Elbrus E2K Speculations
by Alexei Pylkin, Institute of Computational Mathematics and MathematicalGeophisics, Russian Academy of Science
The promising microprocessor Elbrus E2K (Elbrus-2000) developed by a Russian company called Elbrusbecame well-known not so long ago, namely after the article "The Russians Are Coming" by KeithDiefendorff, which had been published in the February edition of the leading technical publication"Microprocessor Report".
The main theses of the article sound as follows:
- The new processor is expected to perform from three to five times faster than Merced processorfrom Intel. E2K will require lower power, and its manufacturing will appear considerably cheaper thanthat of Merced;
- The Elbrus team has the credentials and rich experience to be able to introduce such a processor;
- Elbrus computers have been developed basing on the main principals of the modern progressivemicroprocessor architectures, long before the similar ideas appeared in the West;
- Elbrus-3 was developed in 1991 with the technology, which was already outdated even in thosedays. However, its performance was twice as great as that by a famous Cray;
- The technological novelties implied in the E2K are just as impressive as its architecturalinnovations: E2K boasts a much cooler electronic scheme than that used in the recent record-holder -Alpha 21264 processor;
- CMOS technology used in E2K with lower power demands will very soon provide the world with a pocketsupercomputer;
- As for compilers, the Elbrus team proved highly qualified here as well as in architectures developingand electronic projecting: at present the E2K parallelizing compiler achieves up to 10 instructions perclock cycle, which is almost three times as much as the best existing Alpha compiler manages;
- In the interpretation regime E2K executes the microprocessor code of Intel and Sun only 10-30% slowerthan its own one. Just for a better comparison: FX!32 program, which executes Intel processors code onAlpha computers, sometimes happens to slow down the applications by three times; while E2K allows gettingthe full 100% binary compatibility for any codes of Intel processors and for any operation system, whichis of course to the advantage of E2K compared to FX!32;
- An important feature of E2K is the "impregnable" virus protection of the data and codes being executed.The development in this direction in the West has been ceased together with the failure of Intel 432microprocessor.
The mass media all over the world have reported about the development in Russia of a new processor,which will potentially surpass Merced due to its much more promising features. The interest and wilddiscussions on this matter in the Internet keep on growing. A great number of famous westernmicroprocessor experts rank Elbrus E2K very high. According to them, the Russian microprocessoris a fount of exclusive ideas with great future.
In the beginning of this year Elbrus finished the logical projection of its processor and completedthe system software. The exact model allowed them to get all the performance estimates. And now they arefacing the final and the most important part of their work: the realization of all their ideas in silicon.
The present article tells a bit about the Elbrus family history, about the Russian priorities amongsome architectural solutions widely utilized in today's microprocessors. You will probably agree with methat historical investigations may sometimes help to reveal the future… So, let's go.
The history of the world computer science is connected with the name of Elbrus. This company wasfounded in Lebedev Institute of Precision Mechanics and Computing Equipment, which team had beendeveloping supercomputers for the Soviet Union's defense establishments for over 40 years. E2Kprocessor embodies the developing ideas of the Russian supercomputer Elbrus-3 built in 1991.Today Elbrus-3 architecture is referred to EPIC (Explicitly Parallel Instruction Computing).
According to Boris A. Babaian, chief architect of Elbrus supercomputers, superscalar architecturewas invented in Russia. To quote him as saying: "In 1978 we developed the world's first superscalarcomputer, Elbrus-1. At present all Western superscalar processors have just the same architecture.First Western superscalar processor appeared in 1992 while ours - in 1978. Moreover, our variant ofsuperscalar is analogous to Pentium Pro introduced by Intel in 1995".
The historical priority of Elbrus is confirmed in the States as well. According to the same articlein Microprocessor Report by Keith Diefendorff, the developer of Motorola 88110 - one of the first westernsuperscalar processors: "In 1978 almost 15 years ahead of Western superscalar processors, Elbrus implementeda two-issue out-of-order processor with register renaming and speculative execution".
In 1991 Peter Rosenbladt, the Hewlett-Packard executive, was in Moscow visiting Elbrus (by this timeit was Institute of Precision Mechanics and Computing Equipment) and received exhaustive documentationon Elbrus-3. A bit later it turned out that at the same time HP started a project together with Intel,which resulted into a joint development of EPIC-processor, aka Merced. Its architecture is very similarto that of Elbrus-3, and all the differences deal mostly with the simplifications introduced by Intel intheir processor.
As to Boris Babaian, Peter Rosenbladt offered partnership with HP. But Babaian made up its mind tojoin Sun (the first meeting with Sun administration had taken place in 1989). And so, in 1991 the contractwith Sun was signed. According to Sun official reps, Elbrus took part in the development of UltraSPARCmicroprocessor, optimizing compilers, operation systems (including Solaris), Java tools, multimedialibraries.
From the very beginning the E2K project was financed by Sun. And now it is fully independent, allintellectual resources belong to Elbrus and over 70 USA patents guarantee this property. Boris Babaianexplained: "If we had kept on working with Sun a bit longer, everything could have belonged to it.Although 90% of the work had been done long before Sun appeared".
Between 1992 and 1995 a famous microprocessor architect Dave Ditzel worked together with Elbrus in Sun.Boris Babaian tells: "And then Dave founded his own company - Transmeta, and started working on anew device very similar to our one. We still keep in touch with Ditzel. And he, on his part, alsofeels like being our partner". Unfortunately, there isn't much information about the future productfrom Transmeta yet. All we know is that it will be a VLIW/EPIC microprocessor with low power and x86binary compatibility provided by object code dynamic translation.
And not so long ago I found out that Elbrus developments have very much to do with the most widelyspread universal microprocessors - x86 Intel family.
Russian supercomputer know-hows are carried out in Pentium processors. Former Institute of PrecisionMechanics and Computing Equipment employee Vladimir Pentkovski is Intel leading processor architect.Pentkovski carried to Intel huge experience and progressive technologies developed at the Institutewhere he had worked. According to microprocessor expert Keith Diefendorff, "Elbrus has developedcomputers based on superscalar, shared memory multiprocessing, and EPIC techniques, long beforepapers on those subjects appeared in the West".
Elbrus-1 and 2 and processor instruction set were implemented in high-level language El-76, which wasused in Elbrus even instead of assembler. El-76 looks very much like Algol-68. The main difference seemsto be in dynamic type linking in El-76 with all the required hardware support. El-76 is intended forsystem programming as well as for application one. Before execution complex El-76 operators are translatedto Java-like bytecode. This bytecode is translated to simple micro-operations by a special unit at run-time.
In Institute of Precision Mechanics and Computing Equipment Vladimir Pentkovski took part in thedevelopment of Elbrus-1 (1978) and Elbrus-2 (1984) supercomputers and led the development of El-90superscalar 32-bit microprocessor. The technical statement of work was created in 1987, and the firstEl90 prototypes were built in 1990.The El90 architecture reflects a combination of RISC and Elbrus2ideas.
The main features of El-90 are:
- superscalar architecture capable of executing three instructions per clockcycle (FP, integer and load/store);
- 32-bit architecture;
- simplified instruction set (compared with Elbrus-2), the majority of which can be executedwithin one clock cycle;
- hardware support of high-level programming languages;
- speculative execution;
- out-of-order execution;
- branch prediction;
- register renaming;
- separate instruction and data caches - 32KB each;
- high-performance pipelined floating point unit;
- multi-level memory hierarchy, first and second level cache support;
- 10-way multiprocessing support;
- debug and performance monitoring support;
- "super-accurate computations" mode support (several processors perform the same calculations,comparing results after each step. If the results do not match, processors repeat the calculations).This mode was required in Elbruses because Elbrus and El-90 electronic element base was not reliableenough for some military applications.
In 1990 Pentkovski began working on El91S, the successor to the El90. But the political andeconomic changes in Russia stopped the financing. We don't know anything about Pentkovski between1991 and 1999. In 1999 his name appeared again in Intel Technology Journal. According to the Author'sBiographies section of the article devoted to Pentium III architecture,(developer.intel.com/technology/itj/q21999/articles/art_2who.htm#pentkovski),Vladimir Pentkovski led the development of Pentium III processor architecture.
Also I would like to remind you that in 1993 Intel introduced an absolutely new 32-bit Pentium processorwith lots of new features. Here they are:
- x86 instruction set is very complicated. Pentim has special unit which translates complexx86 instructions into simple RISC-like ones;
- 32-bit superscalar architecture capable of executing two instructions per clock cycle;
- branch prediction;
- separate instruction and data caches;
- pipelined floating point unit;
- multi-level memory hierarchy, first and second level cache support;
- 2-way multiprocessing support;
- debug and performance monitoring support;
- functional redundancy checking support - two Pentiums perform the same calculations, comparingresults after each step. If the results do not match, processors repeat the calculations.
Well, Pentium III and El-90 seem to have so much in common, don't you think so?
It seems that Pentium has been named after its developer - Vladimir Pentkovski. Really, in 1995 Intelintroduced a more decent architecture, Pentium Pro, which has caught up with the Russian microprocessor of1990. Pentium Pro had an improved x86 instruction decoder, improved superscalar core, and additionallygot speculative and out-of-order execution with register renaming. Besides, it also featured betterfloating point numbers processing unit and multiprocessing support.
And now Pentkovski's former chief, Boris Babaian, is going to bash Intel with a revolutionary newElbrus E2K microprocessor.
Elbrus E2K with EPIC architecture is developed by Russian company Elbrus International(www.elbrus.ru). The project is led by Boris A. Babaian, AssociateMember of the Russian Academy of Science. Mass production is planned for 2001 in case the financial basisremains sufficient. We can suppose that E2K prototype will be introduced by the beginning of the year 2000.And it is likely to be manufactured on a mini-plant in Moscow. Now the processor exists only in form of acertain description on a special language called Verilog. This description embodies the final stage of thelogical projection, and then the Verilog description will be transferred into the silicon. The systemsoftware set for E2K is already completed. It includes the parallelizing compiler and binary x86 codetranslator. About two years ago Elbrus showed the running of MS Windows 95 and MS FlightSimulator gameon Sun workstation with UltraSPARC processor. Elbrus E2K is expected to be much cooler than the Americanachievements in all respects. The E2K will deliver higher performance with lower power and less silicon.Moreover, the software binary translation provides absolute compatibility with x86 code almost withoutany performance reduction, which is not much worse than in case of E2K original code. By the way, theprogramming languages C and C++ with E2K static type are very likely to acquire new peculiar features:hardware supported type dynamics and context protection.
The Russian achievement turned out of great interest to Microsoft Corporation, which operationsystems would really need the performance of E2K. The legend of the computer world - Gordon Bell(his personal web-page is located at www.research.microsoft.com/barc/gbell/),who developed DEC's famous PDP and VAX computers, now heads Microsoft Telepresence Research Group.He has greatly contributed to the popularization of the Elbrus E2K project at internationalconferences.
On 26 May he opened the International Symposium on High Performance Computing in Japan with the keynotespeech "The Next Ten Years in Supercomputing" and on 10 June - the 14th Mannheim Supercomputer Conferencein Germany. In both speeches Gordon Bell paid special attention to E2K. He presented a table comparingElbrus E2K and Intel Merced on a slide, which looks as shown below. Note that the results are evidentlynot in Intel's favour:
| ||Elbrus E2K||Intel Merced|
|Clock frequency, GHz||1.2||0.8|
|Die size, mm2||126||300|
|Power dissipation, W||35||60|
|Bus bandwidth, GB/sec||15||n/a|
|Cache sizes, KB||64/256||n/a|
|Peak advertised performance, GFLOPS||10.2||n/a|
|System ship||Q4 2001||n/a|
As we can see E2K beats Merced in every row.
Also Gordon Bell (www.research.microsoft.com/users/gbell/bio.htm) is a computer-consultant-at-largewho devotes time to startup ventures for R&D in advanced computer technologies.
Elbrus E2K project has already won certain scientific and technical weight since there weredeveloped a lot of interesting architectural solutions, and created perfect new generation compilers.Besides, E2K means that Russia still can boast highly qualified specialists. Professor Boris A. Babaian,and other Elbrus employees teach in Russian high schools, which implies the future increase of thespecialists number in our country.
So, my personal opinion is that the project has already succeeded, no matter what will come outof the silicon realization. Though Elbrus will probably disagree with me here, and prove the worldthat it's the best.
Pages: [ 1 ]
Comments currently: 0
Add your Comment
Enter your username and e-mail address. Password will be sent to you.
Latest materials in CPU section