DailyTech - HyperTransport 3.0 Ratified Today

Submit News

Hardware HyperTransport 3.0 Ratified Today
Kristopher Kubicki (Blog) - April 24, 2006 12:45 PM

32 comment(s) - last by Slaimus.. on Apr 28 at 6:12 PM

The largest benefit of HT 3.0 is the near-double bandwidth increase

HyperTransport 2.0 bandwidth woes got you down? Say no more, HT 3.0 is here!

The HyperTransport Consortium just announced (PDF) the 3.0 revision of the HyperTransport interconnect. HyperTransport obtained instant fame when AMD picked up the bus protocol for the Athlon 64 and Opteron series processors. The bus is one of the industry's most open and fastest available already, but 3.0 adds dozens of new features and increased bandwidth.

The most apparent change for HT 3.0 is the bump in the data rate clock. HyperTransport 2.0 had a maximum clock of 1.4GHz; HT 3.0 increases that to 2.6GHz. This brings the total bandwidth available up to 20.8GBps. Additionally, HT 3.0 adds hot-plugging so devices can be inserted and removed from the HT layer on the fly. Power management and AC interconnect mode also played a large part in the newest standard -- HT 3.0 will now transmit up to one meter at the maximum specified clock speed with no signal loss. Effectively, HyperTransport can be used to connect from one machine to another in the correct conditions. DailyTech previously spoke to PathScale, a company focused on making high-speed InfiniBand interconnects. According to PathScale, AMD's HyperTransport interconnect provides greater overall bandwidth and scalability over anything else currently available.

Mario Cavalli, General Manager for the HyperTransport Consortium claims "The added performance and new features of HyperTransport 3.0 extend the applicability of HyperTransport Technology from chip-to-chip and board-to-board, all the way to chassis-to-chassis applications."

HT 3.0 also features an "un-ganging mode". Un-ganging simply means that the HyperTransport links can dynamically reconfigure during operation. For example, a single 1x16 HT link can be reconfigured as a 2x8 virtual HT link. The obvious use for something like this is a processor that could reconfigure itself to run SMT via two logical cores, each with its own HT link. After the SMT operation, the processor could configure itself back down to a single core.

HyperTransport 3.0 was only half of the HyperTransport Consortium's announcement today. Also announced was the HyperTransport HTX interface. HTX is a low voltage differential signaling link designed specifically for chassis-to-chassis HyperTransport interfaces. Essentially, the HTX connector allows the HyperTransport protocol to work over an external interface. One practical use for this may easily become external adaptable DRCs.

Comments

Threshold

Username
Password
remember me

This article is over a month old, voting and posting comments is disabled

am2?

By Quincunx on 4/24/2006 2:06:02 PM , Rating: 2

Perhaps delayed AM2 integration?

RE: am2?

By DigitalFreak on 4/24/2006 2:11:24 PM , Rating: 2

Uh... and what would have been the point of that? You're not even close to saturating the current 1Ghz/16 bit version of HT used in the Athlon64/Opteron.

Parent

RE: am2?

By DigitalFreak on 4/24/2006 2:12:20 PM , Rating: 1

Damn shitty comment software. Was supposed to be a response to the comment below.

Parent

RE: am2?

By Discord on 4/24/2006 2:39:50 PM , Rating: 2

LOL, your right. The present HT is sufficient for Socket 939. There are other features that make HT3 desirable though. Power management, improved integration, HT expansion slots (bye, bye PCIe (You could have used HT1 for expansion slots but they hammered out and added functionalities specifically with the HT3 spec)), HT based grid computing (Imagine that LAN...err I mean HT party! (But seriously imagine adding workstations/servers together in a near seamless grid by just plugging them in together with a cord. No more stupid rack servers) and tons of little things on top of that.
HT3 is cool stuff. Maybe cool enough that I would buy an AM2 based system if it had it.

Parent

RE: am2?

By Trisped on 4/24/2006 7:22:49 PM , Rating: 2

Not quite. The LAN emulation qualities you speak of are limited to computers within 1 meter of the hub. While that works great if you are trying to connect a few servers or racks together, you would actually be better off getting some Cat6 and running a 10Gb/s LAN with full duplexing and ATM interconnects that make 10Gb/s look like 10 year old tech.

quote:
But seriously imagine adding workstations/servers together in a near seamless grid by just plugging them in together with a cord. No more stupid rack servers

I believe you can do that with a LAN too.

Parent

RE: am2?

By mino on 4/24/2006 9:30:00 PM , Rating: 2

Yes, you can.

And the performance will tell, :-)

Pathscale's Infinipath is 2x faster than IB not to speak on 10Gb.

Maybe 100Gb would do, on Myrinet mx maybe. But it would be close.
On TCP/IP ? Not a chance. It is like comparing 32b/33MHz PCI to PCIE 16x.

cheers

Parent

RE: am2?

By mino on 4/24/2006 9:32:01 PM , Rating: 2

meant for Tristed

Parent

RE: am2?

By peternelson on 4/25/2006 3:34:27 AM , Rating: 2

Maybe so, but you need a very special lan like Myrinet to deliver near the exceedingly low latency that hypertransport interconnect has. Plus HT lets you treat the whole system memory as one space (single system image), whereas lan based system needs to map addresses on to the address space of each node (obviously timewasting conversion). I look forward to reading the HT3 spec.

Parent

RE: am2?

By Discord on 4/24/06, Rating: 0

RE: am2?

By meson2000 on 4/24/2006 4:27:36 PM , Rating: 2

Not in the greater than 2 processor arena. Intel is still sticking with their obsolete front side bus. So Intel's NGMA won't be able to touch AMD in performance with greater than 2 CPUS. HT3 will only extend that lead for AMD when it gets put in actual products....

Parent

RE: am2?

By Trisped on 4/24/2006 7:28:52 PM , Rating: 2

AMD seems to do one thing at a time. Now it is DDR2 support, then they will optimize it. After that maybe HT3 followed by Quad-core.

Parent

AMD

By Discord on 4/24/2006 2:07:29 PM , Rating: 2

Well this will definitely keep AMD ahead of the bandwidth curve for a long time.
Too bad that the common consumer won't benefit from HT3 until AMD's next gen desktop CPU comes out. Unless you chip in for that server workstation...
Stupid short-term bridge gap AM2 product release with crappy ass DDRII memory. Oh well, I guess they couldn't wait for DDR3 forever. They could have at least put a newer version of HT to make a little tempting.

RE: AMD

By Furen on 4/24/2006 6:20:56 PM , Rating: 2

Tempting in what way? Having lots of unused bandwith available? HT3 makes sense for socket F because the HT link IS a bottle neck in 4-way+ servers. HTX sounds nice and all but we'll never see industry-wide support for it because Intel is not backing it.

Parent

RE: AMD

By mino on 4/24/2006 9:24:14 PM , Rating: 2

In the 4P+ and cluster/HPC space it does not matter if Intel backs it or not. AMD backs it, and AMD is the leader here. (At least until Q2 2007)
Also HTX is becomming industry standart for HT-based systems(not only AMD) allready. Even while the final draft was not approved systems are around, and selling!

Parent

RE: AMD

By peternelson on 4/25/2006 3:43:51 AM , Rating: 2

HTX was developed as a separate standard BEFORE HT3 (downloadable from the HT consortium page). It was rolled out by IWill on their Opteron motherboard (incidentally they had a lot to do with steering HTX slot specification). I'd say why stop at one HTX slot. If you had two then you could have a kind of daisy chain of hypertransport products eg two motherboards, some storage (yes ramdrive on the hypertransport!) and maybe some Xilinx acceleration on another board.

I think it's good and should be used more (where bandwidth and latency needs justify).

I think you'll find Intel aren't barred from joining the club. It's very cheap with no volume based royalties. Just they would have to design in or bolt on HT to their chips (basically a redesign for them). In a way it would be nice if Intel did adopt it then HT would really take off. In the cluster and server markets it's a great help and will give AMD a continuing advantage.

Parent

RE: AMD

By ceefka on 4/26/2006 5:53:26 AM , Rating: 2

quote:
some storage (yes ramdrive on the hypertransport!) and maybe some Xilinx acceleration on another board.

Someday after AM2 on a HTX slot in your desktop too.

Parent

If this is what I think it is...

By dunno99 on 4/24/2006 3:45:20 PM , Rating: 2

Oh. My. God. This seems like you can hotswap your CPU if it fails! That will basically mean that you don't need specially configured blades that can do hotswapping, reducing costs of multi-CPU servers/supercomputers tremendously. And the dynamic HT splitting for SMT sounds like part of the plan for that reverse hyperthreading thing. Wow!

RE: If this is what I think it is...

By hwhacker on 4/24/2006 4:12:46 PM , Rating: 2

It does that way, about the reverse hyperthreading. It also sounds like a hint of allowance for co-processors, etc.

The bandwidth allowance will surely be helpful come DDR3. For now, the spec of 1.4ghz hasn't exactly been breeched, nor is it really needed yet, but come quad cores it will be...and no doubt this will be implemented in whatever comes after AM2. Quad cores, DDR3, co-processor allowance, and the bus to feed it all. It surely sounds like the products in AMD's future are at least starting to shape up, at least based on the spec of HTT3 and the fact it's completed. This surely is a hint at things to come.

Parent

RE: If this is what I think it is...

By Griswold on 4/25/2006 11:50:34 AM , Rating: 2

Well it seems that AMD is once again thinking way ahead. Go back to the design phase of the K8 CPU - it had multicore in mind when nobody was even thinking about multicores on the desktop/server x86 front.

The pieces fall into place without spreading too much FUD.

Not to say Intel doesnt plan ahead, but what we hear from them and what we get is often not the same - speaking of longterm technologies.

Parent

RE: If this is what I think it is...

By Discord on 4/24/2006 4:18:07 PM , Rating: 2

You could theoretically hot swap your CPU in a multi socket/CPU system. The process of opening up a case and physically removing a CPU without powering down the system would make this unadvisable to say the least. Not to mention you have a good 50-70 watts flowing through those babies... You can try it is you want to though.

Parent

RE: If this is what I think it is...

By Questar on 4/24/2006 6:34:13 PM , Rating: 2

Dude, many, many servers already have hot swap cpu's.

Hell, the good stuff has hot swap everything.

Parent

Release date?

By Haggar on 4/24/2006 1:47:11 PM , Rating: 2

So when will this be released?

RE: Release date?

By bhigh on 4/24/2006 1:59:48 PM , Rating: 2

It was "released" today, that's the point of the article.

If you mean, "When will products using the new version of HyperTransport be available?" then it depends on how long it takes AMD, nVidia, ATI and others to implement it. AMD and one or more partners probably have something in the lab already, but probably nowhere near production ready.

Parent

RE: Release date?

By Discord on 4/24/2006 2:09:43 PM , Rating: 3

You will see HT3 in AMD's new S socket servers.

Parent

Useful for?

By TheLiberalTruth on 4/24/2006 1:56:38 PM , Rating: 4

So this will be most useful for systems with many, many processors...like in servers, right?

RE: Useful for?

By bhigh on 4/24/2006 2:03:23 PM , Rating: 3

It will be useful in multi-core systems, especially as the number of cores per socket increases. While HTT is fine for dual-core, a 4-core chip may require more bandwidth.

The external connections will also allow building a multi-cpu system that has more than one chassis. Imagine plugging two 2-socket servers together and having one system. Now do that 1000 times more and have a supercomputer.

Parent

Some parts sound like hyperthreading

By hstewarth on 4/24/06, Rating: 0

RE: Some parts sound like hyperthreading

By Wwhat on 4/25/2006 8:49:56 AM , Rating: 2

There was a rumor that AMD was working on a sort of reverse hyperthreading, that is making 2 cores seem like one to the system so that you can run old code at increased speed, perhaps this ties into that?

Parent

RE: Some parts sound like hyperthreading

By saratoga on 4/26/2006 5:40:49 PM , Rating: 2

Yeah and that rumor was laughed at by everyone who'd done parallel programming.

Splitting up single threaded code into multiple threads in real time has been the holy grail of parallel processing for decades. Unfortunately, its essentially impossible within the current model of how computers work.

Parent

AMD strikes back ;-)

By peternelson on 4/25/2006 3:29:24 AM , Rating: 2

I've noticed quite a few people saying AMD will be Owned by Conroe coming out.

Well, Hypertransport is an excellent standard, and with this it just took a major speed jump. Of course to really take advantage of that bandwidth you needs multicpu and/or multicore. Guess where the industry is going?

You wondered what could be AMD's response to Conroe? Faster chips of course. Bring on the HT3 enabled Opterons ;-)

Intel could do itself a big favour by adopting Hypertransport too, it's very cheap to join the consortium.

But then they would have to admit they got things wrong by pushing their bottlenecked FSB as an alternative.

RE: AMD strikes back ;-)

By peternelson on 4/25/2006 3:49:00 AM , Rating: 2

AMD have been quiet about Socket F recently.

Perhaps now the HT3 spec is announced and in the open, then they might say something like "All our multicpu opteron systems on socket F will have HT3 giving these features....."

So if they were waiting for something, this is it.

Keep watching the news to see what AMD have to say.

Parent

Much Needed

By Slaimus on 4/28/2006 6:12:31 PM , Rating: 2

This is a much needed speed boost for AMD. Current Opterons do not scale well from 4-way to 8-way, because the HT links are being saturated with cache coherency data at 8-way. The need this speed boost to really allow 8-way dual-core Opterons to not be bottlenecked.