RTX 3*** Super and 4*** cards in 2022

[Jae-So]

My understanding is AMD's next gen stuff might be MCM. But this far as I say no MCM GPU's have ever come out, monolithic only.

So it will be interesting Nvidia going Monolithic again AMD's MCM. I hope AMD wipe the floor with them and get some sorely needed market share.

Noted, AMD have been talking about it for a while so thought they might have released a product to test it out.

Brief search suggests next set of GPUs as you said.

 

[CuriousTomCat]

Does AMD not use MCM? I thought they experimented with it.

Yes they experimented with MCM a few years ago. I think they called it Ryzen.

 

[Jae-So]

Yes they experimented with MCM a few years ago. I think they called it Ryzen.

ah so they used it on the APU/CPU side, just not discrete GPUs.

 

[HRL]

What is MCM, is that the likely price?

£1900

Hahaha. That’s almost a given at the present prices!

 

[Wrinkly]

Multi Chip Module

So not monolithic. Which is what all GOU’s have been to date.

I think Kap was joking. MCM in Roman numerals is 1900.

 

[TNA]

I think Kap was joking. MCM in Roman numerals is 1900.

I missed that. I was thinking though, I do not see him around much anymore so maybe he just has not been keeping up with it all.

 

[Grim5]

Are the 4-series meant to be MCM? Or is that going to be the 5-series cards?

If there’s only an incremental increase for the 4-series I’ll be skipping it and sticking with my 3090FE for another couple of years. Can’t see there being a massive generational leap until multi-core GPU’s are a thing.

I think 4000 will be monolithic. MCM is new and latency between dies is extremely important- with bad latency the extra performance gain from a double core count 2nd die could result in as little as 20% extra performance.

So while MCM is working now on server accelerators, for latency sensitive applications like games we suspect some extra time will be needed to optimise the technology.


Now this doesn't mean that AMD won't go for MCM for 4000 series, they may very well do that. However due to the latency issues mentioned above I would not expect to see amazing performance but I would expect MCM would result in large increases in prices for GPUs

 

[Jae-So]

I think 4000 will be monolithic. MCM is new and latency between dies is extremely important- with bad latency the extra performance gain from a double core count 2nd die could result in as little as 20% extra performance.

So while MCM is working now on server accelerators, for latency sensitive applications like games we suspect some extra time will be needed to optimise the technology.


Now this doesn't mean that AMD won't go for MCM for 4000 series, they may very well do that. However due to the latency issues mentioned above I would not expect to see amazing performance but I would expect MCM would result in large increases in prices for GPUs

You make some very good points about the challenges they face.

On the pricing point however, isn't the whole point of a chiplet design to increase yield to reduce productions costs / increase profitability? Surely that would help bring prices down?

 

[CuriousTomCat]

You make some very good points about the challenges they face.

On the pricing point however, isn't the whole point of a chiplet design to increase yield to reduce productions costs / increase profitability? Surely that would help bring prices down?

Yes, that's the idea. Two RTX 3060 chiplets connected using an infinity fabric would destroy a 3090 (assuming the width of the memory bus is upgraded also). Two 3060s are cheaper to produce than one 3090.

 

[HRL]

That’s what I thought too. After all, that’s what AMD have been doing with their CPU’s for several years now. However, I’d be hoping for a minimum of 4 cores rather than just 2.

Guess we’ll see by this time next year.

 

[pete910]

That's pretty funny

I had one, it wasn't

Never hear mine to be honest !

 

[Grim5]

Yes, that's the idea. Two RTX 3060 chiplets connected using an infinity fabric would destroy a 3090 (assuming the width of the memory bus is upgraded also). Two 3060s are cheaper to produce than one 3090.

And the latency was low enough - to get 100% scaling would require two chips to have the same latency as a single chip

 

[CuriousTomCat]

And the latency was low enough - to get 100% scaling would require two chips to have the same latency as a single chip

It works in Ryzen 5000. They've really nailed the infinity fabric.

 

[newuser]

It works in Ryzen 5000. They've really nailed the infinity fabric.

You could say that they've thrown down the infinity gauntlet.

 

[RBImGuy]

Upcoming replacement for 6900xt = performance will likely have a 128b bus.
so, massive reduction of wattage for the same performance.
Cant even start thinking faster than that with the rest of the line up
and will at least be slightly cheaper than current 6900xt

 

[JediFragger]

and will at least be slightly cheaper than current 6900xt

AMD has proved lately that they certainly are NOT a charity (nor your best tech friend), they'll get as much out of people as they can just like all companies.

 

[TNA]

AMD has proved lately that they certainly are NOT a charity (nor your best tech friend), they'll get as much out of people as they can just like all companies.

Yep. But thing is, unless you are a youngling (or blind) they did that with their FX line CPU’s on the back off the very successful Athlon stuff. It was not until Core2Duo that they were forced to lower prices.

That’s why I say no point in loyalty. I buy what I fancy or whatever is best price for performance. I care little about who the company is for the most part.

 

[Poneros]

https://videocardz.com/newz/nvidia-...-be-announced-mid-december-launch-mid-january

 

[g67575]

AMD has proved lately that they certainly are NOT a charity (nor your best tech friend), they'll get as much out of people as they can just like all companies.

Companies can either sell products in very high quantities (AMD can't with graphics cards), or they sell a lot less at a high mark-up, nothing new to see here.

What AMD has achieved with the 6800 XT and 6900 XT is impressive when compared to the 5700 XT. The 6900 XT has about double the performance of the 5700 XT, with a TDP thats just 75w higher. It's just a shame that more couldn't be produced.

The card has theoretical performance that is greater than a 100% improvement, because the core config (Core count, TMUs and ROPs) is double that of the 5700 XT, with improved clock speeds also.

 

[kevman]

For shoppers who are not familiar with gpus all these mem configurations are a nightmare