I don't see where hUMA itself stops you from passing to the GPU arbitrary pointers to both code and data. Where is the hardware support for memory access protections implemented?
-
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.
AMD: PlayStation 4 supports hUMA, Xbox One does not
- Thread starter ElTorro
- Start date
I don't see where hUMA itself stops you from passing to the GPU arbitrary pointers to both code and data. Where is the hardware support for memory access protections implemented?
seattle6418
Member
I miss Jeff Rigby. He was all over that Kabini Kaveri stuff when most of us didn't knew what it was and how it was related to PS4.
PS4 has a unified memory that work inline with hUMA concept. it's not AMD problem that MS fucked up the design of X1 w/ esram and slow DDR3 memory.
Now MS is angry and started pulling strings? why not equip X1 with future proof specs to avoid all this in the first place instead of being cheap ass and charging premium for inferior tech!
MS is setting on a mountain of CASH, they could've easily created the most powerful console ever conceived ! but they think people are stupid, they don't understand specs and hardware. funny thing that we have Sony which is in the brink of bankruptcy investing in cutting edge and expensive RAM, equip PS4 with better GPU , and devoted a 5 year project to develop PS4 with an eye to industry future trends and needs to future proof the console. also above all that we find Sony charging less for PS4 despite the very superior tech found inside it compared to X1 !
MS just don't care, they said it themselves.
Now MS is angry and started pulling strings? why not equip X1 with future proof specs to avoid all this in the first place instead of being cheap ass and charging premium for inferior tech!
MS is setting on a mountain of CASH, they could've easily created the most powerful console ever conceived ! but they think people are stupid, they don't understand specs and hardware. funny thing that we have Sony which is in the brink of bankruptcy investing in cutting edge and expensive RAM, equip PS4 with better GPU , and devoted a 5 year project to develop PS4 with an eye to industry future trends and needs to future proof the console. also above all that we find Sony charging less for PS4 despite the very superior tech found inside it compared to X1 !
MS just don't care, they said it themselves.
phosphor112
Banned
This should explain what hUMA is.
The only thing I don't get is paging... GPU has to work with paging enabled, so shaders are context-binded to a thread ?
Hmm. Maybe neither PS4/X1 have hUMA then.
PS4 can snoop CPU cache,
I don't think there is any sort of that feature in the X1 at all. Not with what the documents say at least.
Major edit:
Look at this.
Look at the Onion/+ Bus.
It's bidirectional.
Onion = GPU > GPU cache > CPU Cache > Main RAM. Vice versa for other way around.
Onion+ = GPU > CPU Cache > Main RAM. Vice versa.
Once the pipeline hits either the CPU or GPU's cache, the system SHOULD be able to directly access the CPU/GPU without having to go all the way out to main RAM.
So, it SHOULD have coherency...right?
Yes, since both the CPU/GPU probe the CPU's caches via Onion, if you choose to use that bus, CPU and GPU are fully cache-coherent. And thanks to the volatile tag, the GPU can bypass its own caches, meaning that they are not compromised by CPU/GPU interaction. This seems to be different on the XB1 where the GPU cache, from my understanding of the leaked documents, must be flushed.
idrinkalone
Banned
Does this mean I can't have halo at 1080p 60 fps? If not i don't care.
You can ALWAYS have a halo at 1080p/60FPS, no matter the console.
SenjutsuSage
Banned
thunder_snail
Member
But doesn't this mean huma like computing is limited by the relatively narrow bandwidth of onion? (relatively narrow for gpu, not cpu) Do gpgpu computing typically require a lot of bandwidth?
ConteZero76
Member
IIRC Cerny said that you can explicitly flag as "coherent" or "not coherent" different memory zones (probably on a page level) so a clever paging don't engulf onion without a reason.
thunder_snail
Member
Well, with huma, I'd imagine the GPU utilizing its many cores and high bandwidth to work on huge data structures while the CPU takes over infrequent branch conditions (since gpus suck at that as far as I know). So I think the GPU would still benefit a lot from high bandwidth.
But since this is a gaming system and the GPU will probably spend most of its time on graphics, hopefully, onion may be enough. 20 GB/s is an overkill for jaguar cpus anyways.
I'll just have to wait and buy an amd kaveri laptop to play around with this tech. Gonna be very educating.
Manmademan
Member
This is late, but it's worth addressing I think. it's erroneous to say MS doesn't care and prefers to sit on a mountain of cash while giving gamers shit hardware and telling them to go F themselves if they don't like it.
MS is publically traded and answerable to shareholders. they already lost their shirts establishing xbox two Gens back. a second money losing console of that nature isn't happening. Xbone needs to have a clear path to profitability.
second, the xbox division doesn't really make it all that much money for MS. Other divisions are much more important, and at the moment they're struggling with windows 8, are fending off google docs as an office competitor, losing badly with windows phone and getting completely, utterly and totally owned with surface. they can't afford to throw money at the xbox just to kill Sony, there's too much else at stake for them.
amstradcpc
Member
Yeah,until someone says if really there is a real unified memory address instead of a virtual one we will not know if its hUMA or not.I bet is still virtual like Llano or Cerny would have said the contrary.
thunder_snail
Member
Are you sure kaveri won't have an option for gddr5? I'm sure I've read somewhere they are gonna release an apu based on ps4's design. So they already have the memory controller worked out. Gddr5 dimms may not exist yet but a lot of ultrabooks these days have memory chips on the motherboard itself. Also if they go for an apu with say, 6 cu's, may be ddr3 will be enough.
Panajev2001a
GAF's Pleasant Genius
In a console environment there often are additional paths which in a way break nice and easy abstractions but provide additional performance headroom. This could be a way to explain both the use of onion, onion+, and garlic as you were also explaining in your post.
From leaks and interviews, it is apparent that, according to the memory region you store data in, you will access data through garlic or onion and if you want to skip GPU data caches you can access data through onion+.
It is possible to think that for hUMA to work, for addresses to be shared between CPU and GPU, AMD needed both a new GPU core compared to their previous APU's as well as other software customizations including OS support which would take much less time to enable in a custom BSD based OS than on Windows and OS X.
Also, it is conceivable that hUMA works only when CPU and GPU access data mapped on the onion memory region and not when accessed over onion+ or garlic. It would make sense and it would not be too hard to manage either IMHO for developers accustomed to crazier setups
.According to Cerny's Digital Foundry/EG interview and what he was saying about the HSA software stack, it is possible that it might be a feature that will get in the SDK post launch.
The GPU uses 4 memory controllers, each with two 32bit wide channels. [1] Jaguar most likely, like Kabini [2], uses a single 64bit wide memory controller. That must limit the bandwidth on Onion.
[1] http://www.amd.com/us/Documents/GCN_Architecture_whitepaper.pdf
[2] http://www.anandtech.com/show/6976/...wering-xbox-one-playstation-4-kabini-temash/4
/edit: Quote from the second article
So much for the "based on Kabini implies no hUMA" argument.
Panajev2001a
GAF's Pleasant Genius
It would still be virtual, CPU an GPU would share addresses and the same virtual to physical memory pages mapping. GCN can already use virtual memory and access data outside its directly accessible physical memory through paging, what seems to be required for hUMA is an extension (share the same virtual address space and pointers to it as the CPU) of that not a complete revolution I think.
sleeping_dragon
Banned
This thread has no huma.
By the way, the upcoming Kaveri (which officially supports hUMA) seems to be AMD's first APU to use GDDR5. [1] It's not far-fetched to conclude that the PS4s memory architecture is based on that.
[1] http://gamingio.com/2013/04/amd-to-...mory-architecture-for-cpuapus-as-cost-lowers/
[1] http://gamingio.com/2013/04/amd-to-...mory-architecture-for-cpuapus-as-cost-lowers/
Panajev2001a
GAF's Pleasant Genius
Good point.
Kind of puts everything in place when it comes to what AMD said and then had to get from under at Gamescom.
Very interesting!
Thank you for the link.
dragonelite
Member
Isn't the X1 apu presentation today?
Yes it is. 9:30AM PDT afaik.
googleplex
Member
... And the performance gap gets bigger.
Want to elaborate?
dragonelite
Member
So i can expect some slides later tonight or so here in euroland
robertsan21
Member
this is one confusing thread
Is hUMA used for PS4? or are the jury still out on that?
Will Xbox one have it?
Is hUMA used for PS4? or are the jury still out on that?
Will Xbox one have it?
What event is it at?
this is one confusing thread
Is hUMA used for PS4? or are the jury still out on that?
Will Xbox one have it?
We're probably not gonna get a definitive answer one way or another until these systems launch and the NDAs lift. It seems likely that PS4 supports hUMA in some form at least, since it looks like it's got the parts in place to support it. Jury's out on Xbone, though personally I wouldn't be surprised if it had some sort of solution of its own... But that's a tale straight from my ass, so don't put any stock into that.
Curious about this Xbone APU presentation. Wonder if we'll get any interesting insights out of that.
Seems likely based on the modification that Cerny talked about back in April.this is one confusing thread
Is hUMA used for PS4? or are the jury still out on that?
Doesn't seem like it. The OP gives a good rundown.
Thanks, curious what we will get from this. Hopefully something positively surprising.
I checked this table :
http://en.wikipedia.org/wiki/Heterogenous_System_Architecture#AMD_HSA_Implementation
and if I'm not misunderstanding something most of the listed 2013/Kaveri/2014 HSA features are indeed in Xbox One's APU. (Can be validated by a mod if wanted)
- passing pointers between CPU/GPU
- GPU uses pageable system memory via CPU pointers (well that's basically an implication from the above point)
- context switch
- pre-emption (which is basically context switching)
It really seems that only the eSRAM prevents the box from being "hUMA" by AMD's definition:
- Fully coherent memory between CPU & GPU
But MS might have an own solution for this:
from B3D (http://forum.beyond3d.com/showpost.php?p=1777116&postcount=5697)
I guess that's what they will talk about in the hot chips session.
http://en.wikipedia.org/wiki/Heterogenous_System_Architecture#AMD_HSA_Implementation
and if I'm not misunderstanding something most of the listed 2013/Kaveri/2014 HSA features are indeed in Xbox One's APU. (Can be validated by a mod if wanted)
- passing pointers between CPU/GPU
- GPU uses pageable system memory via CPU pointers (well that's basically an implication from the above point)
- context switch
- pre-emption (which is basically context switching)
It really seems that only the eSRAM prevents the box from being "hUMA" by AMD's definition:
- Fully coherent memory between CPU & GPU
But MS might have an own solution for this:
from B3D (http://forum.beyond3d.com/showpost.php?p=1777116&postcount=5697)
I guess that's what they will talk about in the hot chips session.
He goes on immediately after to mention nested page tables, which is what this is probably about. It sounds like, to me, that they have solution where you can pass a pointer from GPU/CPU at an API level using a page walker, but that's going to come with some overhead, and if the cache needs to be flushed every time, it's going to cost a lot. They may have a cache level solution that keeps it from being flushed.
The entire OS seems to sit in its own VM, which is interesting, and which probably means that a huge chunk of memory is going to be dedicated to the OS.
Dedicated GPU (in contrast to iGPU: integrated GPU).
Hexa
Member
Wait what? I thought by definition an APU can't have a dGPU.
That's why I was wondering - but I guess the person in question just did verification tests on APUs/dGPUs and PS4/X1 APUs. I first understood as if these consoles have an APU + dGPU. That was pretty much unbelievable.
Afaik, Richland APU's iGPUs can be used for Crossfire with a dGPU.
edit:nevermind - misreading.
That guy worked with certainty on more than just one project.
Could you provide your evidence for the context switch / pre-emption I have yet to read or even hear anything that suggests the XBONE has it.
For the first two points they are a standard feature of GCN, so it would be surprising if the XBONE didn't have them.
I don't think that compares. Cerny said that the quite specific use case of raytracing for audio could be done by GPGPU. I don't see how the audio processor in the XB1 could do raytracing since this task depends on the representation of the scene's geometry and, thus, needs fast access to main memory.
In addition, the XB1's audio processor explicitly has "pathways" to integrate calculations performed on the CPU indicating that it won't work well as a general purpose processor. I guess it's there to perform programmable tasks on audio streams. In this respect, we don't really know what the PS4's audio chip can do since all we have are Cerny's two sentences on that issue.
Brad Grenz
Member
It should be noted that the audio processing that Cerny has talked about doing on the GPU is not something you can do on the Xbox One's audio processor. If an Xbox One wanted to do the same audio ray casting it would have to use the GPU, too.
EDIT: Beaten
EDIT: Beaten
The only difference I can spot is the lack of finegrained GPU cache control in the XB1 and ESRAM/DME. But that was all known before the HotChips presentation from the leaked documents so I don't think that we have gained that much more information.
If I remember correctly, a 2x2 setup is the standard one in GCN while the added ACEs in the PS4 are among the explicit modifications.