Sign in with
Sign up | Sign in
Your question

HyperTransport 3.0 Specs

Last response: in CPUs
Share
April 24, 2006 6:52:16 PM

I don't know (or care) if it's been posted before.

Looks like the specs are similar to HT 2.0, which is great, it means ULTRA-LOW Latency while pushing bandwidth to new levels.

1.8GHz/2.0GHz/2.4GHz/2.6GHz

Provides up to 20.8GB/s per Duplex (41.6GB/s Effective) while maintaining Dual x8 or Dual x16 Links, as well as the possibility for Dual x32 Links, which means Inter-Device Coherency is brought to new levels and means that scalability to 32P Servers for sF 1207's will have the bandwidth they need and at the same time, have the same latencies or less than that of previous versions.

They've also added in the ability for 8B/10B Encoding, which will increase the integrity of the data transmitting as well as allow it to be transmitted farther then previously.

Also HotPlugging: The ability to add/remove HyperTransport Devices w/o disrupting system state.

~~Mad Mod Mike, pimpin' the world 1 rig at a time

More about : hypertransport specs

April 24, 2006 7:19:00 PM

That sounds pretty sweet.

With the hot plug feature, say you have a system that has 3 processors but has an open slot for a 4th. Would that enable you to plug in a 4th processor without turning off the system? Be pretty sweet.
April 24, 2006 7:35:03 PM

Quote:
That sounds pretty sweet.

With the hot plug feature, say you have a system that has 3 processors but has an open slot for a 4th. Would that enable you to plug in a 4th processor without turning off the system? Be pretty sweet.


It isn't specific on the exact devices, but I would surmise it to be possible. I think the best thing would be in large 32P Servers where you have, say, 16 or more CPU's, and if it's an extremely large dependant server, rebooting it is probably not plausible, so one could add in additional CPU's or Co-Processors into the slots w/o disrupting the system.

~~Mad Mod Mike, pimpin' the world 1 rig at a time
Related resources
April 24, 2006 8:41:38 PM

I dont think they would support hot swapping of Processors... Its just too much money to have a stray charge frying it :(  It does sound cool though even if some one had a 2 CPU system you could leave it up and change out one CPU at a time :)  I do think there was one company that supported the idea though. Might have been Alpha or some other company my brain is failing to remember properly (damned early onset CRS syndrom !) lol anyway I was just reading about HT 3.0 and its support for new ram types and further out HT 4.0 with support for FBDIM2 ! I can hardly wait !!!
April 24, 2006 8:47:26 PM

"Provides up to 20.8GB/s per Duplex (41.6GB/s Effective) while maintaining Dual x8 or Dual x16 Links, as well as the possibility for Dual x32 Links, which means Inter-Device Coherency is brought to new levels and means that scalability to 32P Servers for sF 1207's will have the bandwidth they need and at the same time, have the same latencies or less than that of previous versions. "

Someday i hope to be able to understand 1/2 of that.

8O
April 24, 2006 8:59:54 PM

do you have a link (if there's more info on it...)?

thanks,
Ara
April 24, 2006 9:05:56 PM

Quote:
do you have a link (if there's more info on it...)?

thanks,
Ara


http://hypertransport.org/

~~Mad Mod Mike, pimpin' the world 1 rig at a time
April 24, 2006 9:19:50 PM

On the cpu hot swap, I think an electrical zap would be the least of your worries, Id wonder just how fast I could get a hsf on and running!

I think that because the distance that htt 3 can go, about 3 feet, the hot swap would be used more for things that are external to the mobo.

Maybe be able to have a cable that goes from one server to another and sort of "network" or daisy chain the 2 different mobos if need be, and its all pulg and play.

and MMM it would seem to be you that would know where any more info on this could be found. If you could be so kind as to provide a link that would be great, btw I have read the toms artical on it already.

<edit> seems to have already got that done thanks</edit>
April 24, 2006 9:43:23 PM

Quote:
On the cpu hot swap, I think an electrical zap would be the least of your worries, Id wonder just how fast I could get a hsf on and running!

I think that because the distance that htt 3 can go, about 3 feet, the hot swap would be used more for things that are external to the mobo.

Maybe be able to have a cable that goes from one server to another and sort of "network" or daisy chain the 2 different mobos if need be, and its all pulg and play.

and MMM it would seem to be you that would know where any more info on this could be found. If you could be so kind as to provide a link that would be great, btw I have read the toms artical on it already.

<edit> seems to have already got that done thanks</edit>


Daisy chain mobo's DAMN !!! now THATS an idea and a half ! hmmm I bet it would be cheaper then fibre channel too ? probably faster ? (im super light on clustering tech knowledge lol)
April 24, 2006 10:12:32 PM

So would the daisy chain effect finally allow for hundreds of opterons in a single supercomputer?
April 24, 2006 10:20:18 PM

From what little I understand about fibre channel networks thats what they use those for is clustering (kinda makes multiple machine act like one?)
April 24, 2006 10:33:16 PM

Mike it's so hard waiting for it with the way you describe it!
April 25, 2006 1:05:33 AM

Nice find Mike. Now with HTT 3.0 AMD will choke intel to death with higher bandwidth and low latency. Intel should consider using HTT instead of making some buggy interconnect like CSI but they're to proud and arrogant.
April 25, 2006 4:27:01 AM

Quote:
Nice find Mike. Now with HTT 3.0 AMD will choke intel to death with higher bandwidth and low latency. Intel should consider using HTT instead of making some buggy interconnect like CSI but they're to proud and arrogant.
That is partially true they were'nt so arrogant to go back to pentium 3 architecture when amd was kicking their ass!
April 25, 2006 4:53:29 AM

Hasn't hotswapping being around for ages ?, including processors & RAM ?.

I suspect this would be more for adding stuff to a backplane, etc, without having to power off.
April 25, 2006 9:01:31 AM

Maybe it was Sun's ultrasparc as in
Quote:
up to 12 UltraSPARC III processors, up to 96GB of memory, and hot-swappable CPU/memory boards enabled by Sun's proven Dynamic Reconfiguration (DR) technology.
April 25, 2006 9:33:56 AM

Hotplugging has been around for some time yes, for CPU and RAM too - but not in the Windows OS family: it requires OS support.

Most implementations were proprietary and/or unstable and/or lacked efficiency (Source: CentOS mailing lists).

Since HT is slowly becoming the de facto standard for high bandwidth component interconnects, making its nest installment support CPU and RAM hotswap natively is a great idea.
April 25, 2006 12:31:34 PM

Quote:
I don't know (or care) if it's been posted before.

Are you giving the specs to AMD, so they have to make your wihes to become true or you are just spaming the forumz around?
any link of no fanboy blogger saying a little about your SF story?

BTW Intel CSI specs are above:
2048 GaAs 512bit optic links with 16TB/s bandwidth per link with 0ms access time. So HTT 3 sux a lot. Do you agree MMM or you have another better SF story?
April 25, 2006 1:05:02 PM

Quote:
I don't know (or care) if it's been posted before.

Are you giving the specs to AMD, so they have to make your wihes to become true or you are just spaming the forumz around?
any link of no fanboy blogger saying a little about your SF story?

BTW Intel CSI specs are above:
2048 GaAs 512bit optic links with 16TB/s bandwidth per link with 0ms access time. So HTT 3 sux a lot. Do you agree MMM or you have another better SF story?

Can you possibly put up a reference link for that info? It smells of BS...16TB/s bandwidth...? Did you mean to put a G where the T is in "TB"?? Umm, that's like bandwidth overkill...
April 25, 2006 1:41:00 PM

Quote:
tera bytes no way. tera bits maybe. can people please remeber that capital letters make a difference. one tera equal a million million. so that is 1 x 10 to the power of 12!!! bytes. thats alot of bandwidth if it is true.

1 Terabyte = 2 of the power of 40 bytes.
Not ony 16TB/s, but 2048 links with 16TB/s each, that means 32768TB/s total bandwidth that this technology enabels between the 16384 cores that will be available on the Conroe 2 CPU that will be availabel in q1 2008 or maybe sooner.
Ask MMM and 9-inch for the details, they are using the same sources, but our sources are so confidential and we have to be quiet about them.
April 25, 2006 1:43:01 PM

Any idea when this will be come more than a spec and will be in real products?
April 25, 2006 3:06:38 PM

Quote:
tera bytes no way. tera bits maybe. can people please remeber that capital letters make a difference. one tera equal a million million. so that is 1 x 10 to the power of 12!!! bytes. thats alot of bandwidth if it is true.

1 Terabyte = 2 of the power of 40 bytes.
Not ony 16TB/s, but 2048 links with 16TB/s each, that means 32768TB/s total bandwidth that this technology enabels between the 16384 cores that will be available on the Conroe 2 CPU that will be availabel in q1 2008 or maybe sooner.
Ask MMM and 9-inch for the details, they are using the same sources, but our sources are so confidential and we have to be quiet about them.


They have definitely done something nearly impossible. Do you know how many traces would be necessary to provide that kind of bandwidth? Infiniband isn't even close to that. Not saying you're making it up but you have to admit, 16TB per 4 core chip is rather difficult with mobo traces.

At least HT 3.0 is reasonably fast at 41GB/s. This tech would be 1000x faster.
April 25, 2006 3:21:48 PM

Quote:
I don't know (or care) if it's been posted before.

Are you giving the specs to AMD, so they have to make your wihes to become true or you are just spaming the forumz around?
any link of no fanboy blogger saying a little about your SF story?

BTW Intel CSI specs are above:
2048 GaAs 512bit optic links with 16TB/s bandwidth per link with 0ms access time. So HTT 3 sux a lot. Do you agree MMM or you have another better SF story?

Can you give evidence for that? :?:
April 25, 2006 7:39:27 PM

NO I can not give evidence.
I read the post with no links and I was thinking that is just another stupid SF story of how AMD will be good written by ??? and ????? or written by the AMD fanboy bloggers and rewriten by the same spamers with the stupid smile on their faces.
Like the story of 32nm K8s shipping in Q1 2007, the 4GHz 90nm K8, the Reverse HT, the Athlon 1.8GHz being better than 4 core 2GHz Woodcrest and other no-sence lies that people who don't have much IT knowledge are beliving.
This stupidities are pissing me off. Those idiots are ruining the point of existence of the forums. They should be place where people can learn something right, not the stupid wishesh of unintelegent no-sence limited morons that have nothing to do all the day and are spaming the forumz around. So in their context I am just trying to stop their prulioned minds, becouse it looks like there are no admins and moderators that are filtering the garbage on the forum.
April 25, 2006 9:04:45 PM

Quote:
NO I can not give evidence.
I read the post with no links and I was thinking that is just another stupid SF story of how AMD will be good written by ??? and ????? or written by the AMD fanboy bloggers and rewriten by the same spamers with the stupid smile on their faces.
Like the story of 32nm K8s shipping in Q1 2007, the 4GHz 90nm K8, the Reverse HT, the Athlon 1.8GHz being better than 4 core 2GHz Woodcrest and other no-sence lies that people who don't have much IT knowledge are beliving.
This stupidities are pissing me off. Those idiots are ruining the point of existence of the forums. They should be place where people can learn something right, not the stupid wishesh of unintelegent no-sence limited morons that have nothing to do all the day and are spaming the forumz around. So in their context I am just trying to stop their prulioned minds, becouse it looks like there are no admins and moderators that are filtering the garbage on the forum.


But Hypertransport 3.0 is real.
You can't deny it.

Woodcrest still cannot compete with Opterons in 4-way and up configuration is based on the dual FSB.
April 25, 2006 10:23:58 PM

Quote:
NO I can not give evidence.
I read the post with no links and I was thinking that is just another stupid SF story of how AMD will be good written by ??? and ????? or written by the AMD fanboy bloggers and rewriten by the same spamers with the stupid smile on their faces.
Like the story of 32nm K8s shipping in Q1 2007, the 4GHz 90nm K8, the Reverse HT, the Athlon 1.8GHz being better than 4 core 2GHz Woodcrest and other no-sence lies that people who don't have much IT knowledge are beliving.
This stupidities are pissing me off. Those idiots are ruining the point of existence of the forums. They should be place where people can learn something right, not the stupid wishesh of unintelegent no-sence limited morons that have nothing to do all the day and are spaming the forumz around. So in their context I am just trying to stop their prulioned minds, becouse it looks like there are no admins and moderators that are filtering the garbage on the forum.


But Hypertransport 3.0 is real.
You can't deny it.

Woodcrest still cannot compete with Opterons in 4-way and up configuration is based on the dual FSB.
I was not talking about Woodcrest. I was talkin about the stupidity of lieing stories.
HTT3 will be great when available, but it will be useless if there are no good chips that can use its benefits.
Woodcrest is 2 Conroes linked by the 1333MHz FSB, the same FSB that is used as connecton to the northbridge too. The cores on the Conroe chips have their L2 cache shared and direct connection providing 76GB/s to the L1 cache of the other core. Yes, the two dualcore chips are connected with slower link compared to the 1GHz links of the 4P Opteron server, but the Conroe chips are much more faster and are on one CPU(socket). The Woodcrest, the memory and the motherboard for that server will cost less than half of the 4P Opteron Server and we should wait to see which will be better. The adventage of the 4P Opteron Server is the high memory bandwidth that can be achieved through NUMA and the integrated DDR controlers and HTT links on each CPU. 4P Opteron server still will be the best for memory bandwidth hungry services, but the Woodcrest will be faster for the instruction hungry ones(considering that Conroe masacres AthlonFX-60). I don't have and I don't need 4-cores servers, therefore HTT, CSI or FSB exactly are meaning nothing to me and to the 99.9% of the discutants (including the concerned uneducated having no-point fanboys) here on the forum.
April 25, 2006 10:35:54 PM

Quote:
I don't know (or care) if it's been posted before.

Looks like the specs are similar to HT 2.0, which is great, it means ULTRA-LOW Latency while pushing bandwidth to new levels.

1.8GHz/2.0GHz/2.4GHz/2.6GHz

Provides up to 20.8GB/s per Duplex (41.6GB/s Effective) while maintaining Dual x8 or Dual x16 Links, as well as the possibility for Dual x32 Links, which means Inter-Device Coherency is brought to new levels and means that scalability to 32P Servers for sF 1207's will have the bandwidth they need and at the same time, have the same latencies or less than that of previous versions.

They've also added in the ability for 8B/10B Encoding, which will increase the integrity of the data transmitting as well as allow it to be transmitted farther then previously.

Also HotPlugging: The ability to add/remove HyperTransport Devices w/o disrupting system state.

~~Mad Mod Mike, pimpin' the world 1 rig at a time

Some people might consider it wierd that I both understand what you've said, and am actually feeling my mouth water (though not drooling; I manage to refrain from doing that) at the prospects.

I couldn't care what those people think. I can just hardly wait to see what form they'll try using to put it into desktop machines. (Obviously, workstations and smaller servers first)

Ever since I saw the HyperTransport-based slot, I personally thought it only a matter of time before HTT is developed to the point of being viable as a graphics card slot; after all, both ATi and nVidia are on-board the HTT consortum, and an HTT-based slot would provide bandwidth that PCI-express can hardly dream of, as well as even less latency than AGP.

Obviously, such a solution would be slow to appear, since it represents a much greater departure from the norm than any previous new graphics interface technology. It also shifts the work of adding support for multi-GPU setups to AMD, and away from ATi and nVidia. However, if such connection power is to be brought to the masses, there's hardly anywhere else it might be placed for good use.

Quote:
On the cpu hot swap, I think an electrical zap would be the least of your worries, Id wonder just how fast I could get a hsf on and running!

I think that because the distance that htt 3 can go, about 3 feet, the hot swap would be used more for things that are external to the mobo.

Maybe be able to have a cable that goes from one server to another and sort of "network" or daisy chain the 2 different mobos if need be, and its all pulg and play.

Keep in mind the difference between "plug and play" and "hot swappable." the former merely entails a simple setup; you shut down, you plug things in, you boot up, you're off and running. The latter is even better; you can skip the whole shutting down/booting up part. Pretty much all computer components are, to some extent, P&P.

Quote:
Nice find Mike. Now with HTT 3.0 AMD will choke intel to death with higher bandwidth and low latency. Intel should consider using HTT instead of making some buggy interconnect like CSI but they're to proud and arrogant.

I think, perhaps, some of the problems here might be arising from the fact that Intel is kinda left out of the HyperTransport IP loop. I don't think they could adopt it even if they TRIED.

Quote:
Are you giving the specs to AMD, so they have to make your wihes to become true or you are just spaming the forumz around?
any link of no fanboy blogger saying a little about your SF story?

BTW Intel CSI specs are above:
2048 GaAs 512bit optic links with 16TB/s bandwidth per link with 0ms access time. So HTT 3 sux a lot. Do you agree MMM or you have another better SF story?

Those specs, as I can clearly see, are just about 100% fictitious. To manage to get a total of 16TB/sec of bandwidth from a SINGLE 512BIT LINK, you'd have to have a frequency of 256GHz, which is pretty much imposible with today's (as well as the next few decades') technology. And an optic link with no latency?

Bah, this whole post, even if you DIDN'T use the word "sux," wasn't just covered in flags; the whole thing is red. This appears to be nothing more than a fanboy's post. Not necessarily an Intel Fanboy, buy just made by someone looking to troll and flame.
!