Bitboys and 'Graphics hardware for handheld devices'

Thou shalt not talk about Bitboys...

Nappe1,

my (bad) joke wasn't an attempt at Bitboys bashing and I didn't want to offend you.

But I have to say, it isn't really wise to talk about them in public currently, especially if you're touchy when it comes to others making jokes about them. It isn't a good idea until you are really, really sure that they will, in fact, release something.

Bitboys are way beyond the "put up or shut up" point, and you've got to hand it to them, they have kept rather quiet in the last two years (?), aside from some appearances on exhibitions (heh). So let's just wait and see if they're going to come out in the open by themselves.

Chris

 

I believe there is one transister per cell/bit in eDRAM.

 

I do know that 1T-SRAM (as seen in the Gamecube) works that way, but I'm pretty certain eDRAM is different. Now, I don't have any other argument than it doesn't make sense for modern memory chips to have around 256 million transistors (16 chips on a 512M DIMM).

 

eDRAM and DRAM are different processes.

eDRAM is fabricated in standard CMOS logic. DRAM is fabricated in a different process that allows denser capacitors, which is how you can get oodles of bits in a die. Plus DRAM is generally fabricated on a very very aggressive fab that is designed specifically for DRAM.

But, even then, I am fairly certain that there is at least 1 transistor per cell.

Of course, I'm just going on what material I can find on the net, plus what basic knowledge I learned in school and on the job. Maybe ASICnewbie will speak up and share with us some more concrete info.

 

Yes, it appears that TSMC uses a 1T-RAM technology for their embedded memory. I think the BitBoys were of the belief that they could actually get the same denisities out of eDRAM as normal DRAM can get. Speaking of which, I wonder if technology will ever get to that level?

 

Of course it will. How long are you willing to wait

 

I don't doubt that it can happen, but Quantum technologies are coming...once they make it, what would be the point of furthering silicon technology?

 

I'm still can't wait for their card though, I think it'll rock.

 

Lock please

 

That is also my understanding.

 

I am fully behind you Nappe1.

If everyone chose their words a little more carefully and respected each other a little more then things wouldn't heat up like they often do.

Personally I am not here to offend anyone regardless if they think that a GF2 MX is just fine for todays games. In a situation like that I would respond that if they're happy then that's fine with me although I demand a lot more. Different people => different needs and pleassures. As long as I get what I want I'm a happy lad.

 

I've been told the same thing. Historically any type of ASIC embedded memory (e-DRAM, masked ROM, flash ROM, etc.) has required additional mask steps, which means longer production time and higher manufacturing cost.

1T-SRAM retains the almost the same bit-density (it's slightly less dense) as traditional e-DRAM, but does NOT require those additional processing steps. So it's more "ASIC CMOS friendly"

I think all embedded memories have higher transistor densities than 'regular' logic (analog or digital.) Consequently, embedded memories are more sensitive to manufacturing defects. The larger the die-area of the embedded memory, the higher the added risk.

If a customer design requires a large embedded RAM, TSMC recommends some type of "defect management", to avoid that proverbial "one bad apple in the whole bunch" from basically killing the whole die.

One approach (which is already standard practice for standard memory ICs) is to add 'spare memory' rows/columns to the die. During the testing process, test-equipment probes the die specifically for memory array defects. If there are enough spares to 'repair' the RAM, a laser maps the spare rows into the address-decoder array by burning tiny circuit fuses. This service is available at an extra cost.

Another potential approach (but not always feasible) is to add built-in intelligence to manage the defects. In this case, RAM-defects aren't repaired in hardware, but instead 'mapped out' during the ASIC's power-up runtime. This might work if your chip has an on-board CPU or intelligent controller. During power-up, the CPU first runs a diagnostic on the embedded RAM-array (like your motherboard BIOS's POST), in order to identify defects. Presumably, the CPU could would remember and avoid the memory-defects. As you can see, this arrangement is only practical for certain, specialized applications. But it eliminates the need for the laser-fuse circuitry and the associated laser-repair service.

If the embedded RAM is relatively small, it might be less expensive to simply 'do nothing'.

 

Honestly we are at least 20 or more years away from mainstream Quantum computers. Don't expect them any time soon, and silicon systems won't go quietly. There's a lot of room for improvement in current proven technologies that will be realized long before Quantum computers even come close to reality.

 

Let me just make a couple of comments here.

I don't believe silicon technology will die out any time soon. In fact, I'm pretty certain that they'll be very commonly used as "controllers" for Quantum-based processors. Given the potentially monstrous amounts of data that could be computed by Quantum processors, such controllers could easily be far more complex than today's processors.

Regardless, how far we are away from mainstream Quantum Computing depends almost completely on economic factors (i.e. Even if we have a usable, workable QC in five years, it may be a number of years later before they make any real sense for the consumer sector). I'm sure I'll know far more about it in another year

 

Quantum computers might actually be based on "todays" silicon litography technology, see this EETimes article. And don't forget that programming a quantum computer will be very different from what we have today. Don't expect them to run serialized code all that well. What makes them interesting is that their computing power grows exponentially to the number of qubits it can process, so it's situations of extreme paralellism that will make them shine. Fortunately for us, computer graphics is such an area!

Regards / ushac

 

Yes, a Quantum Computer, for certain types of processing, is essentially capable of infinite parallelism, computing all possibilities at once (i.e. algorithms where you do the exact same processing on different inputs are where this is possible). Physicists are very interested in these possibilities, but personally, I'm not. This is very constraining, and only lends itself well to certain calculations. While it will be interesting to see what sorts of calculations will be done with such computers, I'm more interested in the advancement of general-purpose computers. My current (albeit uneducated) belief is that it should be possible to build a massive array of molecule-sized quantum processors that would all operate together in a fashion similar to today's FPGA's (field programmable gate arrays).