Discussion RDNA 5 / UDNA (CDNA Next) speculation

[Kepler_L2]

Hope there is a halo sku for RDNA 5

If you mean a 6090 competitor there isn't

 

[Bryo4321]

If you mean a 6090 competitor there isn't

Is it far enough along to ballpark the cu count for top die? Unless rdna5 is drastically changed maybe ~96 again? I would think something like ~128 would be too big.

 

[Kepler_L2]

Is it far enough along to ballpark the cu count for top die? Unless rdna5 is drastically changed maybe ~96 again? I would think something like ~128 would be too big.

Yeah, my guess is the lineup will go from 64CU/256bit+32CU/128-bit to 96CU/384-bit+64CU/256-bit+32CU/128-bit.

 

[poke01]

Yeah, my guess is the lineup will go from 64CU/256bit+32CU/128-bit to 96CU/384-bit+64CU/256-bit+32CU/128-bit.

That’s more than fine. Not everyone need mega heaters in their homes anyway

 

[itsmydamnation]

Yeah, my guess is the lineup will go from 64CU/256bit+32CU/128-bit to 96CU/384-bit+64CU/256-bit+32CU/128-bit.

That would be amazing if that happens , with improved "tech" a real upgrade for a 7900XTX user. Maybe im in to much of the adroc AMD consumer GPU ghetto but i expect 64 CU again 🙂.

 

[Kepler_L2]

That would be amazing if that happens , with improved "tech" a real upgrade for a 7900XTX user. Maybe im in to much of the adroc AMD consumer GPU ghetto but i expect 64 CU again 🙂.

With the Xbox(?) APU leak having 384-bit bus and what looks like 80 CU (very similar layout to Navi21), I don't see RX 10K topping out at 64 CU.

 

[SolidQ]

With the Xbox(?) APU leak having 384-bit bus and what looks like 80 CU (very similar layout to Navi21), I don't see RX 10K topping out at 64 CU.

You mean this leak? and it's Xbox?

 

[Kepler_L2]

You mean this leak? and it's Xbox?

Yeah, codename + giant die size/bus doesn't really fit for Playstation.

 

[marees]

Yeah, codename + giant die size/bus doesn't really fit for Playstation.

Also this is 3nm

The same guy just one or two videos back claimed that Sony would use 2nm for PS6

 

[SolidQ]

Yeah, codename + giant die size/bus doesn't really fit for Playstation.

80CU? looks like more powerful than 9070XT + rumors about 1k$ console

~4090 perf?

 

[marees]

80CU? looks like more powerful than 9070XT + rumors about 1k$ console

~4090 perf?

How many CUs fit with a 384ghz bus ?
Or are those 2 independent

 

[RnR_au]

80CU? looks like more powerful than 9070XT + rumors about 1k$ console

~4090 perf?

I believe the current rumour is whatever it takes to match an 80 series. And I guess that would be a 6080. A 5080 is only 25% faster than a 9070XT on average at 4K.

 

[branch_suggestion]

80CU? looks like more powerful than 9070XT + rumors about 1k$ console

~4090 perf?

Lets see, N48 has roughly 25% more gaming perf/FLOP than N31.
RDNA5 can be expected to increase that more, let's say 35%.
At 3Ghz, 80CU matches 2.5Ghz 96CU in FLOPS.
35% faster than the 7900XTX is a 4090.
Really depends on target power, but I expect it should beat the 5080 and be similar to a 4090 at lower power.
Importantly, that is comfortably 2x PS5 Pro.
You can charge $1599 for it, same as the 4090 launch price for the same perf with a fully complete system that fits in a backpack.
You cannot beat that price/perf with a DIY PC until ATX is replaced with a modern holistic architecture, AMD builds a halo to force NV discounts down the stack or just wait 5 years.

I am NOT a crackpot.

 

[Kepler_L2]

80CU? looks like more powerful than 9070XT + rumors about 1k$ console

~4090 perf?

Should be somewhere between 5080 and 4090

 

[Win2012R2]

I don't see RX 10K topping out at 64 CU.

You reckon chance of 160-ish or just mere 120?

 

[Kepler_L2]

You reckon chance of 160-ish or just mere 120?

96 CU with same config as Navi31 is the most likely imo.

 

[Win2012R2]

96 CU with same config as Navi31 is the most likely imo.

Well, still if it's 384 bits and higher freq, IPC, GDDR7 - should be a nice card if they don't screw up again the odd version.

Should certainly exceed 4090 at 4k easily? As long as they don't skimp on ray tracing - they need parity at least

 

[ToTTenTranz]

96 CU with same config as Navi31 is the most likely imo.

N31 config with >3GHz clocks, GDDR7 and higher IPC than RDNA4 sounds like a beast that could bite a GB202 in its heels, to be honest.

 

[ToTTenTranz]

Yeah, codename + giant die size/bus doesn't really fit for Playstation.

It doesn't fit any console within the usual subsidized model either, to be frank.

 

[jpiniero]

It doesn't fit any console within the usual subsidized model either, to be frank.

Being MLID it's likely fake. The MS console is not going to be cheap tho.

 

[marees]

Being MLID it's likely fake. The MS console is not going to be cheap tho.

Since this is console leak others such as Tom Warren, Jez Corden or Michael Henderson should be able to corroborate this

 

[SolidQ]

As long as they don't skimp on ray tracing - they need parity at least

looking at RDNA4 RT, they should just improve PT. I think it's gonna be fine at RT/PT

 

[SolidQ]

Should be somewhere between 5080 and 4090

So 96CU desktop variant should be near 5090, include higher clocks + GDDR7.

 

[ToTTenTranz]

Being MLID it's likely fake. The MS console is not going to be cheap tho.

It might not be fake. Point was this chiplet solution won't be found in a $500 console unless they're going with an ancient process node.
If instead this is a semicustom chip for Xbox-branded pre-builts or handheld hybrids being sold by Asus, MSI, Lenovo, etc. then it's not going to cost $500 nor $600. We'd be looking at over $1000.

Though Sony is the one who started to wind down the subsidized hardware concept, with the PS5 Pro's MSRP.

 

[Win2012R2]

looking at RDNA4 RT, they should just improve PT. I think it's gonna be fine at RT/PT

They need to double it again on per CU level, so more with more CUs. Nvidia got sloppy with Blackwell, but this is unlikely to continue, AMD must catch up proper next time, ideally be better.