Tuesday, March 22, 2005

Best supercomputing quote ever

Apple has put up a new description of Virginia Tech's System X supercomputer. In it, Dr. Kevin Shinpaugh, Director of Research and Cluster Computing, has this to say about System X:
The new Xserve cluster, says Dr. Shinpaugh, is “a very beautiful-looking machine, a combination of black racks and brushed aluminum, finished very nicely. Then you have the nice blue blinky lights. A lot of people who come by like to see blinky lights so they can say, ‘Oh, there is something going on.’
:)

Sunday, March 20, 2005

IBM's 970MP: Top 10 things to know about it

Even though there is now all this 970MP hoopla, it seems that many people don't really have an idea of what it really is. Rumours have been all over the map about it recently, but to the best of our knowledge it seems that the Think Secret article that broke the news about the 970MP a year ago is probably the most accurate.

In this article I will outline the top ten things (in no particular order) we should know about this chip:

1) As the name would suggest, it is a 970 series chip, but dual-core. In other words, it is a dual-core chip that shares the POWER4 heritage of the 970FX. It is not a POWER5 derivative, nor is it a derivative of Cell's PPE.

2) It will not support simultaneous multithreading. If it had been a POWER5 derivative, this feature might have appeared, since SMT is a feature of the POWER5.

3) It will likely not have an integrated memory controller. An integrated memory controller would have been a bonus, but the information available makes no mention of an IMC. (The POWER5 has an integrated memory controller, but the POWER4 does not.)

4) It will likely add support for partitioning. This is a feature absent on the 970FX, but which is present in the POWER4. Don't ask me to explain it though, since I'm a bit fuzzy on the concept myself, especially in terms of how it would fit in with Apple's products.

5) Think Secret said it will have 1 MB L2 cache per core. This makes sense as the cache upgrade is a natural evolution for the 970 series chips. At higher clock speeds, it will be more necessary too. There is no information available about a shared cache, but it's likely the caches will be independent, and not a combined shared 2 MB cache for both cores.

There is still no official confirmation of the 970GX (although one article at IBM, but not written by IBM, did reference this chip). The 970GX is the single-core variant 970MP (with the increased L2 cache), and such a chip would be perfect for the iMac and the PowerBook, and possibly also for a low end single-CPU Power Mac. One wonders if the 970GX could simply be a 970MP where one of the cores is defective and has been deactivated (or even completely removed). Recycling partially defective 970MP chips as single-core chips would reduce overall costs.

6) The 970MP is fabricated on IBM's 90 nm CMOS SOI10K process, and has a die size of 153.8 mm² . For comparison, Intel's dual-core Pentium D Smithfield will be 206 mm², or approximately one-third larger. The current G5 970FX is 66 mm².

7) We don't know about on-chip communication speeds between the two cores of the 970MP. Some dual-core chips to be released this year depend on the FSB for the cores to communicate with each other, but this is not ideal. Preferably the 970MP would have higher on-chip communication speeds between the two cores, but we have no information either way.

[Update 2005-03-21] There is however, this statement from the original 970MP rumour: "The IBM PowerPC 970MP incorporates two complete microprocessors on a single chip, along with some common logic to connect these microprocessors to a system." This doesn't really answer our question, but it certainly does not confirm that the two cores can communicate with each other at higher than FSB speeds. [/Update]

8) We don't know the 970MP's clock speed, but we can make some guesses. Right now Apple states that the 2.3 GHz G5 in the Xserve runs at a maximum of 55 Watts. That is an impressively low amount of power consumption for a chip with its level of performance. (For comparison, the 90 nm Intel Pentium 4 Prescott chips run at around 130 Watts.) There are also hints that the 2.4 GHz G5 might run at 60 Watts. Momentum's 970FX evaluation motherboard is listed to support G5s up to 2.4 GHz, and in their board block diagram it lists a CPU power specification of 60 Watts. We don't know the power specification of the 2.5 GHz G5, but we could guess that it would be in the range of 65ish Watts.

Thus, we can expect the dual-core G5 970MP to debut with a clock speed of at least 2.5 GHz. One can even hope that IBM has got things working well enough to make a 2.8-3.0 GHz dual-core release possible. The information provided by Think Secret previously was that a 3 GHz dual-core (and a 3.0 GHz single-core) release was the goal, and we already have Steve Jobs' 3 GHz promise back from WWDC 2003.

9) We don't know what the bus speed will be in Macs. Think Secret suggests a 3:1 divider, which would mean a 1 GHz bus for a 3 GHz chip. However, it would also mean something like a 733 MHz bus for say a low end 2.2 GHz dual-core Power Mac. This is slower than the current low end 1.8 GHz dual Power Mac, but for most usage it may not be a major issue, since memory speed would not be significantly higher with DDR400, even with dual-channel memory. (As you may suspect, the memory type is not dependent on the 970MP at all. Like the choice of AGP vs. PCIe, the choice of memory used in the next Power Mac is up to Apple. So why not DDR2? It's possible the next Power Mac could get DDR2 support, but DDR2-533 offers little overall performance advantage over DDR400, and DDR2-667 is still quite expensive.)

10) The 970MP is designed to support up to a 4-way configuration, which would support the previous (dual dual-core) quad Power Mac rumours.

Interesting times ahead for Apple...


[Update 2005-03-21]

How could I forget? Let's just add one more.

11) IBM's leaked 970MP article references the Momentum G5 970FX evaluation board mentioned above (#8), and states that it was used to test the 970MP's thermal diode.
Using the PowerPC970FX evaluation board as an example, a Cypress Semiconductor PSOCTM (PSOC), P/N CY8C27243, is chosen to sample the thermal diode’s voltage, convert the voltage to a numerical value, and send the information to the service processor for disposition. The PSOC contains configurable analog and digital blocks to create the needed A/D converter and other analog functions.
This suggests that the 970MP might be pin compatible with the 970FX, although we do not know for sure.