Monday, April 24, 2006

HT 3.0 Spec Released

AMD led HyperTransport Consortium just released the HT 3.0 standard. The new features include:

1) 2.6GHZ clock, up to 20.8GB/s (166.4 Gb/s) per link bandwidth
2) AC mode allows links up to 1 meter length for chassis to chassis connections.
3) Auto link splitting between 2x8 or 1x16 links
4) Hot plugging

The extended HT link length allows one to connect multiple individual Opteron servers into higher order SMP boxes. Fujitsu-Siemens have a 8 way Opteron server which is made of four tighly packed 2P Opteron blades connected via HT. Now, one can have Opteron servers farther away forming larger SMP machines. The additional latency due to 1 meter wire length is less than 4ns. A 32P SMP machine can be built by linking 8 four way 1U servers. Since you can hot plug the HT links, you can even add and remove servers (CPUs) on the fly. This is like a cluster, but you are building a true SMP box, running a single instance of a standard SMP OS within a single memory space. This will revolutionize server computing! SUN and IBM better hurry and do more Opteron.

HT 3.0 is 100% backward compatible with previous generations of HT, which means an old NF4 chipset can be hooked onto a newer Opteron with HT 3.0, even though it's less efficient.

I expect AMD's Rev F Opterons to utilize HT 3.0, and Broadcom to supply a lot of the server chipsets (HT bridges with a whole bunch of features integrated).

Currently, an Opteron CPU has three 1GHZ HT links, all of which can be coherent (Opteron 8xx). A coherent HT link can be used for I/O or IPC, but a non-coherent link can only be used for I/O. A next generation socket 1207 Opteron may have more HT links, coupled with the increase in bandwidth, the next generation Opteron servers will be super strong. Even assmuming a Rev F Opteron has three HT links and one DDR2 800MHZ memory controller, the total bandwidth for a 1P Opteron will be 3* 20.8 + 12.8 = 75.2 GB/s. The total bandwidth for a 2P opteron will be 4*20.8 + 2 * 12.8 = 108.8 GB/s.

In comparison, a 1P Woodcrest will have a total bandwidth of 10.2GB/s with a 1333MHZ FSB (64 bit), not enough to pump DDR2 800MHZ (128 bit).


Blogger Mansoor said...

I would like to know your comments on this :

12:28 AM, April 25, 2006  
Blogger Sharikou, Ph. D said...

There were some discussion on co-processors at here.

In my opinion, both coherent and non-coherent HT will revolutionize computing. ccHT can bring multiple boxes into one big SMP box, and HT can handle the I/O bandwidth requirements. Imagine you hook 8 opteron opteron boxes to form a single SMP box, what about the total disk bandwidth? HT can handle them all.

10:05 AM, April 25, 2006  
Anonymous Anonymous said...

HyperTransport DirectPackets Data Streaming
(peer-to-peer, *16 virtual channels*, native packet handling)

HyperTransport technology has a daisy-chain topology, giving the opportunity to connect multiple HyperTransport input/output bridges to a single channel.

HyperTransport technology is designed to support *up to 32 devices per channel* and can mix and match components with different link widths and speeds.

3:15 PM, April 25, 2006  

Post a Comment

Links to this post:

Create a Link

<< Home