Anandtech Reviews Killer NIC

full source to the linux drivers will be made available (and can be gotten now via closed beta... just pm me in our forums or shoot me an email)

We intend to put the drivers into the main linux distros when they are stable and have had more testing.
 
Xipher said:
10ms is a big difference that may have simply been caused by a change in route due to congestion. To truly test this you need an entirely self contained network.

Scratch the network, just drop a crossover between the client and the server.
Click to expand...

Could be for most people..yes, however not in this case.
;)
 
Unless you've got a direct link to the server (or are staring at trace routes logs) you're not going to be able to tell what route your packets are taking. Don't forget that IGRP is going to forward your packets to different routers depending on current traffic so your actual path is changing on a per-packet basis.

I'd be willing to give KNIC a shot should the price ever become something reasonable ($50USD or so) but at this point I have no need of a third hardware firewall, nor something that can't improve my pings due to my crappy ISP.
 
Gillette, Sorry your ISP is crappy. It is true that a crappy ISP can ruin a gaming experience.. (Killer helps some, by recovering from nasty packet loss better than other cards). I hope you can get a better ISP soon!

However, I don't believe Killer will ever be ~$50: it's not engineering-wise possible. (and not really the point of Killer)

Killer is about being the best, elite, baddest ass network card out there (and lot's of review sites think it is just that http://www.killernic.com/KillerNic/KillerNewsReviews.aspx ).

As a result, Killer has been packed full of hardware that isn't normally part of a simple NIC (a lot of it quite expensive):
400Mhz PPC processor + custom designed ASIC = 1 NPU.
64 MB of DDR memory (no other network card has Ram at all)
8MB onboard flash
USB port (so you can connect USB HD or FLASH to your Linux in Killer)
========
One badass card: capable of running any linux app you can throw at it (and an easy way to write said programs...).

As a sidenote: there are other cards out there (mostly for servers) that cost WAY more and do WAY less... [especially for gaming]
$995 card here... http://linuxdevices.com/news/NS4062041190.html
and even a ton of Intel NICs cost $200+ (even just simple dualport Gig nics)
 
Well to be honest, I am interested in the killer nic as the nic's on my mobo are crap, the nvidia nic constantly drops and the marvell one isnt that much better.

I like having leet shit in my pc and I guess the killer nic could'nt be worse than my purchase of an unused Ageia PPU that has only been used for a total of about 2 hours in 6 months or so.

If it really does show improvements over all other onboard nics and doesnt have browseing problems like I have heard then it is likely that I will shell out for one if the space in my pc permits.

If it really does help with onine gaming then I suppose that can only be a good thing.

I guess its time to look for a UK retailer to see what these things are going for....
 
Why has the card not been reviewed here? I already read the "company evaluation" article by Kyle but I would like to see a review done.
 
My opinion in short -
More than likely good technology that's too expensive compared to performance right now. Add some Fnapps and drop the price a little and it becomes much more palatable. $50 sounds unrealistic, but hey we all watch the price of components, $150 within 1-2 years is likely.
At that point it will leave the early adopters hands, the rest of us will have had a chance to actually use one. Unfortunately Bigfoot has gone into probably one of the hardest to prove subsets - my ping rates can vary 10 percent easily within an hour on a server, so how to prove the card is giving me a 10 percent increase, and it's not just a lucky moment?
 
While latency improvements has been the primary topic, it's not the only one they are advertising. They have also mentioned frame rate improvements. I have read the "white paper", and while I find most of it an annoying use of buzz words, some of the concepts did seem sound in relation to checking for and reading network I/O.

I don't think it's worth the current cost, and it's also not something that can be tested with the usual game benchmark (since network I/O isn't normally involved in that case).
 
"Our current opinion is, without FNapps, improved performance across a wider variety of titles, and a significantly lower price tag, this card is destined to be nothing more than an interesting footnote in the annals of hardware history."


Interesting.
 
tomstomper said:
(Killer helps some, by recovering from nasty packet loss better than other cards).
TS
Click to expand...
I'd love to see a technical explaination of how that works.
 
number69 is quoting the anandtech review, which verified TONS of the results were were suggesting (5+ms ping improvements, 5+fps improvement)...

reyalp, here's a tech. explanation of packet loss recovery.

The basic steps are always:
1.) Detect Packet Loss.
2.) Send Request for missing packets.
3.) Recieve Missing packets.

Killer helps with packet loss because:
1.) Detect Packet Loss --- Killer can detect some packet loss internally, and automatically adjust link speed settings to compensate.
For packet Loss Killer cannot detect, the game itself has to detect (via time-stamps, sequence numbers, etc.). But Killer interrupts the game with packets so if Packets 1,2,3 should have been received, Killer gives the game packet 1 and then packet 3 faster (due to interrupt driven model and windows network stack bypass model).

2.) Send Request for missing packets. (a request for a missing packet is sent FROM Killer faster, because that request BYPASSES the windows stack, and interrupts Killer as soon as the data is ready.).
3.) Recieve Missing packets. (Same as #1, Killer interrupts the Game when the missing packet is received: BAM bypassing the network stack along the way.

Hope this helps, it does help in smoothing out gameplay, as many users report.
 
tomstomper said:
number69 is quoting the anandtech review, which verified TONS of the results were were suggesting (5+ms ping improvements, 5+fps improvement)...

reyalp, here's a tech. explanation of packet loss recovery.

The basic steps are always:
1.) Detect Packet Loss.
2.) Send Request for missing packets.
3.) Recieve Missing packets.

Killer helps with packet loss because:
1.) Detect Packet Loss --- Killer can detect some packet loss internally, and automatically adjust link speed settings to compensate.
For packet Loss Killer cannot detect, the game itself has to detect (via time-stamps, sequence numbers, etc.). But Killer interrupts the game with packets so if Packets 1,2,3 should have been received, Killer gives the game packet 1 and then packet 3 faster (due to interrupt driven model and windows network stack bypass model).

2.) Send Request for missing packets. (a request for a missing packet is sent FROM Killer faster, because that request BYPASSES the windows stack, and interrupts Killer as soon as the data is ready.).
3.) Recieve Missing packets. (Same as #1, Killer interrupts the Game when the missing packet is received: BAM bypassing the network stack along the way.

Hope this helps, it does help in smoothing out gameplay, as many users report.

TS

Hope this helps, it does help in smoothing out gameplay, as many users report.

TS
Click to expand...

How does this explain smoothing gameplay when the whole packet loss + retransmission isn't even a factor in most games by nature of the protocol used? Aren't most games using UDP for the exact purpose of eliminating steps 2 & 3 (UDP, unlike TCP doesn't care about lost packets, hence no request for missing packets)?
 
I think the only way to justify this to the [H] forum would be to get one of us out to your place and set up a real world test.

Get two identical systems that are identical save for the NIC cards. Pick a game that the person has never played before and get them going, they dont even have to be good at it. Let them play for 30 minutes on each system.

At the end of the test ask them which performed faster and why they thought so.

Then allow that user to verify the hardware in each machine.

Plain and simple.

Im not going to argue with someone on the net about something Ive never used, because people are proven wrong every day.

Id love to see how this all works with linux and it's network stack as Im not an avid gamer, If this can be used in some sort of great application on a server level (like maybe an asterisk server with regards to packet loss) Id be all over it for a enterprise level VOIP server.

even dual killers used in a linux based firewall distro would be pretty damn slick if the packet loss is handled as described.
 
Kaos, not sure if there was a question in there... though I think you have some good ideas! (the FNA firewall app is rather cool, and dual with pkt forwarding would be sweet).

da_sponge, I love talking about this stuff, and it is my life! :)

Yes, most games use UDP! And you are correct, the UDP protocol cannot detect packet loss by itself, BUT, nearly all modern games implement some form of packet loss detection for 'reliable' packets on top of UDP (e.g. a layer on top of UDP). That's why only the "game" can detect it as it is running the layer on top of UDP. (and so, it's why interrupting the game with packet 3 as fast as possible improves detection).

Incidentally, Link-level packet loss can also be detected by Killer (e.g. Collision detection, packet link errors, etc.) as this is a layer 2 function, and certain layer 2 actions can be taken (such as trying a lower link speed setting, etc.).

Hope this helps, and i'm frankly quite impressed by the level of knowledge here at HardOCP. Way above par.
 
No question, just posting some thoughts.

Ive never been much of a PC gamer but I live in the world of enterprise services at work and would love to see something thats already stable (asterisk or even VOIP in general ) become improved because of some advancement.

Here's a question or two, Im not sure if I read you correctly but it works with XP's network stack to increase efficiency. Im assuming the linux drivers would include this functionality? and what about vista?

Have you had much resistance in the new way that microsoft is forcing signing of drivers as well for vista (if vista is even on your plate now)?
 
Yah we've been working on Vista for a while, and it is EXACTLY the same as XP (for UDP networking). TCP is changed quite a bit though.

Linux is a different beast: we're not focusing on Stack Bypass or Gaming Optimization because gaming isn't done much in Linux, instead we're working to make the drivers stable (they are pretty good and in Beta right now), and expose FNA in the Linux OS... (so people can run FNApps in their Killer with their Linux OS).

One of the interesting aspects of having a Network Processing Unit in a system is it's proximity to the "wire". What that means is not just savings in CPU, but savings because it is close to the wire.... e.g. reduced Bus and Memory utilization. But I digress! :)

Thanks for the questions,
 
tomstomper said:
Yah we've been working on Vista for a while, and it is EXACTLY the same as XP (for UDP networking). TCP is changed quite a bit though.

Linux is a different beast: we're not focusing on Stack Bypass or Gaming Optimization because gaming isn't done much in Linux, instead we're working to make the drivers stable (they are pretty good and in Beta right now), and expose FNA in the Linux OS... (so people can run FNApps in their Killer with their Linux OS).

One of the interesting aspects of having a Network Processing Unit in a system is it's proximity to the "wire". What that means is not just savings in CPU, but savings because it is close to the wire.... e.g. reduced Bus and Memory utilization. But I digress! :)

Thanks for the questions,
TS
Click to expand...

Tom, not to go off topic here or nag, but please stop double signing. It's rule #25.
 
Roger, no more double signing...

tomstomper said:
Hope this helps, and i'm frankly quite impressed by the level of knowledge here at HardOCP. Way above par.
Click to expand...
 
Yeah, Maximum PC gave it pretty much the same review... Honestly, the first time I saw an banner ad for it, I thought it was a joke..

I can't see spending so much for a NIC, but who knows..

So now maybe if they came out with a folding client for it?
 
tomstomper said:
Hope this helps.
TS
Click to expand...
I suppose it does in a way.

Most packet loss happens somewhere on the far side of the users modem. As you admit, killernic can't do anything about this.

Killer may be able to handle link layer stuff slightly better, but this is a red herring because on a properly functioning home Ethernet, the rate of these problems will be extremely low. If it isn't, the solution is to fix your hardware or upgrade as required to handle your actual usage, not get a NIC that handles errors slightly faster.

If the game is saturating the users connection, killernic can't do much about that either. It could drop packets, but that is clearly not helpful. It could also smooth out bursts, but typical home connections usually have plenty of buffer to do that. Games include their own rate limiting mechanism (usually user controlled, like quake3's rate, snaps, maxpackets cvars) precisely because rate limiting isn't something that can easily be handled transparently for game protocols.

The whole "interrupting the game" business doesn't make any sense. A typical game polls once per frame. The game cannot accept packets from the network stack at any other time. Killernic may respond to the poll faster, but it cannot "interrupt" the game to give it a packet when it isn't ready for one. (In fact, I'd wager that most of the gains from killernic are because it avoids interrupts and context switches to handle incoming traffic, and so is more cache-friendly for the game. The is something the HPC crowd has been doing for years.)

The rest of your description all boils down to "killernic is faster", which may or may not be true, but has little to do with packet loss per-se. Any significant packetloss will be detrimental to the game experience, regardless of what NIC you use. The time required for the missing data to arrive is typically going to be so much larger than the time required to read the packet from the NIC that it isn't even woth mentioning.

In conclusion, I don't see how you can honestly claim that killernic will noticeably improve the gameplay experience for people with bad connections. Your reply seems to be taking something that has some grain of truth, but little or no practical significance, and attempting to use that to sell your product to people who don't understand the technical details.

I'm not saying that killernic doesn't have benefits, but this particular claim, like many others in your marketing materials, seems unsupportable at best and dishonest worst.
 
reyalp,

I'm bummed you see it that way: and I was hoping (and still hope) you were interested in discussion 'truth' of how the technology works, rather than any attacks. I'm an engineer, not a marketing exec., so I can't say if our 'claims' are too aggressive or not in the mktg. team: if they are, then that sux, and I hope they tone it down to reality.

So, what is reality? (Can we keep it technical and civil?)

First, you are correct again, a better solution to "Link level" loss is an equipment upgrade (like a new router, or replace a cable)... but not every user does/knows to do this: so having a NIC that will try to help find link-level solutions, in my opinion, is cool.

Second, on "interrupting" the game. It's possibly an overly simplistic statement, because (you are correct again) games will implement different types of network socket usage patterns.
(Our CEO actually wrote a white paper on this you can find it here if you are interested: http://www.killernic.com/KillerNic/KillerAboutSpecs.aspx
In short, as I understand it, polling IS done in some games as you say, and in those games Killer interrupts a Socket State variable that flips a bit so that the next poll loop the game gets the data immediately.. and that state bit is used to return 'immediately' to the game when no data is available (one of the ways FPS improvements are so profound in some games). [I also think this is cool]
For some games (many modern games) a more multithreaded approach is used, and blocking or asynch calls are made which CAN BE INTERRUPTED by Killer (and thus those games see more ping benefit).

The net of the above, in terms of FPS and Ping improvment are obvious, what might not be so obvious, and was my main point, was that when a packet loss (detectable by the game app) occurs, the effect of the ping improvement would be 3x that of normal traffic (becuase 2 rx's are faster and 1 tx is faster).

So, I think that's fully explained. I appreciate the good hard questions.
Anyone got any others? :)
 
tomstomper said:
First, you are correct again, a better solution to "Link level" loss is an equipment upgrade (like a new router, or replace a cable)... but not every user does/knows to do this: so having a NIC that will try to help find link-level solutions, in my opinion, is cool.
Click to expand...

I assume you're talking about packet loss. So what would you recommend to someone having packet loss, a $280 NIC or calling their ISP to fix the issue? Possibly free to a 50 dollar charge. If there is packet loss and the NIC just makes it more managable, I'm sure fixing the issue would be better for performance. So when the killer does fix or smooth out this packet loss, does it tell the user that the problem is there so that it can be fixed?
 
YARDofSTUF,

just to be clear, regarding link-error corrections of Killer. It's a minor, and non-advertised feature, but a good example of the little things that can be done because there is Linux OS on the card.

With Linux, we (and any users in fact) have the ability to run procs and cron type jobs that can do things like actively detect packet loss (link level). So, we put in a feature (silently without any fanfare or marketing that I'm aware of) to detect when there is a link speed mismatch autonegotiation failure (usually old routers), etc.

Yes, the Killer informs the host OS of the problem, and we'll be adding more status as well for other things.

Bottom line, we're just continuing to try to make this THE BEST online gaming network card available, and little features like this are icing.

I do agree though, that a user, certainly a reader of HardOCP, should just fix the problem with a new router/isp call.... :) But for the rest of the world who don't want/can't do that it's a nice little thing.

The real meat is when packets are lost somewhere outside the link-level last hop, and that's where having a network stack bypass can lead to better performance due to early detection and faster recovery as has already been discussed.
 
If I were building an "ultimate" gaming rig Id put one in if cost were no concern to me, I understand people being doubtful but the truth is until you try one then youll never really know. The unfortunate thing is that I bet ALL of you would love to take this thing for some real world tests but its cost does make it hard to justify "just to see" kind of tests.
 
You must log in or register to reply here.
Back
Top