PS3 GPU Analysis

[one]

One, eDram is not something you can just slap in there, you have to design an architecture around it! If NVIDIA next gen GPU use edram from the start then there are possibilities that we'll see an edram based GPU, otherwise I wouldn't bet on it

Yeah, it's moot at this time without the info about how they started the design But still the info that they use bulk CMOS only for the GPU sounds odd without eDRAM.

 

[sir doris]

kinda OT, but just a quick question: If eDRAM in the PS2 at 150MHz had 40+GB/s bandwidth how much bandwidth would eDRAM on a 500+MHz chip offer? Would just scale in proportion with the chip speed or might it run slower (but still fast enough)?

 

[Panajev2001a]

One, eDram is not something you can just slap in there, you have to design an architecture around it! If NVIDIA next gen GPU use edram from the start then there are possibilities that we'll see an edram based GPU, otherwise I wouldn't bet on it

Yeah, it's moot at this time without the info about how they started the design But still the info that they use bulk CMOS only for the GPU sounds odd without eDRAM.

Well, this is right, but again if nVIDIA did not plan to use SOI... well we will see if they have time changing from bulk-CMOS with low-k materials to SOI. It is not trivial.

 

[Fafalada]

They aren't much good for anything else ? They are stream processors after all.

I am pretty sure the question was referring to "how does one know SPEs will be the only thing available to do vertex processing".
Frankly, we don't know, and I've been hearing mixed messages lately as far as the rumours of any interest go.

In a shared memory scenario, the memory for this would ideally just be shared between CPU and GPU, so there's no need to send it over a bus.

There's damn good reasons TO send it over a bus - storing transforms in main mem not only wastes already precious memory, it also sends all that data over the bus TWICE.
And the potential reuse of transformed data is next to non-existant compared to the bus and memory wastage.

 

[j^aws]

Saying there is no eDRAM is like saying there is no sPOON.

Vysez,

I hope you mean that Xenon EDRAM, 'E' stands for 'E'nhanced' and not 'E'DO'...whatever Enhanced means...

 

[cthellis42]

No, not only is it EDO, but they're using a Cyrix processor!

 

[j^aws]

No, not only is it EDO, but they're using a Cyrix processor!

That's an 'overclocked Cyrix' too...

 

[Gubbi]

They aren't much good for anything else ? They are stream processors after all.

I am pretty sure the question was referring to "how does one know SPEs will be the only thing available to do vertex processing".
Frankly, we don't know, and I've been hearing mixed messages lately as far as the rumours of any interest go.

I'm aware of that (hence the question mark after the statement). It's just that vertex processing is one workload where the SPUs might do very well.

Cheers
Gubbi

 

[Farid]

I hope you mean that Xenon EDRAM, 'E' stands for 'E'nhanced' and not 'E'DO'...whatever Enhanced means...

Now, you're being silly Jaws, EDO would be too expensive for a $299 console, duh!


(Enhanced Dynamic Random Access Memory)--a form of DRAM that boosts performance by placing a small complement of static RAM (SRAM) in each DRAM chip and using the SRAM as a cache. Also known as cached DRAM, or CDRAM.

The Enhanced DRAMs are similar to standard DRAMs, but they send the output of the sense amps to a true RAM read cache (a second port). The SRAM cache and DRAM operate independently to enhance speed.

(Enhanced Dynamic Random Access Memory) A form of DRAM that boosts performance by placing a small complement of static RAM (SRAM) in each DRAM chip and using the SRAM as a cache. Also known as cached DRAM, or CDRAM.

 

[pahcman]

i hope ps3 have next gen edram. one of very few cool thing with ps2 graphics is fast edram. i see dx9 effects demo and i think ps2 still look nice.

funny if Xenon have edram and ps3 dont. 360 turn around is complete.

Xenon = ps2
ps3 = Xbox

 

[marconelly!]

Simply put, 25gb/s is not enough for a next-gen GPU. It won't be able to handle 720p 4x-8x FSAA, HDR rendering/blending/filtering, and tons of highly detailed textures in huge worlds and lots of render-to-texture ops.

OT, but isn't Xbox 2 also using unified memory architectrure? What is it's rumored bandwidth?

 

[j^aws]

Simply put, 25gb/s is not enough for a next-gen GPU. It won't be able to handle 720p 4x-8x FSAA, HDR rendering/blending/filtering, and tons of highly detailed textures in huge worlds and lots of render-to-texture ops.

OT, but isn't Xbox 2 also using unified memory architectrure? What is it's rumored bandwidth?

From the leak, ~ 22.4+ GB/s aggregate, shared CPU+GPU.

http://www.beyond3d.com/forum/viewtopic.php?t=13470

 

[nAo]

From the leak, ~ 22.4+ GB/s aggregate, shared CPU+GPU.

You should factor in GPU onchip dram bandwith also..

 

[j^aws]

I hope you mean that Xenon EDRAM, 'E' stands for 'E'nhanced' and not 'E'DO'...whatever Enhanced means...

Now, you're being silly Jaws, EDO would be too expensive for a $299 console, duh!


(Enhanced Dynamic Random Access Memory)--a form of DRAM that boosts performance by placing a small complement of static RAM (SRAM) in each DRAM chip and using the SRAM as a cache. Also known as cached DRAM, or CDRAM.

The Enhanced DRAMs are similar to standard DRAMs, but they send the output of the sense amps to a true RAM read cache (a second port). The SRAM cache and DRAM operate independently to enhance speed.

(Enhanced Dynamic Random Access Memory) A form of DRAM that boosts performance by placing a small complement of static RAM (SRAM) in each DRAM chip and using the SRAM as a cache. Also known as cached DRAM, or CDRAM.

Thanks...calling it CDRAM definitely makes more sense than EDRAM.

I had another look at the leak and it does mention EDRAM but the 'E' is for Embedded...

...
Xenon is designed for high-definition output. Included directly on the GPU die is 10+ MB of fast embedded dynamic RAM (EDRAM). A 720p frame buffer fits very nicely here. Larger frame buffers are also possible because of hardware-accelerated partitioning and predicated rendering that has little cost other than additional vertex processing. Along with the extremely fast EDRAM, the GPU also includes hardware instructions for alpha blending, z-test, and antialiasing.
...

http://www.beyond3d.com/forum/viewtopic.php?t=13470

 

[j^aws]

From the leak, ~ 22.4+ GB/s aggregate, shared CPU+GPU.

You should factor in GPU onchip dram bandwith also..

What is that bandwidth?

I think Marconelly meant the UMA/ system RAM bandwidth but I maybe mistaken.

 

[j^aws]

kinda OT, but just a quick question: If eDRAM in the PS2 at 150MHz had 40+GB/s bandwidth how much bandwidth would eDRAM on a 500+MHz chip offer? Would just scale in proportion with the chip speed or might it run slower (but still fast enough)?

The eDRAM bandwidth can be increased by clocking higher, increasing bus width or increasing no. of ports. The eDRAM also doesn't have to clock the same as the rest of the system.

Actually, I'm not sure how the 48 GB/s GS bandwidth is worked out, if anyone could elaborate...is the GS eDRAM bus 4096 bit and what clock, 150MHz?

 

[nAo]

anyone could elaborate...is the GS eDRAM bus 4096 bit and what clock, 150MHz?

150 Mhz, 2560 bits -> 1024 (color + z reads) + 1024 (color + z writes) + 512 (texture)

 

[London Geezer]

kinda OT, but just a quick question: If eDRAM in the PS2 at 150MHz had 40+GB/s bandwidth how much bandwidth would eDRAM on a 500+MHz chip offer? Would just scale in proportion with the chip speed or might it run slower (but still fast enough)?

The eDRAM bandwidth can be increased by clocking higher, increasing bus width or increasing no. of ports. The eDRAM also doesn't have to clock the same as the rest of the system.

Actually, I'm not sure how the 48 GB/s GS bandwidth is worked out, if anyone could elaborate...is the GS eDRAM bus 4096 bit and what clock, 150MHz?

Gee that's larger than the whole PS2

 

[j^aws]

anyone could elaborate...is the GS eDRAM bus 4096 bit and what clock, 150MHz?

150 Mhz, 2560 bits -> 1024 (color + z reads) + 1024 (color + z writes) + 512 (texture)

Ah..thanks...I don't know where I got 4096 bits bus from.

 

[marconelly!]

think Marconelly meant the UMA/ system RAM bandwidth but I maybe mistaken.

Yeah, that's what I thought. I always assumed PS3 would use some kind of EDRAM too, so I couldn't understand why DemoCoder seems so staunchly opposed to that, when Xbox 2 already seems to have simillar memory configuration.