Medusa Halo with 48 CUs and 384-bit bus according to MLID. No mention of uarch (hopefully it's 12/12.5/13) and if it's 12-channel LP5 or 8-channel LP6.
The reason is simple why Medusa and further mainstream dies would get R3.5. Transistor budget. RDNA3.5 compared to RDNA4 is much less transistor starved.That's why Medusa Point is still RDNA 3.5 I suppose? in an even worse memory bandwidth scenario, we make things more difficult for the iGPU? it's the magic juice that AMD won't give to the masses?
Doesn't 256bit-bus version pretty much mean it must be LPDDR5X? You can't make that width out of 24bit-interface LPDDR6 chips. (Of course, there is always option of MLID being wrong one or multiple out of those claims, or all.)Medusa Halo with 48 CUs and 384-bit bus according to MLID. No mention of uarch (hopefully it's 12/12.5/13) and if it's 12-channel LP5 or 8-channel LP6.
View attachment 118760
and way more area.With RDNA4 It would be way more usable.
The questions is how much more area.and way more area.
Try using the transistor numbers as an indicator of size. RDNA 4 has a very large transistor budget for 64 CU.The questions is how much more area.
If RDAN3.5 offers better performance per area, then ok, but does It? I don' know.
Transistor count is not that important but transistor density.Try using the transistor numbers as an indicator of size. RDNA 4 has a very large transistor budget for 64 CU.
Not for your concern, which is the RDNA 3.5 and RDNA 4 area issue. For a given node you will be able to see the transistor (area) saving by using either, at least approximately. Or are you comparing the two on different nodes, which would be crazy.Transistor count is not that important but transistor density.
RDNA4 has better PPA than RDNA3 for sure. But in mobile PPW is generally first priority. So you have to balance both.
Almost seems like Apple is caring about wanting to compete now, for a change. Or did they get a really crazy deal on RAM?$3.5 for $128GB. Framework might want to update those slides since it’s not shipping till Q3 lol.
Is this better pricing than the M2 gen? It seems the same to me. And Framework/AMD with 2k for 128GB RAM still seems like a good deal.Almost seems like Apple is caring about wanting to compete now, for a change. Or did they get a really crazy deal on RAM?
If it only was 3.5$ I would buy it myself (sorry couldn't help it)$3.5 for $128GB.
You also get a better GPU and CPU and faster I/O.On more serious note, while it offers 2x the bandwith it cost almost twice as much. And 1.5k over my budget.
Yep, the Framework still has its place and much cheaper.Thus I am staying with framework until better deal on Halo appers. Especially seeing staying on x64 saves me migration time.
Well yeah, it cost $4800 to get 128GB in M2 gen but you got a M2 Ultra but the M4 Max for $3.5K is much better.Is this better pricing than the M2 gen?
Lower perf since half the bandwidth.Yep, the Framework still has its place and much cheaper
RDNA3 vs RDNA2 WGPNot for your concern, which is the RDNA 3.5 and RDNA 4 area issue. For a given node you will be able to see the transistor (area) saving by using either, at least approximately. Or are you comparing the two on different nodes, which would be crazy.
it's not.he says that memory may be LPDDR6.
that's not viable at -halo volumes. Get real.IMO, there is a solution: Go to MoP design, and place a fixed price order with one of the memory makers for the memory - which would shield the OEMs from initial price fluctuations.
it's not.
that's not viable at -halo volumes. Get real.