Ahead in many fronts, AMD is still playing catch up to Intel on hyper-threading.
Some of you have raised the issue on why AMD has not adopted hyper-threading to increase performance of its chips in highly threaded environment such as virtualization.
Recently a TechPulse360 reader commented on why hyper-threading made sense:
- Hyperthreading gives tremendous boost in certain applications. 10-15% is towards the lower end of the scale;
- All modern CPU architectures (except AMD) support multiple threads per core. Look at POWER from IBM, T1/T2/Rock from Sun Micro, Nehalem/Atom/Larrabee from Intel …;
- A proper implementation of simultaneous multithreading (SMT, and what Intel calls Hyper-Threading) requires few additional resources. There is no such thing as a “Normal Pipeline” and a “Hyperthreaded pipeline” – all the functional units that form the bulk of the pipleline are unchanged, certain resources are shared, while certain other resources (like ISA registers) are duplicated.
This week an AMD engineer confided to me that not having hyper-threading available made Opteron look slower than Intel’s low-end chips. The engineer also said that people at AMD have now admitted that not having hyper-threading was the wrong technical choice.
So here’s what Pat Patla, AMD’s server boss had to say when I asked him about hyper-threading during my visit at AMD’s Sunnyvale, Calif.- headquarters this week:
“If you look at our future roadmap and what we’re showing for adressing the threaded market, we believe it is best addressed at full core count this time. And you saw our 2010 time frame when we are talking about 12 cores per CPU and in 2011 with 16 cores per CPU. So we think we are pretty well covered in the 48 to 64 threads environment for the next couple years and we’ll see what 2012 and 2013 brings.”
It sounds to me that AMD’s 2012 chips are going to have hyper-threading!
Here’s a video excerpt of Pat Patla answers on the hyper-threading question:
Posted by TechPulse 360 
