[MLID] Zen 6 Magnus Leak: AMD's MASSIVE APU for next gen console (+ Medusa Point Specs)
- Thread starter Lunatic_Gamer
- Start date
- Rumor Hype
SegaSnatcher
Member
Give me the realistic expectations of how powerful the new Playstation handheld can be on 3nm node.
Like 3x - 4x more powerful than Switch 2?
Like 3x - 4x more powerful than Switch 2?
Loxus
Member
Could of been done with yields in mind from the get go.
There's some questionable stuff there, but things like AT0, AT2, Magnus, Canis are 100% real. Interestingly he didn't mention the one codename I know related to the PS6, but I never heard of Robin (and that does fit as a codename) so it's possible the codename I know was for an earlier design.
So it's 34+36CU? And they'll only bin one WGP from one half the chip? Surely making it 36+36 or something would be the better option in that case.
Last edited:
Proelite
Member
I've never seen unbalanced binning like that. There is a first for everything though!
Loxus
Member
Isn't Robin/Robin+ old?
It was spotted awhile back and related to Cyan Skillfish.
Cyan Skilfishシリーズ、Navi10_LITE、Navi12_LITE に Robin に Ariel
ROBIN+
The patch seems to be an update of the previous one, but the added DeviceID has changed; it was 0x1400 last time, but now it's 0x143F.
The comments are also interesting, indicating that 0x143F is the DeviceID for ROBIN+.
Loxus
Member
Last edited:
Yes, notice one side is bigger than the other.
Looks like 40+30 or something similar.
I mean I also see PHYs with different widths and thicknesses so you probably shouldn't take MS Paint too seriously.
Loxus
Member
Remember I picked up on it being 192bit-bus.
And it's one of the ways to explain that 70CU number.
Proelite
Member
Mostly likely way is that he made that shit up, like he has in the past.
Last edited:
SolidQ
Member
Tobimacoss
Member
So is this definitely PS6 then?
Proelite
Member
You see there no 256bit bus card.
About CU, it's can be custom design for consoles
Consoles isn't going to have shader arrays and shader engines that are different in sizes from each other.
He has some info and is clearly extrapolating the rest, poorly.
Loxus
Member
Made up what?
70CUs?
PS5 had a total of 72ROPs but only used 64.
I can see the same happening with CUs for yields. An extra WGP placed, so 3 WGPs are disabled instead of 2 like with PS5.
Probably the only thing I don't agree with is it being 68 active CUs. 64 active CUs seems more logical.
Proelite
Member
Made up what?
70CUs?
PS5 had a total of 72ROPs but only used 64.
I can see the same happening with CUs for yields. An extra WGP placed, so 3 WGPs are disabled instead of 2 like with PS5.
Probably the only thing I don't agree with is it being 68 active CUs. 64 active CUs seems more logical.
It's more likely that it's a 2 SE x 2 SA x 9 WGP design (72 CU total) with 1 WGP per SE disabled for yield (68 active).Made up what?
70CUs?
PS5 had a total of 72ROPs but only used 64.
I can see the same happening with CUs for yields. An extra WGP placed, so 3 WGPs are disabled instead of 2 like with PS5.
Probably the only thing I don't agree with is it being 68 active CUs. 64 active CUs seems more logical.
Loxus
Member
If you don't believe his leaks, why are you even in this thread?
Proelite
Member
I believed his PS5 Pro leaks because the info looked legit.
Only an idiot would believe all his leaks.
Last edited:
Loxus
Member
If you've been really following various leaks from various people, you'll realize how accurate he is.
But your problem is you want it to have over 100CUs so badly that you refuse to believe anything else.
Loxus
Member
2 Shader Engines.
I can remember someone saying AMD reworked the SE to carry a lot more CUs.
PS5 Pro is a lot more fascinating than I thought.
Proelite
Member
Going from 384 to 192 bit bus isn't something you do for accurate leaks. How many iterations does his specs for Magnus have to go through before it's accurate.
If he leaks the actual die shot, I'll trust that info 100%.
Looking at that chart my assumption would be that the top bin was 64CU only, surely they'd use the full bin for the DC variants. Plenty of cases of AMD calling the top bin XT rather than XTX etc. With a 44CU bin it should be 2SE rather than 4SE, so 60CU for a PS6? Fits nicely with compatibility for the PS5 Pro.
Last edited:
Proelite
Member
I think for RDNA 3, they increased the WGP to SE ratio from 10:1 to 16:1.
Physiognomonics
Member
I could see PS6 having 68CUs (PS5 Pro has 60CUs). It could also have the same amount as PS5 Pro, 60. PS4 Pro had the same amount of CUs than PS5 so why not doing the same with PS6. AT2 with 64CUs is the first GPU to be released in april 2027 so it's the first RDNA5 GPU and it would be an ideal candidate for PS6.
2 shader engines is logical for a Sony console in order to save die area so they could still aim for more compute / RT / ML. PS5 Pro does seem to be limited mostly by main bandwidth (only 27% more than PS5) and CPU. But it's pretty efficient in other cases even with only 2 SEs. In many cases resolution is increased by 60%, which is almost linear with the increase of Tflops.2 Shader Engines.
I can remember someone saying AMD reworked the SE to carry a lot more CUs.
PS5 Pro is a lot more fascinating than I thought.
Last edited:
HeisenbergFX4
Member
I like reading this stuff, I don't understand 90% of what you guys are saying but for some reason I find it interesting
Proelite
Member
Is the table with some specs on AT0 and AT2 real? The CU counts, if representing total unbinned number and if true, indicate some massive changes to how shader arrays and engines are organized.
Last edited:
mckmas8808
Mckmaster uses MasterCard to buy Slave drives
It's real but could be changed still, designs are not finalized. Imo very likely for either AT0 to be downgraded or AT2 to be upgraded or both, as the gap is far too large.
Last edited:
Proelite
Member
From the CUs config, the whole shader engines divide into at least two shader arrays and then WGP layout seemed to be deprecated.
AT0 is 4 SE x 2 SA x 12 WGP and AT2 is 2 SE x 2 SA x 9 WGP AFAICT, the SAs got massively larger due to uarch changes which reduce front-end bottlenecks.
HeisenbergFX4
Member
Proelite
Member
Ah, so it makes sense that 72CUs and 192-bit??? bus would need 264mm2 on 3nm if the SAs get larger. Looks like infinity cache is gone and is replaced with larger L2 caches. Still inconsistency with the diagram showing 12 phy controllers which is 384 bit bus. You can't have 16 bit GDDR7 controllers.
Last edited:
Tobimacoss
Member
Is it safe to say Magnus will be at least 40-50% more powerful than the PS5 Pro? I can see it being $799 easily.
Proelite
Member
3-3.5ghz clock + dual issue + architectural improvements, easily 2-3x imo. I expect the similar from PS6 too.
Is it even possible to have a 16bit GDDR7 Phy controller? K KeplerL2
All You Need to Know About GDDR7
In this blog post, we explore everything you need to know about Graphics Double Data Rate, most commonly known as GDDR. Since its introduction in 2000,
www.rambus.com
Last edited:
SolidQ
Member
We still need to know, how many perf per CU compare to RDNA4, even if 15%, then 64CU DS variant could be easily around ~4090 perf.
Proelite
Member
SolidQ
Member
9070XT 64CU already ~4080 level
Proelite
Member
Even in path tracing with DLSS vs FSR4? I'm not talking rasterization perf. I'm talking 1:1 running the most demanding visual workloads with all AI features enabled.
Last edited:
Jen_yakzua
Member
Probably closer to $1100
SolidQ
Member
RDNA5 is different story. They already improve massive RT with RDNA4, with RDNA5 is more further improvements.
We still need comfirmation about Traversal Engine in RDNA5
Tobimacoss
Member
4080 level on consoles, I wonder if Sony will still keep the general shape of the PS5.
For Xbox, I expect a more chunky Series X shape. Xbox will have a lot more variants with the OEM devices, but I would love to see MS do a 8-10 inch cube, or maybe even a Tesseract.
Last edited:
Proelite
Member
I believe it's tall cylindrical shaped with two big round half-balls as stands at the bottom. Single slit like vent on top. All black. 12+ inches in height.
HeisenbergFX4
Member
Last edited:
Tobimacoss
Member
8" diameter?
On AMD HW (with GDDR6) each UMC (memory controller) is 32-bits wide and serves 2x 16-bit channels. I don't know how their GDDR7 controller works.
As for the PHYs themselves AMD doesn't design them, they use stuff from either Cadence or Synopsys.
Last edited:
Physiognomonics
Member
You are talking about Magnus, PS6, or both?
TheRedRiders
Member
So Orion is the codename for the PS6 now which MLiD is "confirming", and he stated in the video that it was a "high performance AMD chip" so that sounds very promising.
It's only a matter of time now before we start getting spec leaks, I wonder who's going to spill the beans first? Maybe our beloved K KeplerL2 ?
It's only a matter of time now before we start getting spec leaks, I wonder who's going to spill the beans first? Maybe our beloved K KeplerL2 ?