First Next-Gen CELL Processor: 2 PPEs - 32 SPEs - at least 1 Teraflop
|
From: Thor Lancelot Simon on 9 Dec 2006 23:31 In article <_PLeh.901$yC5.27(a)newssvr27.news.prodigy.net>, mike <mike(a)mike.net> wrote: > >If I were a business development guy in the IBM microelectronics group >and looking for the next big opportunity, I would talk to CISCO about >the IP protocol engine and the encryption engines mentioned above for >starters. If that did not fly I am sure there are several other >router companies to talk to or other special SPE's to think about. Well, Cisco can pretty much do what they want, but I doubt they'll do much on the basis of anything anyone says here. I should note that there are already some off-the-shelf network processors (I'm thinking of the Xscale models with their "millicoded" front-ends that execute something roughly like 16-bit MIPS instructions) that are probably more architecturally similar to Cell than anything else designed recently -- though this just occurred to me. I don't think Cisco uses these but it's pretty darned clear that that's who they're aimed at! I believe this is actually one of the two pieces of StrongARM/Xscale that Intel did not shed to Marvell so they must think it's worth something. There are not many sane crypto processors out there and nothing that is really equipped to handle, in an efficient, record-at-a-time way, multiple, complex crypto protocols. So for example you can buy something from Broadcom or Hifn that will decapsulate and decrypt an SSL stream for you, and they might give you firmware that does IPsec, but you can't alter it to do OpenVPN or WPA or SSH packet processing in the same efficient, packet-at-a-time way, at least not without help from the vendor that is usually not offered. And *nobody* likes relying on black-box vendor firmware for this stuff, but doing the packet processing in lockstep between the host CPU and accellerator, as is required to use the accellerator's raw cryptographic transforms, is just too slow at modern data rates. A cryptographic version of Cell seems like an incredibly elegant way out of this box (whether the _whole thing_ is used as an offload processor for another CPU, or the PowerPC core is used as main CPU for the overall application and the SPEs as crypto processors). The question is whether IBM thinks they could overcome existing relationships with crypto chip vendors to sell it, I would think, because the win on technical grounds seems pretty clear. -- Thor Lancelot Simon tls(a)rek.tjls.com "The liberties...lose much of their value whenever those who have greater private means are permitted to use their advantages to control the course of public debate." -John Rawls
From: jsavard on 9 Dec 2006 23:46 John Dallman wrote: > quoting me: > > > What I want to know is if they're ever going to make a chip that > > concentrates on *double precision* (64-bit) floating-point > > performance. > > Unlikely while they're concentrating on graphics and audio. And while > CELL design team concentrates in those areas, it's unlikely to get > take-up in the areas that require double precision. Chicken, egg. True, but that isn't too serious if someone else, concentrating in other areas, makes available a double-precision workhorse at reasonable prices. There already is the 96-element ClearSpeed chip, and there may be a DSP out there featuring double-precision as well. John Savard
From: mike on 9 Dec 2006 23:59 "Del Cecchi" <delcecchiofthenorth(a)gmail.com> wrote in message news:4u1el7F162bfrU1(a)mid.individual.net... | | "Robert Myers" <rbmyersusa(a)gmail.com> wrote in message | news:1165712275.771127.89640(a)j44g2000cwa.googlegroups.com... | > Del Cecchi wrote: | >> | >> An interesting idea. Now, how would IBM make money from this project? | >> Is there some organization that would pay IBM for such a chip? Who | >> would | >> buy it and how many would they buy if IBM built it on spec? | >> | >> If you represent an organization with money, email me and we can get | >> the | >> project underway as soon as the deal is signed. :-) (It's easy to | >> figure out seemingly good ideas, harder to make money off them) | >> | > | > A little testy, no? | > | > IBM is one of the few non-government players in a position to take a | > leadership role. | > | > As we are going now, the US is aiming to be a second or third class | > player in technology and science, and much of the negative leadership | > is coming from players who think themselves the brightest thing in the | > firmament. If the nobodies speak up, they're doing no worse than the | > somebodies doing all the talking. | > | > Robert. | > | | All these folks that love to tell others what they "ought to do" as if | the folks in a position to actually do stuff are dough heads or something | sometimes bug me. I merely wanted to point out that economics enter | into the equation. And a little more rigor than "hey wouldn't it be | great if..... we painted it candy apple red and put on a 6-71 blower" | would be nice. | | IBM did take a leadership role with Blue Gene. It was specifically | developed to be useful for a certain class of problems. It turned out to | also be useful for some others. Now they are taking a leadership role in | finding new applications for Cell, including some that the government is | paying for. | | At the very least those making suggestions should think about and discuss | tradeoffs involved. | | And yes, I suppose I was a little testy. | | del I did not mean to imply that IBM should do this "for the good of all man kind" or build such systems on speculation, but that if a partner / customer can be found, Cell offers a way to use very small simple specialized SPE's to make a very fast system. Mike Sicilian
From: Piotr Wyderski on 10 Dec 2006 09:46 Robert Myers wrote: > As we are going now, the US is aiming to be a second or third class > player in technology and science, and much of the negative leadership > is coming from players who think themselves the brightest thing in the > firmament. So simply do not buy their products and let their competitors fulfill your needs. If you are wrong, nobody will see the difference and the mistake will be your problem. If you are right, the competitors' income will increase, so it will become IBM's problem. I am not an IBM defender (well, I keep away, as far as possible, from their software products, frankly speaking), but here I support their way: how much money can you make from designing and selling such a chip? If not much, then why shoud they bother? Best regards Piotr Wyderski
From: Del Cecchi on 10 Dec 2006 11:40
"mike" <mike(a)mike.net> wrote in message news:_PLeh.901$yC5.27(a)newssvr27.news.prodigy.net... > > "Del Cecchi" <delcecchiofthenorth(a)gmail.com> wrote in message > news:4u113vF16253pU1(a)mid.individual.net... > | > | "Thor Lancelot Simon" <tls(a)panix.com> wrote in message > | news:elfhh4$qq3$1(a)reader2.panix.com... > | > In article <sWFeh.801$yC5.90(a)newssvr27.news.prodigy.net>, > | > mike <mike(a)mike.net> wrote: > | >> > | >>What if IBM abstracted out the SPE interface logic and management > | >>software (once it is perfected) to allow third parties to re-use > the > | >>Cell's high-level design with different specialized SPE > instruction > | >>sets? Just as they re-use the PowerPC processor architecture on > | >>several projects, they could re-use the cell multi-processor > | >>extensions to the PPC. The new SPE's could be mini-PPC machines > or 64 > | >>bit vector floating point engines or IP protocol engines, or > | >>compression / de-compression engines, or encryption engines, etc. > | > > | > That's a beautiful idea. I hope IBM listens -- or already has, to > | > someone who generated it internally. > | > > | > -- > | > Thor Lancelot Simon > | > tls(a)rek.tjls.com > | > "The liberties...lose much of their value whenever those who have > | > greater > | > private means are permitted to use their advantages to control > the > | > course > | > of public debate." -John Rawls > | > | An interesting idea. Now, how would IBM make money from this > project? > | Is there some organization that would pay IBM for such a chip? Who > would > | buy it and how many would they buy if IBM built it on spec? > | > | If you represent an organization with money, email me and we can get > the > | project underway as soon as the deal is signed. :-) (It's easy to > | figure out seemingly good ideas, harder to make money off them) > | > | del > > > OUCH! > > I am a retired CIO and unfortunately I do not have a $billion to throw > at the idea or a business that could make a $billion from it. > However, IBM has already talked about "an open PPC development > strategy". They also used the PPC to develop business with the big > three game console companies. > > If I were a business development guy in the IBM microelectronics group > and looking for the next big opportunity, I would talk to CISCO about > the IP protocol engine and the encryption engines mentioned above for > starters. If that did not fly I am sure there are several other > router companies to talk to or other special SPE's to think about. > > The point is that the Cell architecture if not the current Cell > implementation could provide the basis of a whole new market area for > IBM. There are any number of potential partners that need cheep > cycles to make their next generation products. > > Mike Sicilian > > Cisco is a good idea. Actually it wouldn't surprise me if someone were to already be talking to Cisco about such things. IBM sells them a lot of stuff, and I almost certainly wouldn't know about what they were pitching. And sometimes the best ideas don't come from the marketing folks. del > |