Software vs Hardware
|
Prev: First Next-Gen CELL Processor: 2 PPEs - 32 SPEs - at least 1 Teraflop
Next: old microcode listings
From: Christoph Breitkopf on 13 Dec 2006 05:50 nmm1(a)cus.cam.ac.uk (Nick Maclaren) writes: > In article <1165981337.336317.84290(a)80g2000cwy.googlegroups.com>, > "Robert Myers" <rbmyersusa(a)gmail.com> writes: > |> > |> This whole discussion is surprising. Implementations of the TCP/IP > |> stack in hardware can be bought off the shelf. Every contributor to > |> this thread must know that. There's a card out especially for gamers > |> that implements a network stack. > > The whole stack? I didn't know that. Do you have a reference? He's referring to this: http://www.killernic.com/KillerNic/ Has some features typically only found in dedicated server network cards, but it certainly does not implement the whole TCP stack. AFAIK at least - I don't know anything about network stacks. Regards, Chris
From: Terje Mathisen on 13 Dec 2006 06:31 Robert Myers wrote: > This whole discussion is surprising. Implementations of the TCP/IP > stack in hardware can be bought off the shelf. Every contributor to > this thread must know that. There's a card out especially for gamers > that implements a network stack. > > If there were any edge to anything else, somebody'd be doing it, if > only for the gamers. If it isn't being done, it's because there's no > advantage to it. Robert, that 'Killer NIC' is very close to pure bullshit, i.e. they sel it because they can make money on gullible gamers. If they do shave a few us of the ping time, it is because they cheat, i.e. tweak their TCPIP implementation, not because they have hardware on the card. Terje -- - <Terje.Mathisen(a)hda.hydro.com> "almost all programming can be viewed as an exercise in caching"
From: Benny Amorsen on 13 Dec 2006 08:06 >>>>> "CB" == Christoph Breitkopf <chris(a)chr-breitkopf.de> writes: CB> Has some features typically only found in dedicated server network CB> cards, but it certainly does not implement the whole TCP stack. CB> AFAIK at least - I don't know anything about network stacks. It certainly runs _a_ whole network stack. The thing is an embedded Linux system. How they can get a latency advantage from shuffling packets through the Windows network stack as well as the Linux one is somewhat baffling to me. /Benny
From: Dennis M. O'Connor on 13 Dec 2006 08:53 "Terje Mathisen" <terje.mathisen(a)hda.hydro.com> wrote ... > Robert Myers wrote: >> This whole discussion is surprising. Implementations of the TCP/IP >> stack in hardware can be bought off the shelf. Every contributor to >> this thread must know that. There's a card out especially for gamers >> that implements a network stack. >> >> If there were any edge to anything else, somebody'd be doing it, if >> only for the gamers. If it isn't being done, it's because there's no >> advantage to it. > > Robert, that 'Killer NIC' is very close to pure bullshit, i.e. they sel it > because they can make money on gullible gamers. Actually, HardOCP did some blind testing with and without it, and on some games (WoW, for example) it did give an improvement in gameplay. The article is on at http://enthusiast.hardocp.com/ I'd say the "Killer NIC" isn't really for the gullible, no more than tuned exhaust headers in cars are. Both serve the desires of people with the money to pursue that every last little bit of performance. I'd expect "Killer NIC" users to already be using overclocked, aggressively-cooled systems with maxed out DRAM and dual- or quad-ganged high-end video cards. For people in that category, the cost is no big deal, and there is nowhere else to go for the performance. These are gamers. It's a hobby. That means there is no "reasonableness" criteria to apply: just whether it is fun. -- Dennis M. O'Connor
From: Robert Myers on 13 Dec 2006 11:33
Nick Maclaren wrote: > > The whole stack? I didn't know that. Do you have a reference? > I was referring to the products of http://www.alacritech.com/, and, no, it does not implement the entire stack in hardware, just those parts that it is advantageous to implement in hardware. When I became aware of the "killer NIC," I revisited the subject of TCP/IP offload and satisfied myself that there were already products existing that would do whatever a sensible person wanted to do. I must say that I admire the creators of the "killer NIC" for spotting a niche and occupying it. The subject of TCP/IP offload has been discussed here seriously and extensively, but I don't know that the subject has been exhausted. A serious revisiting of the subject would acknowledge that the problem and the possibilities have long been known (Del's point, I think), examine the current state of the art, and ask what possibilities remain. I'm not so interested in games myself, except as drivers of technology, which they are. Communications limitations, not processor technology, are the limits to a truly scalable supercomputer. At the moment, most of the effort and most of the money is going into a technology where the game has already been conceded before it is started: the best the communications hardware can do is not to muck it up any worse than the off-the-shelf communications protocol and software it must rely on. Robert. |