AMD RX 7900XT, 90% to 130% faster than 6900XT, MCM, Q4 2022.

[g67575]

I'm thinking that the monolithic chip is probably gonna be pretty decent. Maybe it will be similar to a RX 6900 XT / Navi21 in die size, but built on a new EUV fabrication process.

I suspect that all the rumours about the performance of high end MCM graphics cards will be wrong. That's because even if the early prototypes spec if accurate, it doesn't mean it will actually be a good design that functions to required specifications, and within acceptable power and temperature limits. Some of the rumours are saying more than double the CU count for the top RDNA 3 card (vs the 6900 XT), it's silly. Not to mention, it would likely be very expensive.

 

[eddyr]

I'm thinking that the monolithic chip is probably gonna be pretty decent. Maybe it will be similar to a RX 6900 XT / Navi21 in die size, but built on a new EUV fabrication process.

I suspect that all the rumours about the performance of high end MCM graphics cards will be wrong. That's because even if the early prototypes spec if accurate, it doesn't mean it will actually be a good design that functions to required specifications, and within acceptable power and temperature limits. Some of the rumours are saying more than double the CU count for the top RDNA 3 card (vs the 6900 XT), it's silly. Not to mention, it would likely be very expensive.

The MCM design will give them greater flexibility to configure those power, temperature and performance metrics of the package as whole, without being purely restricted by whatever happens to be the final power/temp/performance characteristics of the die that they are able to produce allows (and then have to try and meet market segmentation requirements - they should be far less shackled, which is also worth adding, it's not necessarily going to end up benefitting the consumer).

 

[BigBANGtheory]

Anyone know why AMD are getting away with smaller dies than Nvidia? Is it process or just a different design?

picture is worth a 1000 words and all that

 

[humbug]

picture is worth a 1000 words and all that

32 cores, pha.....

I'll do you one better.

64 cores

Zen 4 will come in two releases, Genoa with 96 cores 192 threads and Bergamo with 128 cores 256 threads. they will run Crysis!

Zen 5 with 256 cores 512 threads and 384 cores 768 threads.

 

[melmac]

The slide says it is the same. If there where any efficiency gains in the node between RDNA 1 and 2, It was probably so small that AMD didn't even bother to mention it.

While it wasn't massive, it still was an increase over what was in RDNA 1. I remember figures of 10% more performance per watt. Something along those lines moving from N7 to N7P.

 

[g67575]

Doesn't that slide basically just show that the improvements in power efficiency from RDNA 1>RDNA2 were due to improvements in the design of the architecture? E.g. per compute unit, less than half the power usage at approx. 1900mhz, with higher frequencies possible as the power is scaled up.

Interesting graph, it suggests that it would be very power inefficient to ramp up the clock frequency, rather than increase the number of Compute units, if RDNA 3 is anything like RDNA2. So, AMD should only increase the clock rate if they hit a limit on the number of CUs in their high end cards.

I've said it before though, higher clocked cards do tend to offer better value to customers (assuming cooling is sufficient), although not everyone seems to agree with me about this conclusion. I point you to the RTX 3050, with a boost clock of 1777mhz.

 

[Chuk_Chuk]

While it wasn't massive, it still was an increase over what was in RDNA 1. I remember figures of 10% more performance per watt. Something along those lines moving from N7 to N7P.

So I looked this up and it is upto 10% less power for the same speed. Though the graph on there seems to imply that the actual gains on average are quite a bit less than the 10%.

7 nm lithography process - WikiChip

The 7 nanometer (7 nm) lithography process is a technology node semiconductor manufacturing process following the 10 nm process node. Mass production of integrated circuit fabricated using a 7 nm process began in 2018. The process technology will be phased out by leading-edge foundries by...

en.wikichip.org

 

[melmac]

So I looked this up and it is upto 10% less power for the same speed. Though the graph on there seems to imply that the actual gains on average are quite a bit less than the 10%.

7 nm lithography process - WikiChip

The 7 nanometer (7 nm) lithography process is a technology node semiconductor manufacturing process following the 10 nm process node. Mass production of integrated circuit fabricated using a 7 nm process began in 2018. The process technology will be phased out by leading-edge foundries by...

en.wikichip.org

Good find. But, I really think it was more than that. I haven't done any searching sorry. I seem to remember that during one of the AMD presentation days they mentioned the improvement they were getting from moving from N7 to N7P was around 15% People were wondering at the time where AMD were going to get the other 35% that they had in that slide. That it was going to be best case scenario and not real world. Of course it turned out to be a little bit better than that.

Sorry, straying too far from my original point. At the end of the Day TSMC process is much more efficient than the Samsung 8nm one. And we can't really make any estimations about power efficiency based off this generation's products and the small amount of rumours we have about the next generation cards.

 

[Chuk_Chuk]

Good find. But, I really think it was more than that. I haven't done any searching sorry. I seem to remember that during one of the AMD presentation days they mentioned the improvement they were getting from moving from N7 to N7P was around 15% People were wondering at the time where AMD were going to get the other 35% that they had in that slide. That it was going to be best case scenario and not real world. Of course it turned out to be a little bit better than that.

Sorry, straying too far from my original point. At the end of the Day TSMC process is much more efficient than the Samsung 8nm one. And we can't really make any estimations about power efficiency based off this generation's products and the small amount of rumours we have about the next generation cards.

I remember people were questioning how AMD was going make such large gains with a "minor" node revisions and people were doubting they could do it on architect alone. I don't remember any figures about performance though

Regarding Samsungs 8nm, I believe there are ampere GPUs on TSMC 7nm and I thought some people had done a light comparison between them to see how much of an improvement they yield.

 

[Gerard]

The "leakers" are already backtracking on their initial waffle. So as usual hedging their bets both ways so they can point to whatever was closer to reality and say "We WeRe RIgHt11111!!!"

Load of balls. Imo

 

[melmac]

I remember people were questioning how AMD was going make such large gains with a "minor" node revisions and people were doubting they could do it on architect alone. I don't remember any figures about performance though

Regarding Samsungs 8nm, I believe there are ampere GPUs on TSMC 7nm and I thought some people had done a light comparison between them to see how much of an improvement they yield.

The figures came from AMD. They came out before that slide that you linked to, that came from a different presentation.

The only chip that Nvidia make on the TSMC 7nm process is the GA100. There were rumours that Nvidia was switching some of their gaming cards to TSMC in 2021. That didn't happen though.

 

[melmac]

The "leakers" are already backtracking on their initial waffle. So as usual hedging their bets both ways so they can point to whatever was closer to reality and say "We WeRe RIgHt11111!!!"

Load of balls. Imo

LOL are you surprised? It's the same before every release. Like somebody trying to suggest to me earlier in this thread that MILD was a reliable source of info. Like you say they all just hedge their bets and give such open ended statements that they can change to make it sound like they knew all along what was coming.

 

[humbug]

I remember people were questioning how AMD was going make such large gains with a "minor" node revisions and people were doubting they could do it on architect alone. I don't remember any figures about performance though

Regarding Samsungs 8nm, I believe there are ampere GPUs on TSMC 7nm and I thought some people had done a light comparison between them to see how much of an improvement they yield.

That was my memory of it.

No way can AMD double 5700XT performance, it would be a 500 watt space heater......

All that sort of nonsense.

 

[Gerard]

LOL are you surprised? It's the same before every release. Like somebody trying to suggest to me earlier in this thread that MILD was a reliable source of info. Like you say they all just hedge their bets and give such open ended statements that they can change to make it sound like they knew all along what was coming.

Same ***** every year that people invariably fall for, you'd think by now these supposed "leaks" would be ignored as they're as credible as jackanory. Occasionally they get some things right but that's a case of throw enough **** at a wall and some will stick.

 

[melmac]

That was my memory of it.

No way can AMD double 5700XT performance, it would be a 500 watt space heater......

All that sort of nonsense.

Well of course they did. AMD had a history of doing exactly that. That AMD did something amazing with RDNA 2 is fantastic. But, they had slides before that didn't turn out to be even close to reality. Poor Volta for example. Or before the RX 480 was released, the slide that nearly got you banned from the forums. Remember you were saying that it was going 980Ti performance. Gibbo had to come in and say that it wasn't going to be close to that level of performance because you were going ballistic with everyone who didn't agree with you.

 

[Gerard]

. But, they had slides before that didn't turn out to be even close to reality. Poor Volta for example.

Volta never became a gaming gpu because according to jensen it was too expensive to produce. So presumably that's what that "poor volta" thing was about as it was widely expected to be a gaming gpu at some point but it never happened.

Found an article with quotes saying as much: https://www.techspot.com/news/70584-nvidia-volta-gaming-gpus-not-foreseeable-future.html

 

[humbug]

Well of course they did. AMD had a history of doing exactly that. That AMD did something amazing with RDNA 2 is fantastic. But, they had slides before that didn't turn out to be even close to reality. Poor Volta for example. Or before the RX 480 was released, the slide that nearly got you banned from the forums. Remember you were saying that it was going 980Ti performance. Gibbo had to come in and say that it wasn't going to be close to that level of performance because you were going ballistic with everyone who didn't agree with you.

AMD have 50 years of history, the majority of it making very good and efficient products, my knowledge of it spans 25 years.

Due to a combination of a few mistakes and a lack of money AMD had a few years of products that were not good, so did Intel, so did Nvidia.

They were working through their problems and getting past them before this current generation of GPU's, i saw no reason why AMD's bad run should continue when they had already got past it.

 

[Random Guy Esq.]

Lads, step away from the keyboard for a bit.

 

[JediFragger]

LOL bloody kids eh

 

[subbytna]

Holy thread deletion batman Lots of ban hammer waving there but no other action