Support NeoGAF

[QuickSilverD]

I know Jaguar should have some advantages being a more modern design and having OoOE, but technically speaking, regarding the number of operations they could do in a real word scenario, which is better?

I'm not too technically minded so try to explain it as simple as you can, I know these two have very different architectures, but I wouldn't be asking this question if every time we talk about these consoles CPU everyone complains how weak their "notebook" cpus are, to the point were I am wondering this now.

Last gen, GPU tasks were being offloaded to the Cell to carry the weight of the (relatively) weak RSX, this gen it seems it is the other way around, with the GPUs now having to carry the CPUs (GPGPU)

 

[mcrommert]

Even with the CPU's being weaker in relation to the tech of the day...the xbox one and ps4 are around 14x-18x more powerful cpu's then their predecessors.

 

[omonimo]

Nope. Jaguar is simply more cpu than the cell. Cell can handle gpu task but it's not more powerful. From what I know.

 

[intheinbetween]

I know Jaguar should have some advantages being a more modern design and having OoOE, but technically speaking, regarding the number of operations they could do in a real word scenario, which is better?

I'm not too technically minded so try to explain it as simple as you can, I know these two have very different architectures, but I wouldn't be asking this question if every time we talk about these consoles CPU everyone complains how weak their "notebook" cpus are, to the point were I am wondering this now.

Last gen, GPU tasks were being offloaded to the Cell to carry the weight of the (relatively) weak RSX, this gen it seems it is the other way around, with the GPUs now having to carry the CPUs (GPGPU)

totally different, 8 cored 1,6ghz x86 CPU VS 1 PPE @ 3,2Ghz + 6-7 SPEs PowerPC CPU

Cell had its strenghts when used properly, which required hard work. Jaguar is a "standard" x86 cpu, so it will be much easier to work with

 

[Jonnax]

I remember that Ubisoft presentation about the dancers showed that the SPEs in the cell were very good for those tasks but isn't it rather that the SPEs were similar to a GPU?

 

[KojiKnight]

totally different, 8 cored 1,6ghz x86 CPU VS 1 PPE @ 3,2Ghz + 6-7 SPEs PowerPC CPU

Cell had its strenghts when used properly, which required hard work. Jaguar is a "standard" x86 cpu, so it will be much easier to work with

In a single threaded instance how do they compare though? Of course games are going to take more advantage of parallelism that has occured in the last decade of PC development, but for the sake of argument, how does a single Jaguar core in a single threaded application fare against the primary PPE of the Cell?

 

[Human_me]

I remember that Ubisoft presentation about the dancers showed that the SPEs in the cell were very good for those tasks but isn't it rather that the SPEs were similar to a GPU?

Heres that slide incase anyone was wanting it.

 

[LukasTaves]

For pretty much anything other than GPU workload the jaguar will destroy last gen CPUs. For the tasks where devs used Cell to compensate for RSX weakness there's no need to use the processor anymore.

There are some workloads in between like complex AI routines (visibility checks, IK etc) that cell would outperform jaguar cores, but those should be able to run on the gpu too.

 

[jelly]

What's a Jaguar compared to i3, i5s or is it chalk and cheese?

 

[Finaika]

I'm not too technically minded so try to explain it as simple as you can, I know these two have very different architectures, but I wouldn't be asking this question if every time we talk about these consoles CPU everyone complains how weak their "notebook" cpus are, to the point were I am wondering this now.

They're not notebook CPUs, they're tablet CPUs.

 

[Mihalis0013]

What's a Jaguar compared to i3, i5s or is it chalk and cheese?

If I'm not mistaken, Jaguar was meant to compete with Intel Atom line, not the core i series

http://m.neogaf.com/showthread.php?t=677101

 

[James1o1o]

They're not notebook CPUs, they're tablet CPUs.

Not even that.

Jaguar is an architecture, not a CPU.

There are tablet and notebook CPU's that use the jaguar architecture. These CPU's are known as Kabini and Temash.

You could have a single core 100mhz processor using Jaguar or an 8 core 5ghz processor using it.

 

[Rolf NB]

They're not notebook CPUs, they're tablet CPUs.

There are tablets with Haswell CPUs. What's your point?

 

[QuickSilverD]

Even with the CPU's being weaker in relation to the tech of the day...the xbox one and ps4 are around 14x-18x more powerful cpu's then their predecessors.

this doesn't sound right to me

Nope. Jaguar is simply more cpu than the cell. Cell can handle gpu task but it's not more powerful. From what I know.

so, how much more powerful the PS4 CPU is in relation to Cell, say in a CPU intensive task, how much better would Jaguar perform?

 

[danwarb]

No. The Jaguar is better suited for purpose, though weak next to a typical newish desktop CPU. Cell in PS4 would be dwarfed to insignificance by the fairly average GPU.

CPUs are I think a big limiting factor this time. Already very slow and probably why you get better than console performance with a 7850, similar to the PS4 GPU part, but typically with a far better CPU. I'd hoped physics would be huge this generation.

 

[R1CHO]

Even with the CPU's being weaker in relation to the tech of the day...the xbox one and ps4 are around 14x-18x more powerful cpu's then their predecessors.

14-18?

No way overall.

 

[Nirolak]

The Cell's SPUs were floating point calculators more akin to what Compute Shaders are on the PS4/XB1 GPUs. For calculations based on floating points, they're going to be more efficient than a Jaguar CPU.

However, for CPU loads that aren't based in floating point calculations, they're notably better.

The idea here is to put floating point workloads on GPU Compute Shaders so you're only focusing workloads that are ideal for CPUs on the CPU.

When Ubisoft showed that chart of CPU vs. GPU performance and included the Cell, it was a floating point based workload, which is why it ran better on an SPU than a Jaguar core, but vastly better on a GPU (which is a really heavily focused floating point calculator since graphics involve tons of floating point calculations).

 

[Henrar]

They're not notebook CPUs, they're tablet CPUs.

With that TDP? Hell no, those Jaguars in XBO/PS4 would fry in a tablet.

 

[Elios83]

I know Jaguar should have some advantages being a more modern design and having OoOE, but technically speaking, regarding the number of operations they could do in a real word scenario, which is better?

I'm not too technically minded so try to explain it as simple as you can, I know these two have very different architectures, but I wouldn't be asking this question if every time we talk about these consoles CPU everyone complains how weak their "notebook" cpus are, to the point were I am wondering this now.

Last gen, GPU tasks were being offloaded to the Cell to carry the weight of the (relatively) weak RSX, this gen it seems it is the other way around, with the GPUs now having to carry the CPUs (GPGPU)

They are very different architectures.
8 core Jaguar at 1.6GHz is much better than the single PPE in the Cell at executing general purpose code (like branching code), it's about 4 times more powerful and way more efficient (PPE didn't have out or order execution as you mention and also cache memory was far less flexible) which means that you can get better performance out of the same theoretical power.
For pure parallelized floating point calculations Cell is much better because it has its seven SPEs.
Fact is that the Cell has these SPEs because it was meant to be the centre of the whole system, they were needed to offload graphics calculations from the GPU and making them work in tandem.
This gen consoles have been conceived with the opposite philosophy, GPU is at the core with the bulk of the floating point capabilities of the system relying on its programmable shaders which, thanks to their GPGPU capabilities,could even help the CPU offloading calculations from it.
So in short, if you take Cell and 8 core Jaguar CPU alone then one is much better than the other at specific tasks and viceversa. But considering that these CPUs exist as part of a vision for a whole system it doesn't make sense to judge them in this way.

 

[QuickSilverD]

The Cell's SPUs were floating point calculators more akin to what Compute Shaders are on the PS4/XB1 GPUs. For calculations based on floating points, they're going to be more efficient than a Jaguar CPU.

However, for CPU loads that aren't based in floating point calculations, they're notably better.

The idea here is to put floating point workloads on GPU Compute Shaders so you're only focusing workloads that are ideal for CPUs on the CPU.

When Ubisoft showed that chart of CPU vs. GPU performance and included the Cell, it was a floating point based workload, which is why it ran better on an SPU than a Jaguar core, but vastly better on a GPU (which is a really heavily focused floating point calculator since graphics involve tons of floating point calculations).

Thanks.

Tangentially, could someone explain what are "Floating Point" workload and how they differ from I guess "normal" workload?

 

[LeleSocho]

Between CPUs trying to be GPUs of the last gen and GPUs trying to be CPUs this gen i really hope that the next batch of consoles come out perfectly balanced on both departments.

 

[LordOfChaos]

Nah. I mean, in some ways, but in ways related to programming a game the Jaguar cores will almost always come out ahead. The Cell wins out on raw floating point related gflops, but it has much bigger drawbacks for actually running a game, ie very little local memory per SPU, having to hand manage every local memory or else have a huge drop in performance, SPUs being outright horrible for branchy code, etc etc.

Jaguar demolishes it for what games actually do. And the little it was better at is now better handled by vastly more powerful GPUs that also demolish Cell at its specialized processing.

Oh, and besides all that the Cell, when taken proper advantage of, was usually just covering up for how shitty a GPU RSX was, very inferior to Xenos. A lot of the GPU load had to be shifted to the Cell to make things pretty.

 

[z0m3le]

This isn't some mysterious CPU, it's an AM1 Kabini Jaguar core with 8 cores instead of 4, you are basically looking at the AM1's Althlon 5150, a 1.6GHz Jaguar quad core, also an entry level APU.

http://www.anandtech.com/show/8067/...thlon-53505150-and-sempron-38502650-tested/10

Here is some benchmarks for a dedicated GPU and that 4 core processor. IIRC PS4 uses 6 CPU cores for games? so the improvement comes from less overhead (windows OS is going to have stuff in the background) and 2 extra cores. It might catch up to the C2750 Atom CPU there, but I don't think it's capable of much.

Cell was a bit lacking in terms of general processing as it only had 1 Core with 2 threads and 7 SPE satellites. In general computing, the 360 might of had a better performance with 3 of these cores, both systems clocked at 3.2GHz and I think each core was around 38GFLOPs giving you 115 GFLOPs for 360's CPU in general compute, 38 in PS3 + well over 100 more GFLOPs from the SPEs that can be utilized for specific tasks. This compares to Jaguar's ~102GFLOPs.

The real issue is that things don't always work out like they do on paper, for instance the intel atom c2750 is just shy of 77GFLOPs iirc but it still exceeds Jaguar in the above bench, despite a flops advantage. The benefit with these cores is the shorter pipeline and the OoOe, though out of order execution might not mean much when you are programming a game directly to be ran on the hardware in the box, you can just give all your processes an order in which to be processed, thus removing the need for such a thing.

 

[Inuhanyou]

Thanks.

Tangentially, could someone explain what are "Floating Point" workload and how they differ from I guess "normal" workload?

"Floating point" operations or FLOPS, are just the general amount of peak raw numbers a particular component can crunch at a time. It has nothing to do with real world performance metrics outside of a handy observation tool to keep scale.

As for your question OP, in terms of FLOPS, the CELL is much stronger than the Jaguar CPUs in the next gen consoles. But its not a 1:1 comparison, as the CELL only could reach its peak output on very limited scenarios in which it was coded for efficiently, as well as taking into account like you said, a portion of its power was also being utilized to make up for the weak rasterizer RSX GPU in the PS3.

The Jaguar CPU's are much less limited in one they can do and are much more 'multipurpose' and flexible at tasks. Thus their performance in real world scenarios will reflect that. Of course, they are still weak and unbalanced in comparison to the GPU's in the new gen consoles, even the XB1's GPU.

 

[mcrommert]

This isn't some mysterious CPU, it's an AM1 Kabini Jaguar core with 8 cores instead of 4, you are basically looking at the AM1's Althlon 5150, a 1.6GHz Jaguar quad core, also an entry level APU.

http://www.anandtech.com/show/8067/...thlon-53505150-and-sempron-38502650-tested/10

Here is some benchmarks for a dedicated GPU and that 4 core processor. IIRC PS4 uses 6 CPU cores for games? so the improvement comes from less overhead (windows OS is going to have stuff in the background) and 2 extra cores. It might catch up to the C2750 Atom CPU there, but I don't think it's capable of much.

Cell was a bit lacking in terms of general processing as it only had 1 Core with 2 threads and 7 SPE satellites. In general computing, the 360 might of had a better performance with 3 of these cores, both systems clocked at 3.2GHz and I think each core was around 38GFLOPs giving you 115 GFLOPs for 360's CPU in general compute, 38 in PS3 + well over 100 more GFLOPs from the SPEs that can be utilized for specific tasks. This compares to Jaguar's ~102GFLOPs.

The real issue is that things don't always work out like they do on paper, for instance the intel atom c2750 is just shy of 77GFLOPs iirc but it still exceeds Jaguar in the above bench, despite a flops advantage. The benefit with these cores is the shorter pipeline and the OoOe, though out of order execution might not mean much when you are programming a game directly to be ran on the hardware in the box, you can just give all your processes an order in which to be processed, thus removing the need for such a thing.

Windows OS has more in the background??? Xbox also is using 6 cores for games...2 are reserved for the Xbox OS

 

[Maninthemirror]

Nearly the same output, which means realistically I think Cell is equav to Jaguar processors

of course we are talking of the difference between RISC and CISC and x86 is MUCH easier to get performance out of.

If anything the CPU performance now is what Cell should have been able to output at the offset if it was easy to program

 

[shandy706]

Heres that slide incase anyone was wanting it.

Xbox One best dance simulator confirmed!

 

[Rolf NB]

The Cell's SPUs were floating point calculators more akin to what Compute Shaders are on the PS4/XB1 GPUs.

They are really not. They are fully fledged CPU cores boxed into a local memory model. They can branch and jump and resolve indirections willy nilly, because they are CPU cores with everything that implies. As such they can do a lot of CPU-y things GPUs still struggle with or simply can't do at all today, much less in 2006. Liike zlib compression, Xpath queries and stuff like that.

That they have significant SIMD computation resources is true, but it's an orthogonal "issue". You wouldn't call Haswell a "floating point calculator akin to compute shaders" just because it has SIMD resources.

 

[Edvardelis]

You can't really say "this" is better then "that" without taking in to account difficulty to code for. It doesn't matter if one is 50% "stronger" than another if the bulk of developers can't write code well enough to show that advantage.

 

[lukeskymac]

Simple answer: no.

To those asking, while the standard Jaguars were used in netbooks, the ones found in XB1/PS4 have double the cores. They are roughly on par with an i3, AFAIK.

 

[AmFreak]

The Cell's SPUs were floating point calculators more akin to what Compute Shaders are on the PS4/XB1 GPUs. For calculations based on floating points, they're going to be more efficient than a Jaguar CPU.

However, for CPU loads that aren't based in floating point calculations, they're notably better.

The idea here is to put floating point workloads on GPU Compute Shaders so you're only focusing workloads that are ideal for CPUs on the CPU.

The SPU's can do everything a "normal" cpu can do.
They are as fast on integer as they are on fp. They are just extremely simple or "dumb" in comparison to a modern cpu core (In-Order, no branch prediction, etc...).
That's what's making them so fast and so hard to code for.
You can (partly) optimize against these "flaws", but this takes time and is worsen by the fact that your normal pc codebase doesn't care about them.

 

[Panajev2001a]

This is a complex question IMHO.

Pro's: I think that an improved CELL BE 2 processor with say 2 PPU's and 16 SPE's running at 4 GHz would have been possible and could have been a great console CPU for both general purpose and vector oriented tasks (especially with an OoOE core in the PPE's). I still think that physics, A.I., and animation should be run on the same core and not split between two processors without a fast, not too tiny, and low latency memory pool/connection to share data with.

Cons: it would have cost more money in R&D, it would have eaten a bigger slice of the overall silicon budget that would probably be taken out from either the GPU or the RAM ones or both, it might have made it tough or impossible to launch with a consolidated SoC with CPU and GPU on the same die, and most importantly it might not have given Sony the edge in ease of programming which helped gained such a strong initial developers support by big third parties and indies alike.

They might have produced a faster/stronger architecture, but it would have taken longer to learn how to tame it efficiently/harder to code for (no unified memory space, lower quality tools, not as familiar as regular PC's to code for, etc...).

 

[Marlenus]

No.

The Cell could be considered an early APU. The PPE is the CPU and the SPEs act as the GPU. In the PS3 it also had a separate GPU as well to help improve graphics performance.

If you look at the Ubisoft benchmark above you see that the 360 CPU is a lot slower than the Jaguar CPUs and that is essentially 3 PPEs in one unit. Given that you can say that the Cell PPE is ~10% of the Jaguar CPU and the Cell SPEs are ~6% of the GPU (The PS4 GPU did ~1600 dancers IIRC).

 

[twobear]

As far as I can see, even if the Cell is marginally better than the Jaguar CPU at some tasks, it's marginally better at tasks it was specifically designed to excel at.

What's we're seeing is that the 'tablet' CPUs in PS4 and Xbone are very nearly as good at the task that Cell was designed specifically as the Cell, despite the fact that they are low end, general-purpose CPUs.

It's a bit like finding out that your average family car from 2014 has almost the same 0-60 time as a supercar from 2008. Yes, the supercar is still marginally faster, but you still can't fit the family dog in the trunk.

 

[Bonfires Down]

Even with the CPU's being weaker in relation to the tech of the day...the xbox one and ps4 are around 14x-18x more powerful cpu's then their predecessors.

How in the world did you arrive at such a number?

 

[James1o1o]

What's we're seeing is that the 'tablet' CPUs in PS4 and Xbone are very nearly as good at the task that Cell was designed specifically as the Cell, despite the fact that they are low end, general-purpose CPUs.

And to put it into perspective, developers at the end of PS3 life were still finding new ways of utilizing that CPU power, the thing that held the PS3 back was the GPU and memory.

 

[Dictator93]

Even with the CPU's being weaker in relation to the tech of the day...the xbox one and ps4 are around 14x-18x more powerful cpu's then their predecessors.

Can we get a tales from my ass for this post?

Like a large butt spewing shit?

 

[lukeskymac]

And to put it into perspective, developers at the end of PS3 life were still finding new ways of utilizing that CPU power, the thing that held the PS3 back was the GPU and memory.

Yes, well, most games as we know them are bound to be GPU limited, unless you throw a lot of advanced AI I guess.

 

[Nirolak]

They are really not. They are fully fledged CPU cores boxed into a local memory model. They can branch and jump and resolve indirections willy nilly, because they are CPU cores with everything that implies. As such they can do a lot of CPU-y things GPUs still struggle with or simply can't do at all today, much less in 2006. Liike zlib compression, Xpath queries and stuff like that.

That they have significant SIMD computation resources is true, but it's an orthogonal "issue". You wouldn't call Haswell a "floating point calculator akin to compute shaders" just because it has SIMD resources.

The SPU's can do everything a "normal" cpu can do.
They are as fast on integer as they are on fp. They are just extremely simple or "dumb" in comparison to a modern cpu core (In-Order, no branch prediction, etc...).
That's what's making them so fast and so hard to code for.
You can (partly) optimize against these "flaws", but this takes time and is worsen by the fact that your normal pc codebase doesn't care about them.

Sorry, by "more akin" I meant that the way you interact with them is more in the realm of a specialized co-processor that can handle floating point very well as opposed to them not being a CPU at all, instead of compute shaders and SPUs being literally the same.

A lot of engines went about converting the types of jobs you did run on SPUs due to their floating point strength and PS3 GPU deficiencies over to compute shaders, like what the Ubisoft slide is about.

But yes you both do make very good points in that they absolutely are CPUs. The Jaguar cores though just aren't built with doing this kind of thing in mind so you really want to focus on GPU offloading tasks like this unless you just have unutilized CPU cores otherwise.

 

[ProfessorMoran]

As far as I can see, even if the Cell is marginally better than the Jaguar CPU at some tasks, it's marginally better at tasks it was specifically designed to excel at.

What's we're seeing is that the 'tablet' CPUs in PS4 and Xbone are very nearly as good at the task that Cell was designed specifically as the Cell, despite the fact that they are low end, general-purpose CPUs.

It's a bit like finding out that your average family car from 2014 has almost the same 0-60 time as a supercar from 2008. Yes, the supercar is still marginally faster, but you still can't fit the family dog in the trunk.

Why would you put a dog in the trunk?

 

[mcrommert]

How in the world did you arrive at such a number?

Didn't realize this was controversial.

It was told to me by someone who is doing dev work on xbox one...ignore if incorrect

 

[Durante]

It is, if you are doing SIMD-style tasks and have an army of John Carmacks at your disposal.

 

[The Llama]

Sony originally wanted the Cell to be the PS3's GPU as well as the CPU, so it's very good at some things and not that good at others. As a pure CPU, the Jaguar is absolutely superior.

 

[Insane Metal]

It is, if you are doing SIMD-style tasks and have an army of John Carmacks at your disposal.

LoL, perfect.

The Cell is incredibly powerful, considering it's an 8 years old CPU. Too bad it's overly complicated.

The thing is, the new generation of consoles has a much simpler architecture to make the developers' lives easier. Also, the choice was made based on power and heat dissipation requirements.

 

[family_guy]

This is a complex question IMHO.

Pro's: I think that an improved CELL BE 2 processor with say 2 PPU's and 16 SPE's running at 4 GHz would have been possible and could have been a great console CPU for both general purpose and vector oriented tasks (especially with an OoOE core in the PPE's). I still think that physics, A.I., and animation should be run on the same core and not split between two processors without a fast, not too tiny, and low latency memory pool/connection to share data with.

Cons: it would have cost more money in R&D, it would have eaten a bigger slice of the overall silicon budget that would probably be taken out from either the GPU or the RAM ones or both, it might have made it tough or impossible to launch with a consolidated SoC with CPU and GPU on the same die, and most importantly it might not have given Sony the edge in ease of programming which helped gained such a strong initial developers support by big third parties and indies alike.

They might have produced a faster/stronger architecture, but it would have taken longer to learn how to tame it efficiently/harder to code for (no unified memory space, lower quality tools, not as familiar as regular PC's to code for, etc...).

But on the bright side maybe we would have had backwards compatibility.

 

[stan423321]

A programmer's analogy.

Cell is like thirty or so drunk teenagers and two average teachers trying to herd them. Jaguar is like eight German Army generals.

The teenagers are so drunk that they have to cluster together to walk, but when it comes to breaking ice with bare fists, they will probably be a better fit. The thing is, GPU is like a swarm of intoxicated crows that's even better at it (except they use beaks, but that's just a little implementation detail).

 

[BLACK AMERICAN PSYCHO]

It is, if you are doing SIMD-style tasks and have an army of John Carmacks at your disposal.

LOL, perfect.

 

[lefantome]

Power is relative.
The cell wasn't suited for gaming tasks, the Jaguar CPU is doing much better in that field.

Somewhere I've read that the next gen cpus are between 3-5x the Xenon but the gpu power difference is much higher.

 

[LordOfChaos]

Heres that slide incase anyone was wanting it.

That...Is confusing. XBO processor is so much faster than the PS4 in that? Even assuming the PS4 is clocked at 1.6GHz (some benchmarks imply it's faster) and the XBO is 1.75, that's 9% faster, which would put it at about 106FPS in theory. So what's the extra coming from?

And Cell at 3x faster than Xenon? It's faster in some regards in theory, but even then 3x is at the upper bounds, close to double is what you would usually expect with absolutely perfect scaling to each SPU.

 

[LordOfChaos]

A programmer's analogy.

Cell is like thirty or so drunk teenagers and two average teachers trying to herd them. Jaguar is like eight German Army generals.

The teenagers are so drunk that they have to cluster together to walk, but when it comes to breaking ice with bare fists, they will probably be a better fit. The thing is, GPU is like a swarm of intoxicated crows that's even better at it (except they use beaks, but that's just a little implementation detail).

That...Is surprisingly pertinent. Though maybe 8 teachers rather than 8 german army generals.