Wrong ..
VGleaks: Orbis Unveiled! [Updated]
- Thread starter ekim
- Start date
Gcn2 needs to be already taped out though, and it's a 2014 gpu.
Can Crusher
Banned
No, it's a 2013 gpu that was supposed to come out first half of 2013 and got delayed for the second half of 2013 and resulted in a OEM rebranding of several AMD cards.
It's not a 2014 gpu at all.
No delay and re-branding wasnt due to it.
Can Crusher
Banned
Alright, I was just guessing since the re branding came out of the left field (to me). And the delay did happen so, I was adding 2+2
Death Dealer
Member
I assumed that was a typo. You wouldn't have alluded to an early predecessor of Xenos without knowing it wouldn't have evolved from R300.
Death Dealer
Member
Yeah I am surprised too. I really thought it would be beyond southern islands. When rumors dropped of 8GB for Xbox, I knew Sony knew they would be dead meat with only 2GB and would go for 4GB. I think what surprised a lot of people was the DDR3. Someone mentioned it last summer, but it seemed really implausible at the time because of bandwidth.
James Sawyer Ford
Banned
I thought the CPUs of the systems were supposed to be fairly comparable?
But aegies is saying that Durango's CPU is substantially different
Is it possible that Durango has a modified CPU that is much beefier than Orbis and accounts for some added GPGPU performance that is currently allocated to the 1.84 TF number for Orbis?
In other words, Durango has 400 GFLOPs more powerful CPU, but that's offset by Orbis' 400 GFLOPs allocated to GPGPU functions...ultimately resulting in systems that are comparable in power?
This is what some people are currently speculating on B3D....is it realistic?
But aegies is saying that Durango's CPU is substantially different
Is it possible that Durango has a modified CPU that is much beefier than Orbis and accounts for some added GPGPU performance that is currently allocated to the 1.84 TF number for Orbis?
In other words, Durango has 400 GFLOPs more powerful CPU, but that's offset by Orbis' 400 GFLOPs allocated to GPGPU functions...ultimately resulting in systems that are comparable in power?
This is what some people are currently speculating on B3D....is it realistic?
Death Dealer
Member
I think it's even more HSA adapted. It might not have been as big of deal in a closed console with APU. Just a hunch. But as other have said, unless it was going to be a huge leap, I think they wanted to stay conservative with something they knew wouldn't give them potential problems with a delay. I guess that won't happen but they had to plan just in case.
AltogetherAndrews
Banned
I think it's a pretty big part of the console's power, actually, given that exclusives are a direct part of the console. Hardware is important, but I'd argue that it's even more important to the console that there is a stable of dedicated developers that know how to work well within the limitations of the system (or as you say, play to the strengths while hiding the weaknesses). This generation it really made the difference for the PS3, hopefully MS can compete next generation.
A CPU with gpgpu capabilities? Would it not make more sense to have a gpu with that ability? I guess we had cell this gen, but that was a 400m investment on Sony and partners part.
When I was doing a rough guestimate benchmark for jaguar 8core based on expected performance levels, I found that there were xeons very similar to jaguar 8core level, which is - performance wise - similar to A10 hardware that Sony went with.
The extra Xeon was probably dsp related or as Aegies said, trying to partially mimic audio processing, etc...
gundamkyoukai
Member
I don't think it realistic for Durango to have a 400 GFLOPs CPU based on Jaguar .
That is crazy talk but you never know .
Neander
Member
Hmmm... I wonder if that's the reason why vgleaks specifically calls the Orbis CPU Jaguar where as they make no such distinction for Durango. Also (not sure if I'm reading it wrong) for Orbis they say the two CPU clusters have a "shared 2MB L2 cache" where as for Durango they say it has a "total of 4 MB of L2 cache".
Death Dealer
Member
I can see some kind of special vector helpers like Xenon. I don't have any clue how many GFlops they could be. I do know the 8 core Jaguar itself is supposed to be a little over 100. So that seems really weird to have a CPU co-processor with 4x the power of the main one. Very Cell like. Some early insiders did say Durango was weird.
JohnnySasaki86
Member
400glops for a CPU?? A Jaguar based CPU? That's even crazier. I doubt that Durango CPU would 4x as powerful as Orbis's. That be nuts.
They'll both have 4mb L2 cache. That's basically a two quad core jaguars put together. I think you are looking too deep if you are taking apart sentences that essentially mean the same thing.
JohnnySasaki86
Member
No I think Orbis has a total of 4mb L2 cache as well. They say there is two clusters. Then the next line below they say that EACH cluster contains 4 cores and a shared 2mb L2 cache.
Once again just like the extra ALU on the each of the 4CU's I'm pretty sure that's the correct way to read it.
For that Durango also need those Orbis 4 CU to match in power. From rumors right now both of them are the same and both 100 Gflop.
Neander
Member
It'd need 40 Jaguar cores total to have an extra 400 GFlops over the 8 core model. A modified CPU that has GPGPU performance... you think MS is putting a couple Cells in the Durango?
It can. If they have a 256 wide AVX2 unit per core, the Durango CPU would sport 410 GFLOPs. In case you have forgotten MS did the same thing with Xenon CPU on the 360 where they increased the VMX128 count.
iamshadowlark
Banned
I really don't get the fascination with GCN2.
What scently suggests wouldn't change it much, I suspect.
Would they really get quadruple the theoretical FLOPs doing that? And if they did, would they see much of a real-world performance boost? It'd be a worthwhile upgrade to the Jags, as the double-pumped 128-bit AVX2 unit seems odd for a gaming console, but I don't think it'd match having an extra four CUs on the GPU.
les papillons sexuels
Banned
new instruction sets / architecture.
like the memexport function and tessellator in the 360's xenos that eventually became the basis for dx10/11 features.
It will get the performance and it would actually be better to have that amount of flops on the cpu as compute jobs, even with the improvements to run them on gpus, are still predominantly a cpu side duty.
Mind you, I am not saying that they are adding it to their cpu but as for whether it is possible for them to do it? yes it is and it will more efficient at it. That is why you hear the suggestion that the 4 CUs that seems to have been reserved for compute jobs would perform better if the ps4 cpu can have access to them directly.
oversitting
Banned
http://www.simmtester.com/page/news/images/image33.gif
ddr1 vs rambus. 10% difference. with ddr2 the difference will shrink to next to nothing.
No shit? jaguar cores are 3mm^2 at 28nm which means they are like 24m transistors each core. the 4mb cache is 24m.
Please find a relevant benchmark of XDR for performance reference. If you can't please print out your post and eat it. GDDR5 is a DDR3 derivative. Google GDDR5 vs XDR.
XDR was about the same speed as GDDR5. But memory speed on a cpu has no really any performance gain since ddr2.
iamshadowlark
Banned
But it doesn't seem to be any of that. Everybody I see talk about it speak in terms of refinements, I doubt its going to be anywhere near the scale you speak of.
JohnnySasaki86
Member
360's gpu could do tesselation??? Since when?
alexandros
Banned
Quick question for aegies: What kind of PC hardware (CPU/GPU) would be needed in order to run next-gen games on PC at a similar level of quality/resolution? No need to be specific, just a ballpark figure would be fine. Also fine if you don't have that information or you don't want to reveal it.
Brad Grenz
Member
ATI GPUs have had rudimentary forms of tessellation hardware going back to the "TruForm" stuff on the old Radeon 7500 series.
Code:
[B]Display ScanOut Engine (DCE)[/B]With Sony's love for 3D & knowing 3D is going to be a big part of PS4, I'm wondering could this chip be for more than just up-scaling , like could it be used to convert all games into 3D even if it's a 30 FPS game by up converting the FPS or by splitting the images up in a way that the TV will see it as 3D.
Brad Grenz
Member
Wouldn't 256 bit AVX2 only double what a standard Jaguar can do? Aren't they already 128 bit AVX units? So double that and you'd get up to 205 GFlops.
They aren't in the normal graphics pipeline and you have to program them directly. But you can use them for what you like. Eg you could run physics for half the time (or on 2CUs), and use the rest of the time to do graphics related stuff like pre sorting or post processing. Not directly pushing polys, but impacting graohics anyway.
Not sure if the CUs are directly accessible (potentially something devs might want) or you just give them code and they divvy it up between them. Shouldn't matter too much - developers would arrange their code to take the correct amount of time to accomplish the same result.
Errr.....since the console was created. It has a Geometry Tessellation Unit though it is not as powerful nor as flexible as the ones found in AMD gpus. Several titles use it actually but not extensively eg. Viva Pinata, Alone in the Dark, and others.
Lord Error
Insane For Sony
The only benchmark relevant to this discussion would be how XDR performed in PS3, vs. DDR of the time. I'm guessing it's practically impossible that they didn't test various memory types on PS3 kit, and whatever advantage XDR gave them, big or small (remember, GPU needed to have indirect access to it as well as the CPU) was probably deemed worth the price hike.
Nope, it was an honest mistake. I know it derived from the R400 but for some reason I thought it was the R300 when making that post. These late nights are killing me. =P
I just want to comment on this - IMHO, everything that is shitty about the PS3 architecture you can pretty much blame on the RSX.
les papillons sexuels
Banned
the very beginning, same with unified shaders. All of amd's r600 cards are based off of the designs that AMD and MS collaborated for the 360. Memexport is also the foundation for compute shaders. Both tessellation and compute shaders were supposed to be in dx10, but were pushed back to dx11.
IIRC on paper the RSX had a higher theoretical output (fill rates, flops etc), was the beefer hardware but the xenos had a better instruction set and feature list.
and the ram...
Cool. That's what I thought.
Panajev2001a
GAF's Pleasant Genius
Those extra CU's might also be used for rendering so they need to have the same feature sets, also mapping vectors and matrices to textures buys you quite a bit of compute performance, seen it first hand years ago with CUDA kernels.
It does make sense if SCE asked for the HW scheduler in those extra CU's and the branch predictor, amongst other bits, to be optimized more for compute tasks than graphics processing so to offer more help to the CPU.
I blame the RAM on the RSX. Seriously.
amstradcpc
Member
Googling, from Sotfopedia:
Jaguar now comes with most if not all of Bobcats abilities, but it also adds complete architectural computational support for SSE4.1, SSE4.2, AES, CLMUL, MOVBE, AVX, XSAVE, XSAVEOPT, FC16, and BMI instructions.
Since most SIMD instructions are 128-bit wide if not wider, AMD decided to double the width of its FPU pipelines from 64-bit to 128-bit.
SSE instructions required two passes on the Bobcat, but they now can be handled in one single pass, and this is no small feat.
Ironically for AMD, Jaguar comes with support for AVX and these would be best fit by 256-bit wide FPU pipelines,but this is not the main concern for a mobile-oriented architecture.
Jaguar now comes with most if not all of Bobcats abilities, but it also adds complete architectural computational support for SSE4.1, SSE4.2, AES, CLMUL, MOVBE, AVX, XSAVE, XSAVEOPT, FC16, and BMI instructions.
Since most SIMD instructions are 128-bit wide if not wider, AMD decided to double the width of its FPU pipelines from 64-bit to 128-bit.
SSE instructions required two passes on the Bobcat, but they now can be handled in one single pass, and this is no small feat.
Ironically for AMD, Jaguar comes with support for AVX and these would be best fit by 256-bit wide FPU pipelines,but this is not the main concern for a mobile-oriented architecture.
Death Dealer
Member
On paper, I think Xenos actually just edged out RSX in Flops. The main advantage was unified shaders. So the developers could decide how to allocate the pixel and vertex loads. PS3 developers were stuck with a predefined arrangement. So Xenos could shift most of its pipes to vertex loads and blow the doors off RSX max 8 vertex pipelines. Xbox 360 developers had a scheduler that would dynamically allocate the shader workload because there was no distinction between pixel and vertex units.
Well...IMHO this article hits on the main issues pretty well.
http://www.theinquirer.net/inquirer/news/1007286/ps3-hardware-slow-broken
Of course, workarounds were found, and I think they are well known to just about every PS3 developer worth their salt.