Posts by UK Jim 16 publicly visible posts • joined 8 Mar 2012
Glad you said "Intel and Arm are to a certain extent rivals", since that may be true, even though they are not direct competitors.
Consider this beer analogy:
Think of Intel as a brewery; they have their beer recipes and they sell beer but not the recipes, while Arm, sells beer recipes, but doesn't sell beer.
Therefore the two companies are in different (but, admittedly, closely related) markets. If you are thirsty, and need a pint, going to Arm won't help, since if you want a pint of Arm beer you can't buy it from them, but have to go to one of the companies that brews using their recipe. Similarly if you want Intel's recipe, tough luck, they aren't selling it.
Clearly, if everyone decides they like Arm's recipe, that is bad for Intel (just as if everyone likes Intel's recipe that is bad for Arm), but they are not direct competitors. Intel's competitors are the companies enabled by Arm (Ampere. Broadcom, Marvel, ...) not Arm itself
"What sunk Inmos was their building their own chips, instead of going fabless (like ARM effectively did)". Maybe if Inmos had started ten years later this would be true, but the whole initial sales pitch of Inmos was to exploit the step change in manufacturing with the the shift to 10 micron. Inmos achieved that successfully, and were ahead of the world in manufacturing for a few years. (Hence the fact that the Inmos SRAM was dominant and used as the main upgrade between a Cray-1 and a Cray-1S).
What sank Inmos was not the Transputer, but the Thatcher government which hated anything associated with the National Enterprise Board, so dumped Inmos to Thorn-EMI who had insufficient capital, so could not make the required level of investment to grow the company and push forward the manufacturing. Before that sale, Inmos had exceeded its growth plans.
Note, too, that the Transputer was the first processor to include on die memory, which would not have been possible without the Inmos manufacturing ability.
"The certainty of the fab going obsolete" only follows from the lack of investment which brings us back to the decisions of the milk snatcher.
Although the Transputer may have been the most visible Inmos product (exactly because it took the most work to use!), it was not Inmos only (or, maybe even main) product. We manufactured memories (both SRAM, and non-volatile) and a "colour look-up table" (a RAM with onboard D/A converters to generate video signals).
TBH I don't care how much AI power it has. There is a well defined measure of High Performance Computing performance (which we hate, but there it is), and if you claim to be "Britain's most powerful publicly known supercomputer", and then quote a position on the Top500 and a Linpack performance number then you are explicitly using that definition. (Which, here, does not support the claim of being the UK's top machine).
If you want to compare "AI power", then that's fine, you;'re very welcome to do that, but whatever you claim there is not comparable with any measure used to evaluate and rate the top supercomputers.
If there was an "AI500" list and recognised AI benchmark to use to rate machines then claiming a position in that is fine. But the claim that performance on FP16 (or BFP16) is comparable with double-precision Linpack is just wrong.
Or, if you prefer: you are absolutely right, NVIDIA is making an apples to tangerines comparison, and that is what I am objecting to!
"The next SuperPOD project is the Cambridge-1 behemoth, planned to be Britain's most powerful publicly known supercomputer"
Umm, no:
"ARCHER2 will have a peak performance estimated at 28 x 10**15 FLOP/s" (https://epsrc.ukri.org/blog/supercomputers-how-archer2-will-increase-the-pace-and-productivity-of-research/#:~:text=ARCHER2%20will%20have%20a%20peak,than%20the%201964%20Cray%20supercomputer.)
vs "eight petaflops of Linpack benchmark performance" (your article).
So Archer 2 is 3.5x more powerful on a the disliked, but standard, measure of HPC machine performance. (And, Archer 2 is "publicly known")
"and in the process bump the “Titan” machine at the USA's Oak Ridge National Laboratory off the podium for just the third time in the history of the TOP500 list of the world's mightiest supercomputers."
Really!? This is the third time that Titan has dropped out of the top 3?
I think what you were trying to say is "and in the process made this only the third time in the history of the list when the USA has no machine in the top 3".
While you can force vectorization of that code (as you do by using compiler flags), doing so is not, in general, safe. Consider an invocation of vectorize_this in which a and either b or c point into the same array. (E.g. vectorize_this(&b[1], b, c);). There is now a loop carried dependence and the results generated by the vector code will be different from those generated by the scalar code.
If you know that the code is used without such overlaps, then the right answer is to modify the code and use the "restrict" qualifier on the arguments to inform compiler of that fact. Though then, of course, you can't claim not to have to modify the code!
(FWIW I work for Intel, and this is an issue for everyone...)
Intel acquired the Infiniband assets of Qlogic about a year ago...
http://newsroom.intel.com/community/intel_newsroom/blog/2012/01/23/intel-takes-key-step-in-accelerating-high-performance-computing-with-infiniband-acquisition
(FWIW I work for Intel, but I do not speak for them :-))
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2024
