Question Speculation: RDNA2 + CDNA Architectures thread

[uzzi38]

All die sizes are within 5mm^2. The poster here has been right on some things in the past afaik, and to his credit was the first to saying 505mm^2 for Navi21, which other people have backed up. Even still though, take the following with a pich of salt.

Navi21 - 505mm^2

Navi22 - 340mm^2

Navi23 - 240mm^2

Source is the following post: https://www.ptt.cc/bbs/PC_Shopping/M.1588075782.A.C1E.html

 

[dzoni2k2]

AMD, AMD what are you up to?

Complete world domination.

 

[jamescox]

Not with a 192-bit bus it won't. At least definitely not at 4K. NVIDIA is using a smaller bus size, but they have GDDR6X to make up for that. By all reports AMD is still using GDDR6. By comparison, the 1080ti and 2080ti both used a 352-bit bus.

According to NVIDIA:

1080ti -> 11gbps
2080ti -> 14gbps
3070 -> 16 gbps

AMD? 8gbps?

EDIT The RX 5700 XT apparently had 14gbps memory, so the 40CU RX 6000 equivalent will have 10 or 12 gbps.

That is where the “infinity cache“ makes up for the lower bandwidth, right? The statements about disrupting 4K gaming seem to indicate that they expect to do very well at 4K.

 

[leoneazzurro]

https://twitter.com/x/status/1308699315804798976

Lol if true

 

[kurosaki]

Navi, Flounders Edition. There, Nvidia. Now we have written it out. Your move.

 

[psolord]

Geforce, Soimpson Edition! There!

 

[kurosaki]

Geforce, Soimpson Edition! There!

As in reference to the GloFo Soi -process? 😁

 

[Ajay]

Wild speculation: AMD might have actually went out of their way to troll/mislead leakers.

I believe that Navi23 (Dimgrey Cavefish) is actually the 80 CU, 505mm², 384b model we were expecting.

Now there is one main problem with this theory and that is Sienna Cichlid (80 CU, HBM2?) which most people assumed was Navi 21.

However it's common for some products to show up in drivers and never be released. Alternatively, Sienna could be a spin off of Navi 23 for the professional market.

Evidence of this theory include the Navi 21 die circulating on AIBs

https://twitter.com/x/status/1303890583123968000

being much smaller than expected

https://twitter.com/x/status/1303986125225095169

Navi 21 being 256b and Navi 22 192b (https://twitter.com/_rogame/status/1306655000454725636) and finally the fact that Dimgrey hexcode is #696969 (https://www.colorhexa.com/696969) because obviously Navi 23 is the 6900XT!

Wow, having a top card (N21) with only a 256b wide dram interface seems nuts; especially so if AMD isn't using GDDR6X.

 

[Stuka87]

NM

 

[Stuka87]

Wow, having a top card (N21) with only a 256b wide dram interface seems nuts; especially so if AMD isn't using GDDR6X.

I am still thinking it will be HBM.

 

[raghu78]

I am still thinking it will be HBM.

If AMD has intentionally planted incorrect firmware info so that leakers don't get the actual specs then I think even 80 CU might be wrong. I would not be surprised to see 96 CU and 2048 bit HBM2E for the 505 sq mm flagship die.

 

[CastleBravo]

With all the old rumors of Navi21 having both a GDDR6 controller and HBM IO built in, would it make sense to have mixed VRAM? Could they do 8/16GB 256bit GDDR6, and 4/8GB of HBM2E on the same card?

 

[Heartbreaker]

With all the old rumors of Navi21 having both a GDDR6 controller and HBM IO built in, would it make sense to have mixed VRAM? Could they do 8/16GB 256bit GDDR6, and 4/8GB of HBM2E on the same card?

Extremely unlikely.

 

[Macros96]

Wow, having a top card (N21) with only a 256b wide dram interface seems nuts; especially so if AMD isn't using GDDR6X.

If we know this is nuts, then AMD knows this is nuts. So there are only so many scenarios:

1.) They are ceding the high end yet again.
2.) They have some secret sauce compression or caching scheme up their sleeve.
3.) They have successfully avoided accurate leaks and there is a big-die plus HBM card at the top of the stack.

I'm fervently hoping for #3, but know it will be expensive.

 

[Gideon]

Going by the "Navi 21 might not be 256 bit" post, it does indeed seem that AMD is trolling the leakers hard.

 

[Konan]

That 40 CU die will be formidable opponent for RTX 3070. VERY formidable .

Navi 23 - RTX 2080 Super performance levels +10%, @150W TBP.

A 2080S is 16% behind a 2080Ti at 4k so N23 in the target you said you got, still won't beat a 2080Ti.
Basically a 40CU is still going to loose comfortably to the 3070. It's target competitor will more be like the 3060Ti.
Nvidia stated that the 3070 will beat the 2080Ti (still pending judgement on that one in a few weeks)
The 40CU 5700XT is 50% behind the 2080Ti at 4K I don't see a 40CU RDNA2 making at least a 40% leap sorry.
The N22 is the the card that will beat the 3070 (until the S or ti comes out that is) but loose to the 3080.

 

[Glo.]

A 2080S is 16% behind a 2080Ti at 4k so N23 in the target you said you got, still won't beat a 2080Ti.
Basically a 40CU is still going to loose comfortably to the 3070. It's target competitor will more be like the 3060Ti.
Nvidia stated that the 3070 will beat the 2080Ti (still pending judgement on that one in a few weeks)
The 40CU 5700XT is 50% behind the 2080Ti at 4K I don't see a 40CU RDNA2 making at least a 40% leap sorry.
The N22 is the the card that will beat the 3070 (until the S or ti comes out that is) but loose to the 3080.

You can't see how its possible.

Hmmm.

And how is it possible that 44 SM RTX 3070 GPU can be on par, or slightly slower than 68 CU RTX 2080 Ti?

So why would it be impossible for 40CU GPU, with much higher clock speeds to compete with RTX 3070, considering that RTX 3070 also has roughly around the same amount of Compute Units?

AMD is extremely underestimated. Ridiculously underestimated, I'd say given their latest track record, both on GPU and CPU side of things.

 

[dr1337]

You can't see how its possible.

Hmmm.

And how is it possible that 44 SM RTX 3070 GPU can be on par, or slightly slower than 68 CU RTX 2080 Ti?

So why would it be impossible for 40CU GPU, with much higher clock speeds to compete with RTX 3070, considering that RTX 3070 also has roughly around the same amount of Compute Units?

AMD is extremely underestimated. Ridiculously underestimated, I'd say given their latest track record, both on GPU and CPU side of things.

Nvidia literally doubled the fp32 capabilities of turing with ampere. The 3070 at 44sm vs an RDNA2 card with 40CU on paper is a total knockout. Even if AMD somehow managed to wring 20% more performance per cu out of RDNA2 they're still going to be behind the 3070 due to its 4 extra streaming modules.

It would take some yet announced unknown technology for AMD to beat a 3070 with less streaming processors. And tbh if they actually can do that then it means the 6900xt might actually be on par with the 3090. All of that said we don't have a single piece of information that would imply any of this. The only speculation that can be confident is saying that a 40cu RDNA2 card will be much more efficient than the 44sm 3070. But faster? Yeah nobody except AMD knows that.

 

[Olikan]

If AMD has intentionally planted incorrect firmware info so that leakers don't get the actual specs then I think even 80 CU might be wrong. I would not be surprised to see 96 CU and 2048 bit HBM2E for the 505 sq mm flagship die.

Are we sure the 505mm2 is not ARCTURUS?
Rumors say they doubled the CUs of Radeon7 (is 331mm2), but removed display engines

 

[Zstream]

If we know this is nuts, then AMD knows this is nuts. So there are only so many scenarios:

1.) They are ceding the high end yet again.
2.) They have some secret sauce compression or caching scheme up their sleeve.
3.) They have successfully avoided accurate leaks and there is a big-die plus HBM card at the top of the stack.

I'm fervently hoping for #3, but know it will be expensive.

It's only expensive if consumers are unwilling to pay for it.

 

[sontin]

A 2080S is 16% behind a 2080Ti at 4k so N23 in the target you said you got, still won't beat a 2080Ti.
Basically a 40CU is still going to loose comfortably to the 3070. It's target competitor will more be like the 3060Ti.
Nvidia stated that the 3070 will beat the 2080Ti (still pending judgement on that one in a few weeks)
The 40CU 5700XT is 50% behind the 2080Ti at 4K I don't see a 40CU RDNA2 making at least a 40% leap sorry.
The N22 is the the card that will beat the 3070 (until the S or ti comes out that is) but loose to the 3080.

It would be a 90% improvement in perf/w over the 5700XT. He is just making his numbers up. At the same time Sony cant even ship a 10TFLOPs console with less than 100w for the GPU and memory...

 

[Macros96]

It's only expensive if consumers are unwilling to pay for it.

You are absolutely correct. My wallet is ready!

 

[Konan]

You can't see how its possible.

Hmmm.

And how is it possible that 44 SM RTX 3070 GPU can be on par, or slightly slower than 68 CU RTX 2080 Ti?

So why would it be impossible for 40CU GPU, with much higher clock speeds to compete with RTX 3070, considering that RTX 3070 also has roughly around the same amount of Compute Units?

AMD is extremely underestimated. Ridiculously underestimated, I'd say given their latest track record, both on GPU and CPU side of things.

You said the target of N23 is 2080S + 10% that is already less than the 3070 (if it comes out as officially pitched). Seems like you just exaggerated performance. The real danger is the Vega style exaggerating...

Here is what I think -
N21 80 CU 256bit bus 16Gb HBM2| or G6 > trades with 3080 or just over (RDNA2 more efficient)
N22 60 CU 256bit bus 12/16GB G6 > wants to get close to 3080 (RDNA2 more efficient)
N23 40 CU 192bit bus 12GB G6 > wants to get close to 3070 (RDNA2 more efficient)
N24 20 CU 128bit bus 8GB? G6

I will be v happy if things turn out better of course.

 

[AtenRa]

You said the target of N23 is 2080S + 10% that is already less than the 3070 (if it comes out as officially pitched). Seems like you just exaggerated performance. The real danger is the Vega style exaggerating...

Here is what I think -
N21 80 CU 256bit bus 16Gb HBM2| or G6 > trades with 3080 or just over (RDNA2 more efficient)
N22 60 CU 256bit bus 12/16GB G6 > close to 3080 (RDNA2 more efficient)
N23 40 CU 192bit bus 12GB G6 > close to 3070 (RDNA2 more efficient)
N24 20 CU 128bit bus 8GB? G6

80 CUs with just 256bit memory not even if hell froze.

Also, I cannot see any 40CU die (less that 300mm2) coming close to 3070 (GA104 400mm2)

 

[Glo.]

You said the target of N23 is 2080S + 10% that is already less than the 3070 (if it comes out as officially pitched). Seems like you just exaggerated performance. The real danger is the Vega style exaggerating...

No. The problem is that people exaggerate the performance of RTX 3070.

It won't be faster than RTX 2080 Ti, already. Neither of the two GPUs will be.

And yes, target for 40CU GPU is 35% more performance over RX 5700 XT.

How come? Higher IPC+ higher clock speeds.

 

[Stuka87]

Nvidia literally doubled the fp32 capabilities of turing with ampere.

This is over simplifying the changes that were made. Thats like saying Bulldozer doubled its INT performance because it could now handle two INT threads at the same time.

There are only very specific work loads that can actually make use of the single precision FP32 capabilities. Benchmarks have not shown these changes to increase gaming performance. The 3080 has vastly more hardware than the 2080 Ti, but is not vastly faster. Why then would a 3070 with LESS hardware than the 2080 Ti suddenly be able to catch it?