New Itanium More Powerful, Power Efficient

Heir Of The Mess writes "Intel have a press release out about their new Itanium 2 Processor. The new processor doubles the performance, and improves performance per watt by 2.5 times compared to the existing single-core versions. The flagship model triples the cache and can execute 4 threads/instructions per processor enhanced by Hyper-Threading. Transistor count is a whopping 1.7 billion. Triples the previous SPEC_int_rate_base_2000 record. Retails for US$3692 for the top of the range.
So yes the Itanium crew are still pushing forward. I wonder if this could help save SGI?"

Comments

Working with the hardware directly

[drachenstern]

I can tell you that the processors are not even in production at the main server manufacturer backing the Itanium family (ie, not Intel, the other one*).

We have seen a few proto style units roll through, but they have all had serious problems and are not running at full speed. The engineering group either cannot or will not give us a reason why these units are running crippled, but we believe it to be a chipset issue. Hopefully we will see the servers rolling through our manufacturing process within the next 60-90 days, but no management timelines have been released.

Here's to all those of us who want that raw power and are looking to pay for it!!!

*Won't disclose the name since I don't know if this info violates my NDA, but screw it, the public should know this stuff. Information does deserve to be known.

-drach!

Re:Working with the hardware directly

[Anonymous Coward]

They are having major problems with flaky frequency shifting. The processor uses an on-silicon ammeter to control its internal clock frequency. So if you are running code that would cause its power usage to go above the rated wattage it would secretly slow the frequency to prevent that from happening. This analog method is a crapy method of power control. It requires almost no RTL and architechture changes, but tends to be falky, limits performance in unknown ways, and causes poor silicon yield. I have a feeling some grad student is going to bust them down the road by doing a detailed performance analysis and finding that while Montecito does well on SPEC, it is not too difficult to generate code that will make the chip perform very poorly since the frequency limiter kicked in.

Re:Triple the cache

[tomstdenis]

24MiB of 6T SRAM is 1,207,959,552 transistors just for the cells. Not including tags, address decoders, etc...

By comparison, an Opteron uses 113,246,208 transistors for the 2048+256KB of cache [assuming they use a 6T structure which I don't know for a fact since I'm not privy to the details and technically I couldn't say even if I were, so don't assume what I said is verbatim fact, yada, I hate disclaimers] and the 4MB Duo (total of 4096+128KB of cache) needs 207,618,048 transistors for its cache.

Tom

Re:Price

[LWATCDR]

"While x64 may be better-suited for web serving, part of the reason x64 seems to be better than ia64 on most other workloads is that you may not have the OS and software optimized for the ia64 architecture"
Or it could be the ia64 really isn't well suited to web serving.
The ia64 has very good double precision floating point. And seems to be very good at minimal branching / very predictable branching code. In other words not web serving. The ia64 seems to have the same problem that they had with the i860 in fact if you read the wikipedia entry for it it sounds like the exact same chip!
"One fairly unusual feature of the i860 was that the pipelines into the functional units were program-accessible, requiring the compilers to carefully order instructions in the object code to keep the pipelines filled."
and this seems to be even more on the mark "As a result of its architecture, the i860 could run certain graphics and floating point algorithms with exceptionally high speed, but its performance in general-purpose applications suffered and it was difficult to program efficiently (see below)."

My goodness it is hard to believe that Intel would go down this road again. I guess if at first you don't succeed try, try, again.