AMD: RDNA 3 Speculation, Rumours and Discussion

Status
Not open for further replies.
I guess the better question is "why wouldn't they"?

Performance wins benchmarks, and from the graph above, the difference between 100% power target and 60% power target is ~11%, or roughly the difference between the original 3080 10GB and the original 3090, to say nothing of the 3080 12GB and the 3080Ti wedged in between, you're basically giving up an entire product tier's worth of performance for free.

It's not free in terms of binning, board costs, or cooler costs. There's also a less tangible cost to brand perception.

There's a reason Maxwell operated in a high efficiency zone, and that's because it was still able to win there. The benefits of a ~10% bigger win did not justify the costs of attaining that performance.

If Nvidia expected that, say, the fastest Navi 31 would only be 40% faster than a 6950XT, then I expect we would have seen slightly lower performance, but also much lower TDPs, less robust power deliver, and saner cooling.

That the 4090 is rated for 450 watts means Nvidia expects a battle. That both the 4090 founders and AIB cards appear built to accommodate something closer 600W tells you that either at some point Nvidia expected AMD to be even more competitive but now believe they can win with 450W, or that even though they don't expect to win (at least not convincingly) at 450W, at the last minute they decided the market wouldn't accept a 600W card.

The fact that the 4090 often finds itself CPU limited and doesn't need to get to its "poor efficiency zone" is probably something of a blessing though. Nvidia gets to look more efficient now (at the expense of performance), and more performant later (at the expense of efficiency). But in terms of perception efficiency matters most on day 0, while performance is continually re-evaluated as new parts are released.
 
I think the 7900xt is tie or be 10% slower than the 4090.. rdna 3 will double the shaders per cu like ampere so in compute it will compete well.. 12288 shaders at 3 200ghz thats 78 teraflops close to 83 teraflops of a 4090
 
I guess the better question is "why wouldn't they"?


One thing I'll be really curious to see with RDNA3 is whether AMD will have fixed their 'scaling down' issue.

Nvidia cards tend to do exceedingly well when artificially limited by either a lowered power limit, or just low loads such as an older game running with Vsync on and a 60hz cap. Very few reviewers run tests like this, so I'm anxiously awaiting TPU's RDNA3 review to see where it sits in the stack.

Somewhat counterintuitively, the 3090Ti with its out of the box 450w power target is one of the most efficient ways to a 60hz VSync lock in that test, even better than wee GA106 in the RTX 3050. The 4090 is even better in this regard even with its 450w out of the box power target, it happily scales all the way down to 76w in the same test and is nearly the most efficient way to play that particular title.

RDNA2 struggles heavily there, with the higher-end cards drawing less than their maximum, but still quite a lot compared to everything else other than the Arc cards and poor Vega.

Historically power efficiency at low-to-medium loads higher than idle but lower than 100% have been one of the areas where chiplet based designs like Zen2-4 on the CPU side have struggled; you still need to have most if not all of the intra-chiplet communications links powered up, and those usually cost an order of magnitude more power per bit transferred than keeping everything on-die. Doesn't matter much at 100% load, but becomes a significant contribution to overall power draw at part load. Here's hoping RDNA3 has that part nailed down!

I made a chart from TPU's data.
Power draw
V-Sync
Power draw
Gaming
Cyberpunk 2077 1080pCyberpunk 2077 1440p
RTX 409076 W ( only 22% of gaming)346 W138.8 FPS133.8 FPS
RTX 3090 Ti102 W (only 22.9% of gaming)445 W129 FPS93.3 FPS
RX 6950 XT173 W (50.7% of gaming)341 W140.6 FPS86.9 FPS

RX 6950 XT vs RTX 3090 TI vs RTX 4090:
rx6950XT.pngRX 3090Ti.pngRX 4090.png
Nvidia's memory is downclocking to 50%, AMD's stays the same, voltage and frequency reduction is also higher on Nvidia.
This is a clear win for Nvidia. I am pretty pessimistic If AMD will do something about It with RDNA3.
 
Glancing at the 4090 performance numbers, assuming the top end AMD card makes full use of 384bit/24gbps ram it should come in somewhat ahead of the 4090 depending on title, eg doubling the 6950xt's performance on Cyberpunk Psycho RT it'll pull a bit ahead.

Side note: I wonder if they'll restart the branding as "AMD 2XX(X?)". Keeping 7XXX would directly overlap with Zen 4 CPUs, how does the average consumer search for which AMD 7900 they want? That's just going to lead to confusion and possible mixups when there's an obvious and easy solution to fix it.
 
Enthusiast Citizen, the usual Chinese AMD leaker on bilibili, seems less confident now on RDNA3 performance:

My wife is Chinese and she confirmed Google translation these 2 info:
-Availability for second half of December
- Top Navi31 can't beat Ada in both raster and RT

choo choo train will slow down 😢
 
Glancing at the 4090 performance numbers, assuming the top end AMD card makes full use of 384bit/24gbps ram it should come in somewhat ahead of the 4090 depending on title, eg doubling the 6950xt's performance on Cyberpunk Psycho RT it'll pull a bit ahead.

Side note: I wonder if they'll restart the branding as "AMD 2XX(X?)". Keeping 7XXX would directly overlap with Zen 4 CPUs, how does the average consumer search for which AMD 7900 they want? That's just going to lead to confusion and possible mixups when there's an obvious and easy solution to fix it.
Competing in rasterization is possible but we’d need a 3-5x performance increase over a 6900xt to be competitive in RT which is not happening.
 
Competing in rasterization is possible but we’d need a 3-5x performance increase over a 6900xt to be competitive in RT which is not happening.
??? Doesn't AMD's RT scale linearly with compute?
Assuming they kept the same ratio of RT cores, it should be ~2.4x increase. Add in the ~1.3x clock increase(Navi31 @ 2.8ghz) and we are slightly above ~3x RT performance.
If AMD made any other changes to increase RT performance/efficiency, it could be +3x RT performance over Navi 21.
 
??? Doesn't AMD's RT scale linearly with compute?
Assuming they kept the same ratio of RT cores, it should be ~2.4x increase. Add in the ~1.3x clock increase(Navi31 @ 2.8ghz) and we are slightly above ~3x RT performance.
If AMD made any other changes to increase RT performance/efficiency, it could be +3x RT performance over Navi 21.
Since when real world gaming FPS scale linearly with spec FLOPs?
 
Enthusiast Citizen, the usual Chinese AMD leaker on bilibili, seems less confident now on RDNA3 performance:

My wife is Chinese and she confirmed Google translation these 2 info:
-Availability for second half of December
- Top Navi31 can't beat Ada in both raster and RT

choo choo train will slow down 😢
If this true, it would be funny to see the efficiency battle between Navi 31 and the 300W power limited RTX 4090.
 
Enthusiast Citizen, the usual Chinese AMD leaker on bilibili, seems less confident now on RDNA3 performance:

My wife is Chinese and she confirmed Google translation these 2 info:
-Availability for second half of December
- Top Navi31 can't beat Ada in both raster and RT

choo choo train will slow down 😢
I mean hasn't AMD been hinting on "faster isn't always better" or something for some time now?
If they'll be able to put the pressure on 4080s this will be good enough really as those are the ones where prices don't make much sense.
 
I mean hasn't AMD been hinting on "faster isn't always better" or something for some time now?
If they'll be able to put the pressure on 4080s this will be good enough really as those are the ones where prices don't make much sense.
Yeah AMD starting to communicate on efficiency may be the sign that they won't take the crown... But still, lot of space to move. It should be easy to position N31 against 4080 12GB and kill it in value for money. The only question is how aggressive will be AMD. They have proven recently that they are chasing high margins, not people's heart...
 
Yeah AMD starting to communicate on efficiency may be the sign that they won't take the crown... But still, lot of space to move. It should be easy to position N31 against 4080 12GB and kill it in value for money. The only question is how aggressive will be AMD. They have proven recently that they are chasing high margins, not people's heart...
If you said N31 will be positioned against 4080 16GB then I can still understand that, but against the 12GB version, that's impossible.
1. N31 costs more to produce than 4080 12GB.
2. 4080 12GB should perform between 6950XT - 3090Ti.
3. N31 will be significantly faster than 6950XT, I think at least +50% is a given.
Unless you meant in RT, that's a different story.
 
Status
Not open for further replies.
Back
Top