlands 3, Gears 5 or WD: Legion results, overall I think it will be a bit better), having +4GB VRAM and being cheaper? Well, this won't last for long even if miners wouldn't buy it.




What are you talking about? This will be a bit faster than the 3070 (around +35% compared to the 5700 XT) given the benchmarks + having +4GB VRAM + have lower MSRP.

Here's the issue. The 5700xt was $399 at launch. This is "promised" at $479.

So you get 35% more performance, but it costs 20% more. Real world price will likely be closer to 40-50% mroe then a 5700xt. so IRL pricing this will likely end up another turing level generation where price/perf barely moves.

 

[MxPhenom 216]

AMD has surely lost any remnants of conscience. A 192bit wide GPU for a nice price of $500.

OMFG.

This world has gone crazy. Competition? Customers first? Modest margins? Never heard of it!

AMD is the new Apple. It became obvious when they announced the pricing for the Ryzen 5000 series.

Holy f**** the exaggeration. Chill, go take a xanax or something

 

[RedelZaVedno]

It's a bigger die than that. VideoCardz and Igor's Lab have both had confirmation that the die area is 335 mm^2
Don't forget, there's a bunch of additional cache that the 5700XT never had, and there's also all of the hardware raytracing units.

Polaris was only 232mm^2 and on top of that GloFo 14nm was an unpopular node at a time when wafer supply was good and GloFo had no other major committments to other customers; AMD managed to get chips out of GloFo at a good price. TMSC's 7FF is a constrained, sought-after, premium node with a lot of bidders pushing the price up and TSMC basically getting whatever they ask for it.

Polaris = small die on a cheap node with no competition and plenty of supply, cheap PCB and cheap cooler for a 150W part.
Navi 22 = 44% larger die on the most expensive node currently available, bidding against many competitors during a time of supply constraint to make a high-end PCB and expensive coolers for 230-350W cards.

Then RDNA2 arch has to have a very large design. 3060TI has 38CU and is 'only' 392 mm2 in size on de facto 10nm die. I'm still standing behind my estimation of 250 to 275 mm2. We'll see who's right soon enough.

What are you talking about? This will be a bit faster than the 3070 (around +35% compared to the 5700 XT) given the benchmarks + having +4GB VRAM + have lower MSRP.

I don't know where you got 35% number, but from my calculations, based on AMDs own figures (40CU/2560SP up to 2424Mhz) it will have 12,4 TFlops performance at best... that's around 25% faster than 5700XT... Benchmarks AMD showed were with SAM enabled on 6700XT. We'll have to wait for 3rd party independent benchmarks first, but again math tells me it'll probably be on pair with 3060TI (or within 5%) with near zero room for OC.

I'm far away from being Ngreedia's fanboy (I bought 2900 GT, 4870, 6870, R290, RX480 in the past), but I have to say it: $479 MSRP is a joke. No DLSS, worse RT, worse encoder, no supported Adobe Premiere/After effects GPU acceleration... The only plus over 3060TI I can see is +4 gigs of VRAM. This GPU should be priced at $349-379 imho.

 

[Manoa]

AMD drivers suckx: All You Need to Know​

 

[Tom Yum]

Then RDNA2 arch has to have a very large design. 3060TI has 38CU and is 'only' 392 mm2 in size on de facto 10nm die. I'm still standing behind my estimation of 250 to 275 mm2. We'll see who's right soon enough.


I don't know where you got 35% number, but from my calculations, based on AMDs own figures (40CU/2560SP up to 2424Mhz) it will have 12,4 TFlops performance at best... that's around 25% faster than 5700XT... Benchmarks AMD showed were with SAM enabled on 6700XT. We'll have to wait for 3rd party independent benchmarks first, but again math tells me it'll probably be on pair with 3060TI with near zero room for OC.

We'll have to wait until reviews, but you are also seemingly forgetting 5700XT is RDNA1, 6700XT is RDNA2 which brought in significant changes that increases the performance of each CU. You can't such scale 5700XT by the 6700XT clock speed to make an estimate of performance. You also can't translate Tflop figures to framerate on different architectures.

 

[Camm]

AMD has surely lost any remnants of conscience. A 192bit wide GPU for a nice price of $500.

+ 96MB of Infinity Cache, of which AMD detailed in the 256bit + 128mb 6800 series effectively extends bandwidth capacity, reducing the need for such wide buses. That 96mb of cache isn't exactly friendly to put on die (chews up a fair bit of space), so your assertions are pretty baseless.

The bigger issue IMO is that the card is 10w more than 3070 without quite matching its performance, likely pointing to these cards being pretty heavily clocked. I need a card for the HTPC so I might grab one before the prices skyrocket like every other card launch, but I have serious questions about how good this card actually is.

 

[Chrispy_]

Then RDNA2 arch has to have a very large design. 3060TI has 38CU and is 'only' 392 mm2 in size on de facto 10nm die. I'm still standing behind my estimation of 250 to 275 mm2. We'll see who's right soon enough.


I don't know where you got 35% number, but from my calculations, based on AMDs own figures (40CU/2560SP up to 2424Mhz) it will have 12,4 TFlops performance at best... that's around 25% faster than 5700XT... Benchmarks AMD showed were with SAM enabled on 6700XT. We'll have to wait for 3rd party independent benchmarks first, but again math tells me it'll probably be on pair with 3060TI with zero room left to OC it.

Ampere doesn't have an additional 96MB of cache on die, so yes, it's a smaller design "per compute unit" than Ampere. All Ampere cards are Samsung 8nm, btw - the only 10nm GPUs I'm aware of are Intel's Xe dGPUs.

You don't need to see who's right - I wasn't guessing the die size, Navy flounder (codename for Navi22) was confirmed a 335mm die when information intended for regulators was leaked last year. Since then we've had additional confirmation from manufacturing leaks and photos. Sapphire, who make the 6700XT reference cards, have had these things out in the wild for over a month already. For it to be available to buy in the US in three weeks time, the finished product had to leave the factory a month ago

 

[RedelZaVedno]

We'll have to wait until reviews, but you are also seemingly forgetting 5700XT is RDNA1, 6700XT is RDNA2 which brought in significant changes that increases the performance of each CU. You can't such scale 5700XT by the 6700XT clock speed to make an estimate of performance. You also can't translate Tflop figures to framerate on different architectures.

True, but I can extrapolate from 6800(XT)/6900(XT) numbers. I know most games prefer higher frequency over larger number of shading units for now, but I still can't come to +35% over 5700XT. All I come up with is +30% with lots of math and mental gymnastics. That's why I still stand behind within 5% 3060TI performance number.

Ampere doesn't have an additional 96MB of cache on die, so yes, it's a smaller design "per compute unit" than RDNA2. All Ampere cards are Samsung 8nm, btw.

You don't need to see who's right - I wasn't guessing the die size, Navy flounder (codename for Navi22) was confirmed a 335mm die when information intended for regulators was leaked last year. Since then we've had additional confirmation from manufacturing leaks and photos. Sapphire, who make the 6700XT reference cards, have had these things out in the wild for over a month already. For it to be available to buy in the US in three weeks time, the finished product had to leave the factory a month ago

Samsung's '8 nm' roughly equals TSMC's 10 nm in density (Corteks had quite in deep analysis of both nodes). Yeah I see it now, it's 334.54 mm2, but still not a whole lot bigger than full RDNA1 die.

 

[windwhirl]

AMD drivers suckx: All You Need to Know​

Not gonna lie, everyone always complained about a million problems while I was always just fine here lol

 

[Chrispy_]

Samsung's '8 nm' roughly equals TSMC's 10 nm in density (Corteks had quite in deep analysis of both nodes). Yeah I see it now, it's 334.54 mm2, but still not a whole lot bigger than full RDNA1 die.

Oh right, that's where you were going with the 10nm comparison....
I wouldn't get too worried about die size comparisons that aren't on the same process. Transistor count is a more accurate gauge of how "high-end" any given product is, though you do have to account for major shifts in architecture, such as Pascal vs Turing that added an enormous amount of extra transistors for RT cores and Tensor cores.

 

[medi01]

IMO 230W is too much for my tastes on a 335 mm2 die.

I'm interested in seeing how well this performs at 150W. If it can run north of 2GHz whilst sipping power, I'll buy one - assuming I can actually buy one.

I wonder what you say about 3070.

I bet performance figures are with SAM

When performance figures are with SAM, AMD explicitly states so.

 

[Bruno_O]

IMO 230W is too much for my tastes on a 335 mm2 die.

I'm interested in seeing how well this performs at 150W. If it can run north of 2GHz whilst sipping power, I'll buy one - assuming I can actually buy one.

agreed
happy I went with the Powercolor 6800 Fighter (non xt), with some undervolting I'm running stock <200W total on my HTPC (great card btw, awesome value in AU/NZ).

 

[Chrispy_]

I wonder what you say about 3070.

Same thing. The 3070 actually has a reasonable performance/Watt but I'm simply not interested in >200W cards.
I like my PCs to be small and inaudible, so low power consumption is a pretty high priority for me.

I have plenty of super hot and noisy boxes at work with threadrippers or multiple 2080Ti cards in them but I game and watch movies with speakers, not headphones. I like maximum dynamic range and to hear all of the subtlest, quietest details over the sound of fans whirring.