AMD RDNA3 Specifications Discussion Thread

which seems to marry up with Super-SIMD as I wrote about here:

This just reminded me of an old discussion about terminology from 2011 (!) where this was said:
supervector or supersimd makes sense but good luck getting people to use that!
:LOL:
Sorry for the flashback.
 
For those who missed it, I know the 0.4 pJ/bit was talked about a few pages back, but AMD also provided a direct comparison to the power consumption of keeping the infinity cache on the same die. Roughly 4x the power consumption per bit transferred, by moving the GDDR6 + Infinity cache to chiplets. Certainly lowers the price and increases the yield, but as I've mentioned before, you pay the price in power consumption.

InterconnectPicojoules per Bit (pJ/b)
On-die0.1
Foveros0.2
EMIB0.3
UCIe0.25-0.5
Infinity Fabric (Navi 31)0.4
TSMC CoWoS0.56
Bunch of Wires (BoW)0.5-0.7
Infinity Fabric (Zen 4)???
NVLink-C2C1.3
Infinity Fabric (Zen 3)1.5 (?)
 
It seems there's another slide deck, as seen here:


image.php

It's not to bad comparatively. to the 4080 . I think it will look even better against the more similarly priced 4080 12 gig
 
It's not to bad comparatively. to the 4080
7900XTX will be about 25% slower than 4080 here according to my napkin math while being just 17% cheaper.

I think it will look even better against the more similarly priced 4080 12 gig
"4080 12G" will probably be close to 7900XTX in RT while costing some $800 or something.

7900XT seems like a more interesting product for those who "don't care about RT" as it will likely be on par with 4080 in "rasterization" while being 25% cheaper.
 
There's no 4080 12GB anymore. They'll probably call it 4070 Ti when they relaunch it

I know but we don't have a new name for it so since we all know it by the 4080 12 gig I keep calling it that. That will be slower than the 4080 16 gig and cost roughly the same as the amd cards
 
I suspect you're right, but it wildly contradicts the "extract maximum value from each transistor" statement on that first slide I included.

If the clockspeed could reach 3ghz+ then performance per area would go up, by a low amount. As the switch to 5nm should've helped with that anyway it seems like the CU restructuring is close to a wash in terms of any improvement in perf/area or perf/watt.

That would make the 4080 33% faster on average, which seems too high, given its margin over the 3090 Ti.

The leaked 4080 benchmarks don't place it as super performant or anything. The 7900xtx will probably be fine in terms of perf/$
 
The leaked 4080 benchmarks don't place it as super performant or anything. The 7900xtx will probably be fine in terms of perf/$

Leaked? Official 4080 reviews and benchmarks came out today. Are you referring to the 4080 12 GB? The post I responded to was talking about the 16 GB card.
 
Leaked? Official 4080 reviews and benchmarks came out today. Are you referring to the 4080 12 GB? The post I responded to was talking about the 16 GB card.

Yeah! Oop just saw the reviews are out today.

Anyway, glancing at the reviews and benchmarks, looks like the 7900xtx is priced perfectly fine to compete with the 4080.
 
I bet amd wished the 4080 12gig card was still releasing in Dec. That would have been great for them since the 790xt would have been the same price and the 7900xtx would have been just $100 more
Almost makes one wonder if NV pulled a reverse jebait here…
 
Looks like 4080 will be about 25% faster with RT while 7900xtx will be about 25% faster without.

Seems to heavily depend on games. Stuff built for Nvidia, like Control and Cyberpunk, look to heavily favor Nvidia cards. While stuff just built to work well even if it's pretty heavy and even straight up dependent on RT, Metro EE is probably the prime example here, the 4080 looks about 16% faster at 20%+ the cost (depending on how many FE models there are, which are cheaper than board partner ones). So going into the future, even on RT titles the 7900xtx is probably solid for like 95% of titles in price/perf. Though "almost a wash" isn't what's wanted for price/perf improvement in generations here though.

Hopefully Navi 32 is delayed until the clockspeeds really can hit 3+ghz or so. That would put it at 80% of the performance of a 7900xtx at like, $649 (or less).
 
Last edited:
Back
Top