Toshiba samples Cell-based HD GPU
PS3-sourced video-oriented chipage
Toshiba has begun sampling the graphics chip it hopes will beat Nvidia and AMD at the GPU game: the Cell-derived SpursEngine SE1000.
Toshiba's SpursEngine SE1000:
Oddly named it may be, but the SpursEngine packs in four of Cell's Synergistic Processing Element (SPE) cores. Cell has up to eight SPEs to handle vector data, along with a PowerPC-based general purpose computing core for everything else.
Each SPE is a SIMD Risc processor specialising in single- and double-precision floating point maths. The SE1000 also has a memory controller on board that links the SPEs to 128MB of Rambus XDR memory over a bus capable of pushing data at 12.8GB/s.
The reference board
Toshiba also said the chip is capable of MPEG 2 and H.264 encoding and decoding at 1080p full HD resolution, and those are the applications it's pitching the product at rather than 3D graphics. It said it's partnering with the likes of CyberLink, Leadtek and Corel - all makes of widely used PC-oriented video playback apps - to add support for the SE1000 to their software.
The SE1000 reference card fits into a 1x PCI Express 1.1 slot and runs the GPU at 1.5GHz. The chip's capable of 48GFlops, Toshiba said - 12GFlops per SPE. The card consumes just 10-20W of power.
The Japanese giant said it hopes to sell 6m SE1000s in the first three years of the chip's general release. It announced SpursEngine in September 2007.
COMMENTS
Re: 1x PCIe
Well, the SpursEngine was originally stated to be used in low power applications like TVs and laptops.
But, I'm quite interested in this particular board as it looks fairly sparsely populated. If that chip can indeed be passively cooled with a thin heat sink, wouldn't it be neat if they packaged it down to fit in a mini-PCIe slot for use in laptops? Finally, a piece of tech to use in my laptop's spare mini-PCIe slot next to the WiFi card.
Alternatively, perhaps they could fit it into an ExpressCard form factor and use the PCIe x1 connection there.
Though after thinking about it, they probably can't do it any time soon with that power envelope of 10-20 watts. It might be doable with a lower voltage, lower clock speed variant in a smaller manufacturing process in the near future, though.
Cell bandwidth?
Methinks Michael Sanders is confused over the legendarily slow read speed the Cell has on the RSX's memory pool.
Not that it matters anyway, as the RSX itself has really fast write speed to the Cell's main memory pool, where it can obviously read at the 23GB/s mentioned by others.
This strikes me as very handy for media PC boxes, but then if a 1080p Cell-based media player is what you want, a PS3 is pretty reasonable already.
@Michael Sanders
please tell us how the PS3 Cell's 23GB/s bandwidth is slow. We are waiting...
Or have you got your coat already?
shame its not the Cell PPC with Altivec on a card
as stated its a shame it's not got the PowerPC-based core or Altivec onboard.
rather than take the lower bin (ps3 failed cores)PPC/Altivec based Cell yealds and use them in this space, this spursengine totally removed the PPC and Altivec as found on the PS3 so no using the exising (linux)PPC codebase for lots of fun and thats a real shame.
however, lets be clear , this chip has been demoed to be able to do some really clever playback DECODING of many DH video streams at the same time.
and more than that ENCODING is fast on them, so you can hopefully soon get a DVB-* card that can then feed its input to one of these spurs and encode the whole transport stream inside to 4.1 grade AVC perhaps in realtime and then push it out your LAN as a multicast stream, to be selectively played back on any conected device.
thats the 'whole stream at once' (on average 3to5 channels), not mearly one single DVB channel if you will.
Slow memory access?
Say what now? Cell as implemented in the PS3 has something like 23GB/s bandwidth. How is that slow exactly?
