True, but that is also true of all (I hope) types of programming models if the µArch is properly designed to be future proof.
True, but that is also true of all (I hope) types of programming models if the µArch is properly designed to be future proof.Sure it is already supported. But if the GPU has enhanced dynamic execution abilities (reordering, out of order, dynamic allocation, dynamic wavefront sizes) it should get more powerful.
none of that is news and yes it looks like Cray Shasta (EX235a) for obvious reasons.About Mi450
I find it interesting, that MI450X shall get its own LPDDR memory-pool. I speculated for a while now, that this could be introduced with CDNA4 or CDNA5 and would lead to a big benefit. 8x LPDDR packages would also fit on a OAM PCB, there is enough space for that.
the former.Is that in addition to or as a replacement for HBM?
This implies it’s built on the RDNA 4 ISA? That would be sooner than expected, I figured RDNA 5 at the earliest. They’ve been adding architecture features useful for AI and datacenter since RDNA 3. I figured that was their game plan, slowly expand the RDNA ISA until it’s ready. If they think it’s ready, who am I to argue?CDNA 5 ?
9.0.0 -> 9.0.6 -> 9.0.8 -> 9.0.10 -> 9.4.2 -> 9.5.0 -> 12.5?
It's the opposite, they un-ghetto'd DC CUs.That would be sooner than expected, I figured RDNA 5 at the earliest. They’ve been adding architecture features useful for AI and datacenter since RDNA 3. I figured that was their game plan, slowly expand the RDNA ISA until it’s ready. If they think it’s ready, who am I to argue?
You’ll have to elaborate on that for me.It's the opposite, they un-ghetto'd DC CUs.
Seems to be a "fatter" version of RDNA4 with more compute/matrix throughput. RDNA5/gfx13 might not have that much in common.You’ll have to elaborate on that for me.
My understanding is that they’re driving to a unified modular CU and ISA. Then just insert additional IP as appropriate for the target market. With GFX 9 having so much legacy baggage, it would seem prudent to use RDNA as the basis. I can’t see AMD throwing out everything for a clean-sheet design again.
They aren't, but ISAs will converge to a point.My understanding is that they’re driving to a unified modular CU and ISA.
What is even legacy baggage here.With GFX 9 having so much legacy baggage, it would seem prudent to use RDNA as the basis
GCN had terrible PPA in later iterations which had knock-on effects for efficiency of course. IIRC The memory subsystem was also pretty atrocious and caused a lot of issues getting full theoretical performance. I was under the impression CDNA still had to work around these problems despite improvements. Thanks to chip-lets they more or less brute-forced the PPA problem.They aren't, but ISAs will converge to a point.
What is even legacy baggage here.
No it didn't, baby vegas in Renoir/Cezanne was really dang good.GCN had terrible PPA in later iterations which had knock-on effects for efficiency of course
No it was alright, just tricky to scale.The memory subsystem was also pretty atrocious and caused a lot of issues getting full theoretical performance
No.I was under the impression CDNA still had to work around these problems despite improvements
You do understand that MI100 and MI200 are monodie, don't you.Thanks to chip-lets they more or less brute-forced the PPA problem.
Yeah but Vega20 wasn't a good config for gaming, a 48CU with higher clocks would be less area and would do the same there.Navi 10 was 25% smaller than Vega 20 while delivering similar gaming performance (HBM still gave the older card an advantage.)
It just added MALL.Then of course RDNA 2 introduced the full realization of the new memory subsystem
The PPA improvement of RDNA1 was actually quite a disappointment.RDNA is the exact opposite. Navi 10 was 25% smaller than Vega 20 while delivering similar gaming performance (HBM still gave the older card an advantage.) That’s a pretty impressive increase in PPA and efficiency due to the new architecture.
It's two GPUs on the same substrate.I must be hallucinating, because this doesn't look like a single die package...
We're not talking about stacking multiple dies here, MCM is MCM.It's two GPUs on the same substrate.
No different from K80 et al.
It's not MCM, these are two discrete GPUs on the same substrate.We're not talking about stacking multiple dies here, MCM is MCM.
Could be applicable only for CDNA 5 & not RDNA 5, I thinkKepler seems to be claiming UDNA has 256 ALUs per CU, which is 2x from RDNA4 based off this diagram:
I was under the impression that beyond using the significantly less bugged Vega v2 that those iGPUs also used some components from RDNA1?No it didn't, baby vegas in Renoir/Cezanne was really dang good.
It didn't, it's a chopped off Vega20 basically with 1/16DPFP.I was under the impression that beyond using the significantly less bugged Vega v2 that those iGPUs also used some components from RDNA1?
no such thing as UDNA.Kepler seems to be claiming UDNA has 256 ALUs per CU
yea.Could be applicable only for CDNA 5 & not RDNA 5, I think
Because it does not exist.AMD not shared any stuff about UDNA so far (neither any leaks).
Is it even a shader core anymore when the µArch doesn't have any fixed function gfx silicon?Client and DC shader cores live on completely separate tracks.