RSX = Stream Processor!?!

[aaronspink]

It's not just a marketing name:

oh, yes it is.

Since it's refresh free

Its not refresh free, they are just relaxed the timing to the point where they can provide hidden refresh. There is controller logic sitting with the embedded dram array macro which performs a refresh of the dynamic ram cells.

even though eDRAM stands for embedded dynamic randon access memory. eDRAM has traditionally been a term used for embedded memory that requires refresh.

No, eDRAM has traditionally been a term used for embedded DRAM. End STOP, FULL STOP.

It's much more accurate to say eSRAM, but that's an unfamiliar term for people to relate to. How about 1T-SRAM?

It would be fundamentally less accurate to say eSRAM, since it IS NOT SRAM. HAS LITTLE IN COMMON with SRAM. IT IS DRAM.

Aaron Spink
speaking for myself inc.

 

[aaronspink]

Mosys 1t SRAM isn't just straight DRAM with a fancy name, it's tweaked DRAM.

No dram is just straight DRAM. All DRAM is specifically optimized for a combination of error resistance, refresh rate, cost, latency, power, and bandwidth. People have been doing small bank DRAMs well before Mosys existed. And certainly have been doing it with eDRAM for a long time.

Aaron Spink
speaking for myself inc.

 

[flick556]

I think subdivided surfaces will become widespread in both next gen consoles. It makes good since on rsx if this stuff about it’s link to cell is true and xenon has dedicated hardware for it. You definitely stand a chance to save tons of memory but you will still need as much bandwidth as possible. Working with good reduction algorithms and normal/displacement maps you can take a million poly model and reduce it to 10k or so polys that when subdivided and the normal maping applied will produce a near identical model. This is a huge step up from just working with low poly models and normal maps like doom and other new engines do. If a few spus can take that 10k model and crank out 100k or so worth of polys for rsx to render I would definitely consider that a stream processor. That type of subdivision rendering can produce some amazing results and has been used in software rendering for years.

 

[Brimstone]

No dram is just straight DRAM. All DRAM is specifically optimized for a combination of error resistance, refresh rate, cost, latency, power, and bandwidth. People have been doing small bank DRAMs well before Mosys existed. And certainly have been doing it with eDRAM for a long time.

Aaron Spink
speaking for myself inc.

Before the Gamecube, Mosys provided Multibank Memory for Tseng Labs ET6100 2d graphic accelerators. Not only does the Gamecube use Mosys technology for eDRAM, but also as main RAM. The performance/cost ratio for the Mosys DRAM architechture was compelling enough to have NEC specifically fab it just for the Gamecube.

The way you're classifying things, SGRAM, MBRAM, VRAM, and WRAM would be labeled just DRAM. This seems overgeneralized to me.

 

[Barbarian]

I'm pretty sure Barbarian is referring to DXT1, perhaps he can clarify?

Umm, yeah sorry I meant DXT1.
Having said that my original post was actually a question. I don't know how many megs of texture cache would provide this kind of guarantee. The thing is, the claim that any DXT1 fits in texture cache was made by an official source at a recent devcon. It sounds a bit absurd to me but for now I don't really have any other information that would contradict this statement.