Question Speculation: RDNA3 + CDNA2 Architectures Thread

[uzzi38]

Man I have been dying to make this one for a while now.

First rumours for RDNA3 are here so new thread time!

Just going to start off with this one for now: kopite7kimi on Twitter: "@VideoCardz Ah, I mean a simple mcm design with 10240 cores is not enough. Because the lift from RDNA2 to RDNA3 is much bigger than from RDNA1 to RDNA2. We should expect many big improvements in GFX11. 🤔" / Twitter

 

[Ajay]

Another downright nutty figure is the TDP. In a few generations (maybe just one?) it'll be measured in kilowatts.

And the VideoCardz article says 'passive cooling' - lol. No, they are tower coolers with massive airflow coming from the front. I wouldn't want to forget my earplugs when a rack of those servers start up

 

[Ajay]

https://twitter.com/x/status/1457929587069837315

Just nuts if this is true. It's almost like, what do we need this for? 8k gaming? 240 FPS at 4K (we don't even have an interface that can drive think, IIRC)?

 

[leoneazzurro]

Well that depends on how heavy UE5 will be on our PCs lol. Next-gen engines may tax VGAs even more than today,.

 

[Saylick]

Just nuts if this is true. It's almost like, what do we need this for? 8k gaming? 240 FPS at 4K (we don't even have an interface that can drive think, IIRC)?

3x performance (on paper) of N21 is looking super legit. The move to Infinity Cache is just genius in hindsight, especially since a wider memory bus + more power hungry GDDR6X is just not feasible if you want to scale to higher performance targets. AD102 is supposed to be 384-bit G6X memory just like GA102, but I don't see how they can get 2x the bandwidth to feed 2x the execution units without also using some form of on-die cache to relieve pressure on the memory bus. 512MB of Infinity Cache is going to provide so much more effective bandwidth than a traditional approach. That's 4x the IC for 3x the compute units, so the bandwidth/CU is still increased over RDNA2.

 

[biostud]

https://twitter.com/x/status/1457929587069837315

Just nuts if this is true. It's almost like, what do we need this for? 8k gaming? 240 FPS at 4K (we don't even have an interface that can drive think, IIRC)?

DP1.4 with DSC can do 5120x1440x240Hz, which the G9 Odyssey can push and is close to 4k.

but otherwise I would think DP2.0/hdmi2.1 is a given for next gen cards.

 

[biostud]

3x performance (on paper) of N21 is looking super legit. The move to Infinity Cache is just genius in hindsight, especially since a wider memory bus + more power hungry GDDR6X is just not feasible if you want to scale to higher performance targets. AD102 is supposed to be 384-bit G6X memory just like GA102, but I don't see how they can get 2x the bandwidth to feed 2x the execution units without also using some form of on-die cache to relieve pressure on the memory bus. 512MB of Infinity Cache is going to provide so much more effective bandwidth than a traditional approach. That's 4x the IC for 3x the compute units, so the bandwidth/CU is still increased over RDNA2.

I wonder if HBM2e or HBM3 is going to appear on a gaming card.

 

[Justinus]

https://twitter.com/x/status/1457929587069837315

Just nuts if this is true. It's almost like, what do we need this for? 8k gaming? 240 FPS at 4K (we don't even have an interface that can drive think, IIRC)?

I've got a 4k160hz display and a 6900XT, and more games than I'd like run well below 100 FPS. Shadow of the tomb raider only runs at 50 fps with Ray tracing enabled. Even tripling the performance wouldn't hit 160hz over my setup now.

More realistically though, most games at 4k ultra are running 70-90 FPS, with a few special games running much faster (Doom Eternal, FH4, etc). That config would lead to me being able to actually approach or maintain 160hz in most games.

 

[Ajay]

I've got a 4k160hz display and a 6900XT, and more games than I'd like run well below 100 FPS. Shadow of the tomb raider only runs at 50 fps with Ray tracing enabled. Even tripling the performance wouldn't hit 160hz over my setup now.

More realistically though, most games at 4k ultra are running 70-90 FPS, with a few special games running much faster (Doom Eternal, FH4, etc). That config would lead to me being able to actually approach or maintain 160hz in most games.

Well, I didn't think about that - I never run Ultra settings. Most of the games I play are fast paced shooters, so pretty scenery just doesn't matter to me. Plus, I haven't bought any frame killer games since I'm still on a GTX 1070. My 'skip every other generation' preference got blown out of the water with the current crazy prices.

 

[Stuka87]

My 'skip ever other generation' preference got blown out of the water with the current crazy prices.

Yeah, my idea of "I will get a 5700XT to hold me over for now" turned into "Looks like I will be running this until it dies".

 

[biostud]

Yeah, my idea of "I will get a 5700XT to hold me over for now" turned into "Looks like I will be running this until it dies".

Same with my gtx 1070, I had planned for a rtx 3080 to go with my G9...

 

[Saylick]

Well, I didn't think about that - I never run Ultra settings. Most of the games I play are fast paced shooters, so pretty scenery just doesn't matter to me. Plus, I haven't bought any frame killer games since I'm still on a GTX 1070. My 'skip ever other generation' preference got blown out of the water with the current crazy prices.

At the rate we're going, that statement might end up being true, except it won't be in reference to graphics cards generations but genealogical generations. Your grandkids, if you have any, better start saving up now if they want to afford a graphics card down the road.

I can picture it now... children in primary school will brag to their friends that their grandpops once owned an RTX 3080. All their friends will go "wooooah" in astoundment.

 

[eek2121]

DP1.4 with DSC can do 5120x1440x240Hz, which the G9 Odyssey can push and is close to 4k.

but otherwise I would think DP2.0/hdmi2.1 is a given for next gen cards.

Fun fact: The G9 itself couldn’t handle 240hz (factory defect, it interlaces the second the monitor is pushed somewhat hard.

Well, I didn't think about that - I never run Ultra settings. Most of the games I play are fast paced shooters, so pretty scenery just doesn't matter to me. Plus, I haven't bought any frame killer games since I'm still on a GTX 1070. My 'skip ever other generation' preference got blown out of the water with the current crazy prices.

The 3090 handles ultra settings like a champ, but STILL struggles with RT stuff.

Games will also get more demanding as time goes on.

 

[GodisanAtheist]

At the rate we're going, that statement might end up being true, except it won't be in reference to graphics cards generations but genealogical generations. Your grandkids, if you have any, better start saving up now if they want to afford a graphics card down the road.

I can picture it now... children in primary school will brag to their friends that their grandpops once owned an RTX 3080. All their friends will go "wooooah" in astoundment.

-Whole new meaning to "Legacy" Graphics Cards and Drivers...

 

[eek2121]

Yeah, my idea of "I will get a 5700XT to hold me over for now" turned into "Looks like I will be running this until it dies".

When the RTX 30 series came out, I made the controversial decision to upgrade from my 1080ti to a 3090. Something that I never do (I would usually skip 1 more generation at least, or not buy high end at all). I bit the bullet and I am glad I did every day. Now I have a 3090 and a 1080ti, both of which I purchased for MSRP.

I would love to get a 6900xt for my Linux machine, however.

Bonus points: I am on EVGA’s queue system for a 3080 and 3090. was going to cancel, but I might buy and use for trades instead.

 

[Mopetar]

Frankly any card purchased for MSRP is a good deal in this day and age.

 

[GodisanAtheist]

Frankly any card purchased for MSRP is a good deal in this day and age.

-Depends on the MSRP though.

Reading about people getting their MSI or Gigabyte 3080s for "$1300 MSRP" when the founders edition was $800 makes my head spin a bit.

I weep for the MSRP of even "founders" Edition type cards for this next round.

 

[gdansk]

It's worse, FE was basically $700.

 

[Jwilliams01207]

The Steam hardware survey has its issues, but even as unscientific as it is, it's still unlikely to be wrong by such a margin that 4K displays are really 10% of the market.

People on tech forums get nice new shiny toys a lot more frequently than the average person. 4K TVs are really cheap if you don't give a damn about picture quality, but most people don't even know they can be used as a monitor. To them a TV is a TV and a monitor is a monitor.

 

[Jwilliams01207]

The Steam hardware survey has its issues, but even as unscientific as it is, it's still unlikely to be wrong by such a margin that 4K displays are really 10% of the market.

People on tech forums get nice new shiny toys a lot more frequently than the average person. 4K TVs are really cheap if you don't give a damn about picture quality, but most people don't even know they can be used as a monitor. To them a TV is a TV and a monitor is a monitor.

 

[Mopetar]

The Steam hardware survey has its issues, but even as unscientific as it is, it's still unlikely to be wrong by such a margin that 4K displays are really 10% of the market.

People on tech forums get nice new shiny toys a lot more frequently than the average person. 4K TVs are really cheap if you don't give a damn about picture quality, but most people don't even know they can be used as a monitor. To them a TV is a TV and a monitor is a monitor.

You've got something going on because this is the third thread I've seen you have a double post in.

 

[GodisanAtheist]

You've got something going on because this is the third thread I've seen you have a double post in.

-Gotta pump up that post count somehow...

Its this or neffing in every single active thread he can find.

 

[thecoolnessrune]

Another downright nutty figure is the TDP. In a few generations (maybe just one?) it'll be measured in kilowatts.

600W, allegedly. Water-cooled.

https://twitter.com/x/status/1375443623324491785

In the end, in these applications, only perf/W matters at the end. Even if now a single package has 560W of max power draw, it matters nothing if it has a performance >2x a part with 300W draw. PC environment is different.

For the past year or so we’ve been getting briefed by our Partners indicating ~300W CPU / ~400W GPU configurations air cooled. 400W CPU / 600W GPU configurations liquid cooled. This is why company’s like Cisco released the UCS-X Chassis. 2023 is supposed to see higher wattage GPU configurations than that. 1kW is not far off.

 

[Stuka87]

For the past year or so we’ve been getting briefed by our Partners indicating ~300W CPU / ~400W GPU configurations air cooled. 400W CPU / 600W GPU configurations liquid cooled. This is why company’s like Cisco released the UCS-X Chassis. 2023 is supposed to see higher wattage GPU configurations than that. 1kW is not far off.

12th gen Intel chips are already at the 300W level, which is absolutely nuts. nVidia is already at 400W for GPUs, so it would seem that even higher draws are possible. But even with liquid coolers, 1000W (CPU/GPU combined) is going to be tough to reliably cool. Even my current PC, which is way under that, heats my room up. I would certainly not want a 1000W space heater in here.

 

[thecoolnessrune]

12th gen Intel chips are already at the 300W level, which is absolutely nuts. nVidia is already at 400W for GPUs, so it would seem that even higher draws are possible. But even with liquid coolers, 1000W (CPU/GPU combined) is going to be tough to reliably cool. Even my current PC, which is way under that, heats my room up. I would certainly not want a 1000W space heater in here.

Yep, the big thing here is these partners are scoping 300W Air for 1U Rack Servers, in dual CPU configurations. 600W total for 2 CPUs And still leaving space for the massive amount of RAM modern servers can contain. A couple of years ago when partners were getting ready to introduce their new Xeon Scalable lines, they lamented that it was incredibly difficult to arrange the system in a way to be able to hold the massive sockets as well as the increased RAM Channels. That’s one of the reasons why we say almost all the vendors get rid of things like Mirrored SD Card Readers and settle on M.2 Card modules Only. After all this they released a platform that they were pleased with getting 205W TDPs out of. As TDPs increased across Xeon Scalable, most have adapted by decreasing Ambient Temps required (for instance, instead of 90-95 degree ambients, requiring 75-80 degree ambients). With 3rd gen Xeon Scalable, we’re at 270W TDPs on the flagships, or 540W across 2 CPUs, and most partners feel we’re simply at the limits of what we can cool with air in a 1U form factor in a CPU socket. Hence why we expect to see 300W TDPs on next gen on air (likely Sapphire Rapids), but the CPU itself will likely be cable of 400W TDPs (800W across 2 sockets). Next-Gen servers are where I really expect to see Closed Loop and Facility level liquid cooling to start taking off, because it will no longer be a feature relegated for ultra-density, but rather a feature required to get the biggest return on investment of the silicon purchased.

 

[nicalandia]

Could one use one of these for gaming?

Behold the AMD Instinct MI250X

Behold the AMD Instinct MI250X OAM at SC21

HPE-Cray and AMD have the Shasta EX-series node filled with AMD Instinct MI250X accelerators on the SC21 show floor

www.servethehome.com