- Status
Nebuchadnezzar
Legend
That's literally the figure for the 2017 EPYC .... things happened long time ago.
I'm still not holding my breath for consumer chiplet gpu's. Datacenter on the other hand I think is inevitable. I'm happy to be wrong in this one though if consumer chiplet gpu's with no downside on perf turn out sooner than later.
Rootax
Veteran
But they're not talking about that. RDNA 3 is their futur gaming arch.
CDNA X is there compute / science arch.
It can be same kind of red herring as gddr6x turned out to be. Sometimes internet thinks it knows and it doesn't know.
Bondrewd
Veteran
But N31 went past the power-on, lol.
I wouldn't know. Perhaps you can link the material that tells what n31 is so rest of us would also know? In internet posts like this are easily ignored without solid sources.
ah, I guess this is the source:
80cu chiplet in 5nm sounds odd. It should be possible to fit those 160cu's to single die. Maybe a test vehicle that isn't necessarily intended to be sold to consumers? 80cu chiplet in 7nm could make more sense considering increasing prices of more advanced processes and that 80cu chip is already in production in 7nm
Last edited:
DegustatoR
Legend
They are also hitting 300W power so putting two of them in one product sounds kinda unrealistic. And that's where 5nm comes in.
Drop the clock little bit, 500W water cooled is doable. 3090 cooler probably could do that on air as well.
DegustatoR
Legend
3090 is 350W card though. Quite a stretch to 500W.
It goes to pretty insane power draws without cooler giving out when overclocked. In it's default config it's very silent card and cooler is not sweating.
Bondrewd
Veteran
LoL Lol, uh, rofl double lol. You win, I'll press ignore button.
I guess I was the ant looking at spectacular light effect that turns out to be sun and magnifying glass. Should have known better.
Frenetic Pony
Veteran
Ok, here we go, again.
A reminder: Bandwidth for desktop GPUs is insanely cheap compared to their compute. The entire GDDR6X bus on a 3090 uses up maybe a handful of watts at most. 7.25 picojoules per byte. Picjoules is ^-12 joules and 1 joule = 1 watt without time. So even a terabyte isn't that much (like 7 watts). Bandwidth is only dear compared to mobile power usage. Compared to desktop/HPC stuff, where compute frequencies hit the exponential growth curve hard while bandwidth costs remain constant, bandwidth power usage is negligible. Remember when citing math, to actually do the math.
A reminder: Bandwidth for desktop GPUs is insanely cheap compared to their compute. The entire GDDR6X bus on a 3090 uses up maybe a handful of watts at most. 7.25 picojoules per byte. Picjoules is ^-12 joules and 1 joule = 1 watt without time. So even a terabyte isn't that much (like 7 watts). Bandwidth is only dear compared to mobile power usage. Compared to desktop/HPC stuff, where compute frequencies hit the exponential growth curve hard while bandwidth costs remain constant, bandwidth power usage is negligible. Remember when citing math, to actually do the math.
It's important to remember the difference between what die to die, or die to chiplet, energy costs using a propriety in socket interconnect, vs what an external interface like DDR or HBM costs.
Also while GPU workloads, probably have higher inter-thread communication, they are also more able to deal with more latency in these workloads - compared to CPU's anyway.
While I'm no expert in these things, I think that taking the existing 80CU RDNA2 core - with 256Mb SRAM, a slight upgrade to the Raytracing functionality, and modification to the memory interface to work with a host I/O die,
shrinking it to 5nm, putting 2 of them, and a I/O die, is a realistic option for a potential 7900 type card.
That 256Mb SRAM on each chiplet helps saves a lot of the bandwidth, some smartish cache management, would allow for some texture duplication in both the SRAM caches.
Also while GPU workloads, probably have higher inter-thread communication, they are also more able to deal with more latency in these workloads - compared to CPU's anyway.
While I'm no expert in these things, I think that taking the existing 80CU RDNA2 core - with 256Mb SRAM, a slight upgrade to the Raytracing functionality, and modification to the memory interface to work with a host I/O die,
shrinking it to 5nm, putting 2 of them, and a I/O die, is a realistic option for a potential 7900 type card.
That 256Mb SRAM on each chiplet helps saves a lot of the bandwidth, some smartish cache management, would allow for some texture duplication in both the SRAM caches.
As Quesa wrote, it's 7.25pJ per bit, så a 3090 uses 19.5*10^9*384*7.25*10^-12 J/s = 54.3W. Not a deal breaker, but not a trivial amount either.
Cheers
- Status
