xool
Member
mmh
I just noticed that the MP2888A uses "NVIDIA’s Open VReg specification" which is an open standard, but I couldn't find any evidence of AMD ever having used it ..
Needs checking
-
Hey, guest user. Hope you're enjoying NeoGAF! Have you considered registering for an account? Come join us and add your take to the daily discourse.
- Status
mmh
I just noticed that the MP2888A uses "NVIDIA’s Open VReg specification" which is an open standard, but I couldn't find any evidence of AMD ever having used it ..
Needs checking
LordOfChaos
Member
ArabianPrynce
Member
when i think about i dont think 7nm+ is gonna be a thing next gen. the cpu is zen 2 which is regualar 7nm and 7nm + is for zen 3. How are gonna have a cpu with 7nm and gpu that is 7nm+ and make them compatabile???
SonGoku
Member
Consoles chips are a semi custom design they can be on any node they choose to be.
You said it yourself Zen3 is on 7nm+ the ground work is already laid to port Zen2 to 7nm+. Besides RDNA2 is on 7nm+ it would be more expensive to port the GPU to 7nm than the CPU to 7nm+
Note: Im not claiming 7nm EUV is a done deal but it is within the realm of possibility
Last edited:
Negotiator
Member
Next-gen consoles will most likely use a monolithic APU design (judging by Scarlett), which means you can only use one process for the entire chip.
We're not talking about MCM/chiplets, since that's not possible for APUs (yet). In that case, we could have a 7nm CPU, a 7nm+ GPU and a 12nm I/O die.
Just because Zen 2 uses 7nm (chiplet) + 12nm (I/O die) on PC, it doesn't mean that consoles are going to use those nodes and I explained why above.
luca_29_bg
Member
Uhm interesting, maybe gonzalo it's a laptop apu ?
Last edited:
xool
Member
(1.6GHz / 3.2GHz)
Single core 93.8 Pts
Quad core 375 Pts
Multi-core 1,018 Pts
Compared with a Ryzen 7 3700X (3.6 GHz, turbo 4.1 GHz (avg) ) https://cpu.userbenchmark.com/AMD-Ryzen-7-3700X/Rating/4043
Single 135 Pts
Quad 526 Pts
Multi 1,399 Pts
Also sub Ryzen 2700X performance @3.7/4.05GHz
(seems about right for the low clocks though)
[edit] check the float performance in the links though - the Ryzen 3700X part is getting ~+20% float scores over the int scores, whereas the mysterious sample has lower benchmark metrics of float relative to int .. is this even a Zen2 part ???
[edit2] relative float/int scores look more like the inferior Zen+ part the Ryzen 2700X https://cpu.userbenchmark.com/AMD-Ryzen-7-2700X/Rating/3958
Last edited:
luca_29_bg
Member
It would be really funny if Gonzalo it's not next gen related ahahaha XD All this talk for nothing!
Last edited:
xool
Member
The product code AMD Eng Sample: 100-000000004-15_32/12/18_13F9 doesn't match the previous supposed Sony/Gonzalo part eg AMD Gonzalo ZG16702AE8JB2_32/10/18_13F8
I think it's something else.
I think it's something else.
- It doesn't have Zen2 performance
- Maybe someone spoofed the test with a fake product name for giggles
- Could be something else
- Maybe PS5 isn't full Zen2
CrustyBritches
Gold Member
It would be funny if it was something random like an Alienware steam machine.
xool
Member
The AtariBox ?
sonomamashine
Member
it's a speculation thread, anything is possible, gonzalo was a reference of what was possible and made sense for next gen .
Darius87
Member
of course it's not related there's to much discrepancies to suggest that it is and still some people spin other way around.
i've never believed in gonzalo or second sku for xbox it's fan fiction, from rumor leaks.
Quite a bit you can deduce from the AMD Flute userbench (which is almost certainly the same chip as Gonzalo).
* GameCache™ (L3 Cache) has been cut down from the 32MB in Matisse. Looks like there'll be 16MB total for both CCXes.
* Chip looks to be using 16x8Gbit RAM chips (16GB total).
* Memory latency is very bad at 154ns. About 2x what you'd get on Zen 2 with decent DDR4. This is expected if you use GDDR6 as it trades latency for bandwidth.
* CPU performance is a good chunk below even Zen 1.0 in those tests. About ~75-80% of a 1700X.
Unless AMD has plans to release a Gaming APU this is almost certainly a chip for a next-gen console. If you're still someone expecting 384-bit memory buses and +10TFLOPS you should really rein in your expectations at this point or you'll be disappointed come holiday 2020.
* GameCache™ (L3 Cache) has been cut down from the 32MB in Matisse. Looks like there'll be 16MB total for both CCXes.
* Chip looks to be using 16x8Gbit RAM chips (16GB total).
* Memory latency is very bad at 154ns. About 2x what you'd get on Zen 2 with decent DDR4. This is expected if you use GDDR6 as it trades latency for bandwidth.
* CPU performance is a good chunk below even Zen 1.0 in those tests. About ~75-80% of a 1700X.
Unless AMD has plans to release a Gaming APU this is almost certainly a chip for a next-gen console. If you're still someone expecting 384-bit memory buses and +10TFLOPS you should really rein in your expectations at this point or you'll be disappointed come holiday 2020.
Last edited:
luca_29_bg
Member
Yes something is not right, OPN means qualification sample so it should be soon launched...it doesn't make any sense :/
No, its almost certainly related.
Zen2 was OPN in January, so 6 months. But be aware that devs got final APU dev kits in January last time around, so for consoles I would think its different and this is most likely a norm.
Last edited:
sonomamashine
Member
True but remember it's up to Sony & Ms not AMD .
sonomamashine
Member
Give us your thoughts, spec wise .
luca_29_bg
Member
Yes i know, well we'll see, it's a matter of time!
luca_29_bg
Member
https://pastebin.com/y8qXme7b This was posted ? :/
xool
Member
Erm. Doesn't Matisse have 16MB per 4 core CCX too (with no connection between L3 caches) - ie no change
Yep looks slower than DDR4 (so GDDR obv.).
I disagree (not Gonzalo) - also lower performance because clocks I think.
I also commented that this looks like a pre-Zen2 part based on float vs int performance - which contracticts Cerny's statement on PS5 being Zen2 - it's hard to tell but I think the multicore float performance shows it is inferior to Zen2 and more like Zen+
Last edited:
xool
Member
Built-in with DirectX 12.X, DXR, DirectML and Havok instructions into the chip.
BS or someone took a degree in marketting spin (on existing SIMD capabilties)- no way did Sony or MS pay for extra machine learning instructions on Zen
[edit] 275W power supply enough for 56 Navi CUs .. plus everything else - nice dream .. but nope
Last edited:
Compare Flute
versus 3700X (2x16MB):
and you can clearly see the L3 cache is lower. 3700X can't access the L3 of the other CCX other than through the IO die, hence the big spike after 16MB.
Here's a 2700X (2x8MB):
The increase at 8MB vs 2700X is interesting though, I think it's probably more like 4MB per 2 cores. So 4x4MB.
Unknown U... 16GB
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 MHz
1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024 MB
Is a pretty dead giveaway that there's 16x8Gbit modules. Now I very much doubt consoles will have 512-bit buses so we can deduce that it's likely a 256-bit bus with 2 chips connected to each memory controller.
Lower clocks and if the benchmark hits memory (and hence is affected by much higher memory latency) probably explains the delta.
xool
Member
Yeah you're right - I was reading it wrong 2700X results match 8MB per CCX L3, 3700 results match 16MB per CCX L3, and Flute seems to correspond to 4MB L3 available to a core.Compare Flute
versus 3700X (2x16MB):
and you can clearly see the L3 cache is lower. 3700X can't access the L3 of the other CCX other than through the IO die, hence the big spike after 16MB.
Here's a 2700X (2x8MB):
The increase at 8MB vs 2700X is interesting though, I think it's probably more like 4MB per 2 cores. So 4x4MB.
Sorry saw that after I posted - yeah it's right there in black and white.
[edit] so clamshell configuration just like early PS4 versions.
.. Thing looks like/similar Raven Ridge (Zen APU) with GDDR6 memory ..
..gotta say I really don't like the look of this hardware for next gen
Last edited:
LSWilson
Member
Funny how this leak matches PCB reddit leak where 32GB of RAM at 256bit bus in clamshell where described (for Devkit).
Mind you, chips mentioned there where Samsung 18Gbps, which meant 648GB/s of bandwidth (but I doubt they would be able to use Samsungs 18Gbps in retail, probably 16Gbps therefore 512GB/s).
Nothing out of ordinary. Maybe sony would go for 256 bit bus for smaller die requirements but slightly faster RAM, instead of MS 320bit bus with 14Gbps chips.
Mind you, chips mentioned there where Samsung 18Gbps, which meant 648GB/s of bandwidth (but I doubt they would be able to use Samsungs 18Gbps in retail, probably 16Gbps therefore 512GB/s).
Nothing out of ordinary. Maybe sony would go for 256 bit bus for smaller die requirements but slightly faster RAM, instead of MS 320bit bus with 14Gbps chips.
SpinningBirdKick
Banned
Sony's fortunes are heavily dependant on PlayStation.
The new consoles must offer compelling software until 2028 and beyond.
The best way to do that is with high performance hardware from the initial release date. The majority of the user base will be gaming on the base model. Games are taking longer to develop due to increased complexity and high fidelity asset creation. The console lifecycle will also extend, not shrink, to offer a stable target platform that these huge titles can rely upon to be static from concept, through to release and beyond.
A 14.2TF beast in 2020 will just about be able to hold its own against a 40TF PC in 2028. A 8-9TF weakling, not so much.
If you still think we're getting weak sauce 9TF/16GB/256 bit bus consoles then I don't know what to tell you.
The new consoles must offer compelling software until 2028 and beyond.
The best way to do that is with high performance hardware from the initial release date. The majority of the user base will be gaming on the base model. Games are taking longer to develop due to increased complexity and high fidelity asset creation. The console lifecycle will also extend, not shrink, to offer a stable target platform that these huge titles can rely upon to be static from concept, through to release and beyond.
A 14.2TF beast in 2020 will just about be able to hold its own against a 40TF PC in 2028. A 8-9TF weakling, not so much.
If you still think we're getting weak sauce 9TF/16GB/256 bit bus consoles then I don't know what to tell you.
GermanZepp
Member
Next gen Open World games should look like this. Or not? This match original watch dogs reveal? Or is Better?
Last edited:
Mass Shift
Member
Seriously, that would be like bringing a water pistol to a tank fight.
ResilientBanana
Member
PS5 / Scarlet will most likely be laptop APU's also.
Like bringing 1.2TF machine vs 4TF in 2013. I would look at hard limits of console design (TDP, die size) before going onto "omg they CANNOT have less then 15TF I dont care how they do it".
Because if they went with 9TF that would be top end of what AMD can provide. Much better then back in 2013 when you could have MUCH better GPU then 1.2-18TF ones
Because if they went with 9TF that would be top end of what AMD can provide. Much better then back in 2013 when you could have MUCH better GPU then 1.2-18TF ones
SpinningBirdKick
Banned
At the end of the lifecycle, exclusives will be punching well above their weight. As the strengths and weaknesses of a mature platform will be well known and fully exploited.
Ports from 40TF PC titles will obviously suffer in comparison but that's why exclusives should be picking up the slack at the tail end of the generation. Close out the generation with your best work to give people an appetite for the coming generation and win their hearts, minds (and wallets!) before the next gen arrives.
Sounds familiar, no?
There will not be a 14.2TF console next year. At least not in RDNA Flops.
And a 1.84 TF weakling is doing just fine against 5x stronger PC cards atm.
SpinningBirdKick
Banned
We learn from history that we don't learn from history.
This isn't like PS4 where Sony's back was against the wall after the huge misstep with the PS3.
PS5 has more in common with the PS2 launch. Huge, profitable customer base on the back of the previous console's success, backwards compatibility to lock in the users to the ecosystem. What did we get with the PS2? A great performance jump at a fantastic price.
But we got a weak sauce PS4 previous gen. Suddenly a beast that is up to par with 2080ti is the only accaptable outcome for the next gen? Seems like Sony did very well with their weak sauce console that had CPU for tablets in it.
Last edited:
BTW now 3 leaks (Gonzalo, PCB and this new one) tie in perfectly, independently.
As in, Gonzalo and userbenchmark leak show same APU, that is Navi 10 Lite and 3.2GHZ 8 core CPU. While UB leak gives us CPU score, it also gives us memory (clamshell) arrangement which means 16GB of GDDR6 (look at latency yikes) on 256bit bus. You can see 33GB/s speed for chip, which means memory speed is >16Gbps (and only Samsung 18Gbps bring that, 16GB/s is max 32GB/s).
This in turn ties in with PCB leak that specified 32GB of RAM in dev kit with Samsung 18Gbps chips on 256bit bus, which would result in 512GB/s - 528GB/s.
I am now 99% certain Gonzalo/PCB is legit and, since UB leak tells us L3 cache is halved, full 40CU + RT and 8 core Zen2 on 256bit bus would fit in 316mm² die (again, number that would tie in with exact number from PCB leak - leaked before Zen2 and Navi where shown and before we knew die sizes for each).
As in, Gonzalo and userbenchmark leak show same APU, that is Navi 10 Lite and 3.2GHZ 8 core CPU. While UB leak gives us CPU score, it also gives us memory (clamshell) arrangement which means 16GB of GDDR6 (look at latency yikes) on 256bit bus. You can see 33GB/s speed for chip, which means memory speed is >16Gbps (and only Samsung 18Gbps bring that, 16GB/s is max 32GB/s).
This in turn ties in with PCB leak that specified 32GB of RAM in dev kit with Samsung 18Gbps chips on 256bit bus, which would result in 512GB/s - 528GB/s.
I am now 99% certain Gonzalo/PCB is legit and, since UB leak tells us L3 cache is halved, full 40CU + RT and 8 core Zen2 on 256bit bus would fit in 316mm² die (again, number that would tie in with exact number from PCB leak - leaked before Zen2 and Navi where shown and before we knew die sizes for each).
SpinningBirdKick
Banned
On this I feel we must agree to disagree.
Slow, wide, power efficient, 7nm EUV, 14.2 Navi TF.
But, it seems like 1.8GHZ is sweet spot for RDNA and going down to 1.4.-1.5 wont result in expected TDP gain as in years of GCN.
Look at GCN1 power curve. 850-900MHZ is sweet spot. Sounds familiar? Yea, because these are clocks Sony chose for PS4 and Pro.
Negotiator
Member
Regarding the GDDR latency myth:
forum.beyond3d.com
Is PS4 hampered by its memory system?
I think by giving CPUs a fixed budget, no matter what solution is picked, it should be fairly easy to avoid contention ...
forum.beyond3d.com
sonomamashine
Member
How did previous generations came out ? ... exactly, i think you should learn from history how big companies operate .
If we pursue your logic, why Microsoft launched a weaker console ??????????????
knowing that they have more money than Sony ...
Ms's back was not against the wall, so it means Ms wanted to shoot herself on the foot, also you cant compare Ps2 with todays market, we are not talking about the same parts .
40Tf PC's in 2028 ... what tha hell are you smoking .
Darius87
Member
isn't navi 10 lite should mean cut down version of 5700xt or 5700? but you suggest it's 40 cu's like xt version wich doesn't fit "lite' meaning.
Last edited:
As suggested by Apisak (back in January) Navi 10 could have been aimed as big chip (40CU+) but was scrapped as mid sized one. 10 for AMD is used for first chip on market, doesnt have to be biggest, but usually is. So plan may have been changed to putting out mid tier GPU while Navi 10/Lite remained.
Honestly far, far too many clues point into "Gonzalo is console APU direction), especially after latest leak.
Look at CPU score, it has half the L3 cache (which was predicted for console, as 8MB of additional cache takes ~16mm²).
Look at memory score, 16GB of GDDR6 (SC write score tells us its downclocked 18Gbps chip ~528GB/s could be in retail console if 256bit bus is used).
Pay attention to GPU code - 13F9/Gonzalo. What APU can have 20k FS score, 8 core Zen2 and 16GB of fastest available RAM? Only console and it is PS5, because Scarlett is using 320bit bus and 14Gbps chips so it will have bigger die but only 32GB/s more of bandwidth.
Last edited:
SpinningBirdKick
Banned
Previous generations are easy to summarise.
MS has failed to lead any generation. The closest they came was with the 360, in which ultimately they came last.
MS failed with every aspect of the One. The market responded accordingly.
If you can't see the parallels of the success of the PS1/PS4 and the PS2/PS5 strategy then I feel the reveals are going to be a big shock for you.
Approaching 20TF PC cards next year. Then going on to double that performance over the next 8 years to 40TF seems a reasonable assumption even with Moore's Law breaking down.
IceManCat
Member
Id say 8.3TF - 9TF. 9TF would be absolute max I would expect of these consoles.
Keep it in mind that Navi XT, which easily beats 2070/1080, rarely boosts to 1905MHz (9.7TF) and mostly stays around 1.8GHz range (so around 9.2TF). So it would match Navi XT at best, and be between 5700 and XT in worst.
Racer!
Member
Well its not Sony or Microsoft. This is a qualification sample, so its got to be for a product soon launching.
Last edited:
...Or devkits with actual silicon will have to be provided relatively soon (beginning of 2020 certainly) as it was always the case. You cannot have QS chip for console half a year before launching console like you would with ordinary CPU/GPU.
Racer!
Member
Final devkits 1,5 years before launch? That would be a first.
No, dev kits with final chips. OPN is generally found in chips 6-7 months before retail (in GPU/CPU space) but you would certainly want to have final APU 7-8 months before retail in console space because, unlike retail GPUs and CPUs, you have to provide it to actual developers.
Last edited:
- Status