Question Speculation: RDNA3 + CDNA2 Architectures Thread

[Leeea]

Which in turn results in nVidia not giving them any details on anything until 30 days before launch, as was the case with the 3090.

The more they clamp down, the more that squeezes out the sides. It is a self amplifying circle now.

 

[GodisanAtheist]

Which in turn results in nVidia not giving them any details on anything until 30 days before launch, as was the case with the 3090.

The more they clamp down, the more that squeezes out the sides. It is a self amplifying circle now.

 

[linkgoron]

wrong thread...

 

[leoneazzurro]

This may be more related to RDNA4, however...

DIE STACKING FOR MODULAR PARALLEL PROCESSORS - ADVANCED MICRO DEVICES, INC. (freepatentsonline.com)

Original source post

AMD: RDNA 3 Speculation, Rumours and Discussion

Hmmm... 384bit bus, 192mb cache, twin 96CU GCDs, 2.3ghz clocks, 24gb 24gbps memory, $1599. 384bit bus, 192mb cache, 2 80CU GCD, 2.2ghz clockspeed, 12gb 20gbps memory, $999. 384bit bus, 96mb cache, 1 96CU GCD, 3.2ghz clockspeed, 12gb 18gbps, $749 256bit bus, 128mb cache, 1 60(64?)CU GCD, 3.4ghz...

forum.beyond3d.com

 

[Joe NYC]

This may be more related to RDNA4, however...

DIE STACKING FOR MODULAR PARALLEL PROCESSORS - ADVANCED MICRO DEVICES, INC. (freepatentsonline.com)

Original source post

AMD: RDNA 3 Speculation, Rumours and Discussion

Hmmm... 384bit bus, 192mb cache, twin 96CU GCDs, 2.3ghz clocks, 24gb 24gbps memory, $1599. 384bit bus, 192mb cache, 2 80CU GCD, 2.2ghz clockspeed, 12gb 20gbps memory, $999. 384bit bus, 96mb cache, 1 96CU GCD, 3.2ghz clockspeed, 12gb 18gbps, $749 256bit bus, 128mb cache, 1 60(64?)CU GCD, 3.4ghz...

forum.beyond3d.com

That looks like CDNA3 / Mi300

 

[eek2121]

Unsure about current samples, however it appears that 3.8 ghz is the current boost? We will see.

 

[Tuna-Fish]

AMD really knows how to build anticipation for their launches. Less than a month away from a potentially sea-change GPU launch and we got just about nada here. NV was leaking like a sieve at this point.

I wouldn't say that at all. Ångstronomics wrote up a fairly extensive leak on it. So far, they have been exactly spot on with all their stuff, so I'd expect it to be pretty accurate.

 

[TESKATLIPOKA]

Unsure about current samples, however it appears that 3.8 ghz is the current boost? We will see.

And you heard this from where? Is this based on that tweet about 4GHz?

 

[TESKATLIPOKA]

I wouldn't say that at all. Ångstronomics wrote up a fairly extensive leak on it. So far, they have been exactly spot on with all their stuff, so I'd expect it to be pretty accurate.

I hope they are wrong about only 3 SE for N32 and 30WGP.
4SE and 32WGP would be in line with N33 and N31.

 

[Kaluan]

Some potential insight into the TBPs and tiering (and first batch of SKUs) of the soon to launch RX 7000 GPUs from Enermax's PSU calculator:

(deduction math done by VCZ, not me)

• Radeon RX 7950XT: 822W PSU (GPU TBP ~420W)
• Radeon RX 7900XT: 730W PSU (GPU TBP ~330W)
• Radeon RX 7800XT: 700W PSU (GPU TBP ~300W)
• Radeon RX 7700XT: 598W PSU (GPU TBP ~200W)

Edit: Looks a lot like their high-end hovers around 300-350W, which is good, but they also can scale power up to compete with AD102 SKUs.
Hope this also means cards like 7900XT have ample OC headroom. I dont think 1 or even 2 stacks of SRAM on the MCDs would require that much more power, if almost any.

 

[TESKATLIPOKA]

Some potential insight into the TBPs (and first batch of SKUs) of the soon to launch RX 7000 GPUs from Enermax's PSU calculator:

(deduction math done by VCZ, not me)

• Radeon RX 7950XT: 822W PSU (GPU TBP ~420W)
• Radeon RX 7900XT: 730W PSU (GPU TBP ~330W)
• Radeon RX 7800XT: 700W PSU (GPU TBP ~300W)
• Radeon RX 7700XT: 598W PSU (GPU TBP ~200W)

I saw that too, but I have some unanswered questions.
1.) 79** should be based on N31. 420/330 = 1.27 or 27% higher TBP can be true, higher clocks and 48 vs 40 WGP.
2.) 7800 and 7900's TBP are very close to each other, yet one should be a cutdown N31 and the other full N32 in my opinion.
3.) Is 7700XT a cutdown N32 ? Then the difference in TBP is pretty huge. If It's N33 based, then 200W TBP is too high.

 

[biostud]

So maybe 7950X is a LC version as we have seen before from AMD?

 

[Mopetar]

I could believe higher power figures if Zen 4 is any indication. The N5 process seems more capable of accepting more power or at least AMD has changed their designs with this in mind. NVidia won't be shy about drawing deep from the well of watts, and AMD isn't so good as to be able to go head to head without drinking fro the same well. I'm pretty sure they want to butt heads and show that they're not just a second rate GPU choice for budget-oriented consumers.

 

[Kaluan]

I saw that too, but I have some unanswered questions.
1.) 79** should be based on N31. 420/330 = 1.27 or 27% higher TBP can be true, higher clocks and 48 vs 40 WGP.
2.) 7800 and 7900's TBP are very close to each other, yet one should be a cutdown N31 and the other full N32 in my opinion.
3.) Is 7700XT a cutdown N32 ? Then the difference in TBP is pretty huge. If It's N33 based, then 200W TBP is too high.

1. Also probably stacked SRAM on the MCDs, but like I said earlier in my edit, I don't think that in itself contributes meaningfully to TBP increases.
7900XT is probably 20GB/320bit/5xMCD.
2. I think 7800XT is full/near full fat N32 with clocks cranked up. They made much better graphics resource distribution on the dies this gen (while nVidia did the opposite), they can do it IMHO.
3. 90% sure 7700XT is not N33 based. There were rumors N33 isn't coming out very soon, but they have incentive to release a 7700 class GPU sooner rather than later. Particularly if it stacks up well against N21 SKUs. The pricing on it scares me tho. Probably $600 or more.

I don't think a 192bit 7800XT or 128bit 7700XT make any sense (they're going overall wider this gen after all). But who knows.

 

[Joe NYC]

I saw that too, but I have some unanswered questions.
1.) 79** should be based on N31. 420/330 = 1.27 or 27% higher TBP can be true, higher clocks and 48 vs 40 WGP.
2.) 7800 and 7900's TBP are very close to each other, yet one should be a cutdown N31 and the other full N32 in my opinion.
3.) Is 7700XT a cutdown N32 ? Then the difference in TBP is pretty huge. If It's N33 based, then 200W TBP is too high.

7800 might be very high clocked N32.

Edit: Looks like @Kaluan just said the same...

 

[jpiniero]

3. 90% sure 7700XT is not N33 based. There were rumors N33 isn't coming out very soon, but they have incentive to release a 7700 class GPU sooner rather than later. Particularly if it stacks up well against N21 SKUs. The pricing on it scares me tho. Probably $600 or more.

If anything, there were earlier rumors that N33 was originally intended to be released already. Now who knows.

 

[RnR_au]

The pricing on it scares me tho. Probably $600 or more.

AMD need to compete on price unless they have something similar to the DLSS 3.0 frame interpolation. This will be a real nice option to have for single player games where you can crank the pretties on your 4K monitor while still keeping a decent frame rate. Yeah its fake pixels all the way, with some latency, but this won't matter in single player games.

And from what I understand the N5 GCD's could be smallish is size with everything else on dirt cheap N6, so AMD has the chance to build market share with a decent price despite facing an economic downturn and DLSS 3.0.

 

[Kaluan]

If anything, there were earlier rumors that N33 was originally intended to be released already. Now who knows.

Oh indeed, but if I'm not mistaken they came in the context of "mobile first", so both rumors might be true, in a sense.

RX 7700M/7800M (N33, mobile) later this year or CES 2023 launch and RX 7600 (N33, desktop) mid-next year.
Either way, don't think RX 7600 SKUs are coming out anytime soon on desktop.

AMD need to compete on price unless they have something similar to the DLSS 3.0 frame interpolation. This will be a real nice option to have for single player games where you can crank the pretties on your 4K monitor while still keeping a decent frame rate. Yeah its fake pixels all the way, with some latency, but this won't matter in single player games.

And from what I understand the N5 GCD's could be smallish is size with everything else on dirt cheap N6, so AMD has the chance to build market share with a decent price despite facing an economic downturn and DLSS 3.0.

Well I don't see a cut down N32 7700XT having any trouble beating the pathetic joke that is "4080 12GB", particularly if priced several hundred dollars less. Fake frames or not.

If anything nVidia did terrible optics by naming an underwhelming SKU "4080", when it will potentially be solidly outperformed by a x700 class AMD SKU... that'll also be cheaper. I hope the crap they keep pulling finally starts biting them in the behind. Deep recession and all.

It would also be shortsighted to think AMD is not interested in doing some interesting things with their new RDNA3 feature set as well. I just hope it's at least partially compatible with older architectures.

 

[TESKATLIPOKA]

Oh indeed, but if I'm not mistaken they came in the context of "mobile first", so both rumors might be true, in a sense.

RX 7700M/7800M (N33, mobile) later this year or CES 2023 launch and RX 7600 (N33, desktop) mid-next year.
Either way, don't think RX 7600 SKUs are coming out anytime soon on desktop.

It's mobile first from rumors, so CES 2023.
N32 should also be for laptops, so at least 7800M is N32 based.

 

[Timorous]

I saw that too, but I have some unanswered questions.
1.) 79** should be based on N31. 420/330 = 1.27 or 27% higher TBP can be true, higher clocks and 48 vs 40 WGP.
2.) 7800 and 7900's TBP are very close to each other, yet one should be a cutdown N31 and the other full N32 in my opinion.
3.) Is 7700XT a cutdown N32 ? Then the difference in TBP is pretty huge. If It's N33 based, then 200W TBP is too high.

1) indeed.
2) Looks like 6700XT vs 6800 redux to me with a very close TDP. In actual gaming there is very little difference between 6700XT and 6800 power consumption (2W in the TPU test)
3) Lower clocks, less ram, fewer MCDs and less functional units can make a large difference. Look at N23 vs N22 with 32CUs vs 40CUs and 8GB vs 12GB vram.

 

[TESKATLIPOKA]

1) indeed.
2) Looks like 6700XT vs 6800 redux to me with a very close TDP. In actual gaming there is very little difference between 6700XT and 6800 power consumption (2W in the TPU test)
3) Lower clocks, less ram, fewer MCDs and less functional units can make a large difference. Look at N23 vs N22 with 32CUs vs 40CUs and 8GB vs 12GB vram.

2) I totally forgot about that. I imagined RDNA3 SKUs to be clocked similarly, but that's too naive.
3) I had my doubts If It's N32 based, because of the large difference in TBP(-1/3 of TBP) and the likely cutdown specs(pretty much everything -20-25%), but lower clocks(voltage) could make up for that as you said.
A cutdown N32 with 24WGP, 6144SP, 12GB 192bit GDDR6, 48MB IC, 200W TBP could be a pretty interesting chip.
The question is If those TBP values are correct.

 

[Stuka87]

Some potential insight into the TBPs and tiering (and first batch of SKUs) of the soon to launch RX 7000 GPUs from Enermax's PSU calculator:

(deduction math done by VCZ, not me)

• Radeon RX 7950XT: 822W PSU (GPU TBP ~420W)
• Radeon RX 7900XT: 730W PSU (GPU TBP ~330W)
• Radeon RX 7800XT: 700W PSU (GPU TBP ~300W)
• Radeon RX 7700XT: 598W PSU (GPU TBP ~200W)

Edit: Looks a lot like their high-end hovers around 300-350W, which is good, but they also can scale power up to compete with AD102 SKUs.
Hope this also means cards like 7900XT have ample OC headroom. I dont think 1 or even 2 stacks of SRAM on the MCDs would require that much more power, if almost any.

How would enermax have any of these numbers early? They have zero access to early samples. If anything, this is akin to the place holders that retailers put in before they know the real prices. Which tend to be over the actual prices.

 

[Aapje]

@Stuka87

They don't need early samples to be told the expected wattages. That is exactly what one would expect Nvidia/AMD to tell the PSU makers, so they can plan their production to have the right PSU's ready in sufficient quantity.

 

[Stuka87]

@Stuka87

They don't need early samples to be told the expected wattages. That is exactly what one would expect Nvidia to tell the PSU makers, so they can plan their production to have the right PSU's ready in sufficient quantity.

Ok, but the chart shown was for AMD cards, not nVidia. As far as we know, AMD cards don't require exotic power supplies. So I am still not sure why AMD will tell PSU makers the wattage of specific GPUs way ahead of release.

 

[GodisanAtheist]

Ok, but the chart shown was for AMD cards, not nVidia. As far as we know, AMD cards don't require exotic power supplies. So I am still not sure why AMD will tell PSU makers the wattage of specific GPUs way ahead of release.

Doubt AMD would say for any actual sku, but they might tell a PSU maker "Hey, our top end is going to pulling 420w off the 12v rail, please be wary" and then someone within enermax or just some rando then going about assigning product values etc.