5 posts • joined Wednesday 4th April 2007 22:57 GMT
AMD Marketing Horrible
AMD has been marketing the ghost of Opteron future (Barcelona) for about six months. It needs to sell what is on its truck.
Opteron is pretty good against Woodcrest today. In many areas Opteron performance is solid or better than Woodcrest. There are others where Woodcrest has no or limited benchmark results, such as TPC-H and SPECjAppServer.
Opteron also outperforms Woodcrest as a VMware virtualization platform. See:
Clovertown only provides a 1.3X to 1.5X improvement on many workloads, the exception being workloads which are highly threaded and can reside in cache, like SPECjbb. Most customers I deal with are not buying Clovertowns, but are buying Woodcrest over Opteron.
AMD should never have lost one point of market share to Woodcrest in the HPC market. Opteron floating point performance is solid compared to Woodcrest, Barcelona, with four floats per clock, will be stellar for HPC, and the whole "hybrid computing" concept behind the ATI acquisition is well-suited for HPC. But AMD just completely dropped the ball.
AMD could have easily pidgenholed Clovertown into a niche, and positioned it as the equivalent of the Pentium Ds of a few years ago. Then it could have picked its fight with Woodcrest on its terms. It could have clearly attacked Intel's Tulsa XeonMP chips, especially in the VMware space.
Instead, AMD chose to sell futures. The problem with selling futures is they don't generate revenue. And now AMD has to deal with all of the negative press due to its financial performance.
Comparing quad core to quad core (the only fair comparison, because we don't know how the dual-core Conroe would compare to the QX6800 on these apps).
25% more FSB bandwidth (1333 MHz vs. 1066 MHz)
13.7% more clock rate (3.33 GHz vs. 2.93 GHz)
New SSE4 instructions
Differences in cache sizes are not known.
Except for one app (DivX 6), the typical speedup is 25%. I don't count the 3D Mark Pro Overall either, because it is probably very dependent on the graphics card.
So Intel is getting about 10% performance beyond clock. I wonder how much of this is the faster FSB, how much may be due to cache size differences, how much is SSE4, and how much are other microarchitectural improvements.
I would like to see benchmarks Penryn dual-core vs. the fastest dual-core Conroe.
I don't know how much more Intel will improve the performance, or if they will offer higher clock rates than 3.33 GHz, but in its Quad-Core form Penryn does not appear to offer any significant advantage over Barcelona.
I expect Intel will offer Penryn Xeons with very large caches to maximize performance.
2008 will be the year of the great quad-core parity.
Get ready for the price war.
Given its a 144 port switch, it probably uses a 24-port switch ASIC. Fulcrum has a 24-port 10GE switch ASIC. Broadcom has a 20-port ASIC.
The Fulcrum ASIC is also used in Force 10's 24-port 10GE switch.
Barcelona will crush Clovertown in Floating Point
So Intel's response to the 3.0 GHz dual-core Opteron's floating point performance is to point out the quad-core Clovertown beats the dual-core Opteron by 9%?
Barcelona will be capable of four floating point operations per cycle, and expects to get about 50% better FP performance per core at the same clock rate as current Opterons. That suggests Barcelona will more than double Clovertown's SPECfp_rate performance, while likely equaling or slightly exceeding Clovertown's SPECint_rate performance.
I can see why Intel is defensive. They basically have a minimally engineered, quad-core "marchitecture". When faced with the lemon of the Front Side Bus, they made Clovertown lemonade.
I still think most of Intel's recent sales are of dual-core Woodcrest parts (and Woodcrest is a very good chip, mind you), but Intel chose to make the thought leadership debate about quad-core, and that is about to come back and bite them.
In HPC clusters at least, Barcelona will trump Clovertown.
What would really be interesting is if AMD "Clovertowned" Barcelona into an 8-core/die module.