Russian Electronics: Semiconductor and Processors

He is being a petulant retarded idiot. He is comparing apples with oranges and also makes all sorts of nonsensical comparisons which aren't even chalk with cheese.

I will give you an example. He dismisses SPEC CPU as a benchmark and claims 7-zip and Blender benchmarks are more important to gauge performance. Well guess what. SPEC CPU is what we call a composite benchmark. You run a whole suite with a bunch of benchmarks and then compute an aggregate composite number. A single number which roughly describes performance on average for a whole set of applications. SPEC CPU has a compression benchmark (gzip) and ray tracing benchmark (ray) in the suite. The compression benchmark is roughly what 7-zip does, and the ray tracing one is roughly what Blender does. He claims focusing on a single benchmark is better. Well it isn't. You always look at composite benchmarks to gauge general purpose usability. You go into application specific benchmarks if that application matters to you.

As for the merits of VLIW vs OoO architecture processors. These are well known. OoO is better than VLIW at branch heavy multi-threaded applications with thin compute per thread. And VLIW is better than OoO for technical compute where you need to perform a lot of arithmetic operations. The major benefit of VLIW is that you reduce the amount of transistors in the die used for control and branch prediction and the like and then you can use those to make more functional units instead. The problem with VLIW is that in applications where parallelism is hard to extract the performance craters more quickly. But that is only the case because a VLIW processor is typically wider than an OoO processor made with similar manufacture process and transistor budget. Once you get to wider OoO processors you start bumping into the same issues.

Elbrus was originally designed as a supercomputer processor. It was used on the Don2 radar for example. Elbrus2K comes from that and is well suited for both technical compute and workstation purposes. Also well suited for aircraft avionics like for radar signal processing or other data parallel tasks. Now, if you want to run multi-threaded Java apps, or SQL databases, it won't be as good as an OoO processor designed with the same resource base. Because of the reasons I mentioned. But this is when we talk about traditional VLIW processors. The fact is, I do not even know the specific minutia of how the latest Elbrus2K architecture works. In theory there is no reason why you can't put dynamic scheduling on VLIW just like you can on an OoO processor. One way to do it is to have a software dynamic scheduler. The Transmeta CPUs did this. Another way is to add specific hardware support for branch prediction, instruction reordering, speculation, and the like. Itanium CPUs tried to go this route.

Given the Russian resource base, where they have older manufacturing processes, the VLIW design makes sense. And the military and technical applications typically used by the design are suitable for it. It this was a processor designed for a bank or a web services company then they wouldn't have chosen that kind of design.

He also does other nonsensical comparisons. He keeps comparing the IPC of things which are not comparable. Like x86 and Elbrus2K processors made with different semiconductor budgets with different transistor budgets.

The difference in clockspeed is kind of irrelevant. Typically wider processors, like Elbrus2K, have lower clockspeeds but compensate for it with more peak instructions per cycle. What matters is how many instructions you can compute per second. Comparisons with AMD Epyc are nonsense. Made with different process and multiple chiplets. If you made an Elbrus processor with similar manufacture resources the Elbrus2K would be way wider than the AMD x86 one.

The only things he mentions which I can relate with is are lack of availability of the hardware for consumers to buy and lack of openly available compilers and tools. This was also a problem with Itanium. I can sort of understand it since this is the secret sauce of MCST. But it will cause problems for the success of the architecture itself in the long run.

I can also relate that he is comparing two things he supposedly can buy. But, the thing is, for the military they need to buy things made in Russia. And given the types of applications the military needs, Elbrus is the ideal processor and design for them. Baikal does not even design their own CPU cores. So I do not consider them to be on the same playing field as MCST. Not in a long shot.

Russia has a fabless semiconductor company which designs RISC-V cores for licensing to third parties. So you could say this is the embryo of a Russian ARM like competitor. In the long run this company should have good chances for embedded applications and who knows what else. The Mikron manufactured embedded chip mentioned a couple posts above uses one of their core designs.

https://syntacore.com/

I think companies like Baikal will have no future in the new conditions where Russia can't access foreign semiconductor fabs.

There was/is similar software in the US for binary translation. Digital Equipment Corporation used to have something called FX!32 which was used in Windows NT for Alpha processors. It translated x86 Windows applications to run on Alpha Windows. Apple had something similar, called Rosetta, which they bought to translate PowerPC applications to run on x86. I assume the latest Apple Macs, the ones with the M1 chip, use something similar to run x86 Mac applications on ARM Macs.

lancelot wrote:There was/is similar software in the US for binary translation. Digital Equipment Corporation used to have something called FX!32 which was used in Windows NT for Alpha processors. It translated x86 Windows applications to run on Alpha Windows. Apple had something similar, called Rosetta, which they bought to translate PowerPC applications to run on x86. I assume the latest Apple Macs, the ones with the M1 chip, use something similar to run x86 Mac applications on ARM Macs.

It is exactly what Apple is using to translate x86 applications. It does it very successfully actually. I was actually very impressed with the M1 when I normally am critical of Apple.

MacOS X is derived from the NeXTSTEP operating system. So it has something called "fat binaries". You can compile a composite binary targetted at multiple CPU architectures. And when you click to run the binary on the operating system it selects the binary most appropriate for your computer. Later versions of Windows also have something similar.

On top of that you have Rosetta 2, which can either translate a x86 binary into an ARM one when the application is installed, or JIT compile a X86 binary so it can run on an ARM platform.
Of course MacOS X itself only does this for running MacOS X86 platforms on MacOS ARM platforms.

To run a Windows application on a Mac you used to need something called Parallels, which was a machine emulator, installed.
When the CPU was an x86 that was probably pretty trivial. You still needed a copy of Windows to install on the emulator. It was probably something kind of like VMWare. Now that Apple moved to ARM I don't know how you can run a Windows application on a Mac.

Video about Elbrus and the prospects of computing in Russia in the current situation.

Good video. He forgot to mention about the Ministry of Industry and Trade on their investment for production of their on Fabrication equipment but overall he is spot on.

The ending part was kind of an eye opener regarding the open source community openly attacking at Russia. It makes the open source community completely contradicting and him explaining that there needs to be a Russian Github alternative is spot on. It was never a secret in the past that most of these open source groups are part of intelligence communities and thus work for groups or governments. So it cannot be trust anymore. But at least a repository on software and code that works just fine for Russians is needed.

And I am not entirely sure on his last comment regarding RISC V. RISC V is fine enough to have in development for various systems while Elbrus can still be worked on to deal with HPC enviornment.

Anyway, a lot of good points brought up.

That article is misguided in several ways. Elbrus already has a 90nm CPU design manufactured at Mikron. The Elbrus-2SM from 2014. At best they might make a new Elbrus design more optimized for the 90nm process specific to Mikron which takes into account current technology and requirements. MCST does have another processor design which was originally produced at TSMC, the MCST-R1000, which uses SPARC instead of Elbrus architecture. It might be a good idea to redesign that for manufacture at Mikron as well. It has like half the transistors than Elbrus and for a lot of applications it is good enough.

Mikron also manufactures simple RISC-V chips which would be useful for embedded applications.

The article's claim that Elbrus is the only chip design team in Russia is false. There are other chip design teams. For example the ones who did the ELVEES RISCore32 architecture or the NeuroMatrix DSP family. There are also fabless companies like Syntacore who design CPU cores.

Given the limited size of the chip market in Russia it makes no sense for Elbrus to have their own fab. The state just needs to give a swift kick in the pants to Mikron and Angstrem-T.

lancelot wrote:That article is misguided in several ways. Elbrus already has a 90nm CPU design manufactured at Mikron. The Elbrus-2SM from 2014. At best they might make a new Elbrus design more optimized for the 90nm process specific to Mikron which takes into account current technology and requirements. MCST does have another processor design which was originally produced at TSMC, the MCST-R1000, which uses SPARC instead of Elbrus architecture. It might be a good idea to redesign that for manufacture at Mikron as well. It has like half the transistors than Elbrus and for a lot of applications it is good enough.

Mikron also manufactures simple RISC-V chips which would be useful for embedded applications.

The article's claim that Elbrus is the only chip design team in Russia is false. There are other chip design teams. For example the ones who did the ELVEES RISCore32 architecture or the NeuroMatrix DSP family. There are also fabless companies like Syntacore who design CPU cores.

Given the limited size of the chip market in Russia it makes no sense for Elbrus to have their own fab. The state just needs to give a swift kick up the pants to Mikron and Angstrem-T.

Chip usage and demand may grow but even then doesn't make sense for MCST to have their own but rent.

I figure they can look at 3d chip design as well.

GunshipDemocracy wrote:

sepheronx wrote:

Chip usage and demand may grow but even then doesn't make sense for MCST to have their own but rent.

I figure they can look at 3d chip design as well.

I dont get that. If all of above mentioned Russian microprocessor designers could buy manufacturing  services from  Taiwan why they cannot share one fab in Russia too?!

I'm sorry?

They already do use Russian FAB's for their microprocessors but Russia doesn't really have anything below 65nm for more higher production rate.  They are just working out what to do now that they dont have access to Taiwan's FAB's and so they are working with Mikron and possibly others within Russia that can produce chips of various types.  They may even design newer chips using older lithography tech that can be fully utilized at lets say 65nm or 90nm but without skimping on too much in terms of performance.  Like for instance 90nm but 3d chip design.  I am unsure about this that maybe @lancelot may be able to explain better than I can.

90nm with 200mm wafers is more than enough for military, space, and automotive chips. Existing fab capacity at Mikron and Angstrem-T is also more than enough for those applications. It is a matter of organization and keeping those fabs at high usage rate.

Once the public lens is away from Russia and demand for 28nm worldwide craters over the next two years I expect Russia to be able to find some place to fab 28nm chips outside Russia. There will be like three dozen fabs which can make those chips by then. Russia will be able to find someone to make them the chips. You will see. It is all a matter of getting the right corporate scheme. As long as Russia keeps those chips inside civilian government institutions and they aren't sold worldwide no one will care.

If Russia wanted to they could probably make a 65nm with 300mm wafers fab with all Chinese equipment right now. Problem is most designs Russia wants to produce which they currently outsource production are 28nm or better. It will take the Chinese at least two years to get that kind of equipment operational. And who knows what will be the production rate for it.

I think Russia should take a multiple pronged approach.

Like I said before Mikron and Angstrem-T production models should be revamped and it should be possible to switch production of a design from one factory to the other with minimal fuss by making them have some common fab processes. New suppliers for materials need to be found now that Europe isn't available as a supplier anymore. The existing production facilities need to be optimized and if necessary new 180nm and 90nm production processes should be licensed from China with minimum fuss. It is quite obvious that the 90nm production process Mikron currently uses, which was licensed from STMicro, is not optimal for making CPUs. Elbrus-2S+ made at TSMC Taiwan at 90nm has 500 MHz clockspeed, while the Elbrus-2SM made at Mikron Russia at 90nm has 300 MHz clockspeed. The lowest voltage for the Mikron chip is also higher. That isn't something you can just ignore. It is quite evident that the Mikron process or the design they made for Mikron process is performing worse. Angstrem-T supposedly got licenses for using 90nm IBM processes but I never heard of them actually putting them into production. The state also needs to fund the creation of chip designs to be manufactured in Russia to replace those chips which cannot be imported anymore. I am talking about things like control chips for vehicles, embedded systems for industrial applications, and the like. The consumer electronics sector is kind of beyond the scope of existing production facilities I think.

Russia should invest in production facilities for advanced packaging of chips. All the equipment for that can probably be imported from China right now. And Russia should drive production volume for the fabs up by creating some sort of incentive for using chips either packaged or diffused in Russia. The previous program which made incentives for chips designed in Russia was clearly not enough.

Russia should get Turkish and Chinese companies to open factories to make consumer electronics products in Russia proper. I am talking about washing machines, clothes dryers, freezers, air conditioners, and stupid things like that which use a lot of bulk steel and not as advanced electronics. They should push production of both those and cars and boats and airplanes. That will I think take the most advantage of the capabilities of Russian industrial output. They should make those companies either open factories in Russia to make chips, or use Russian factories to make chips for them.

Russia could also consider increasing their production output of lower complexity ICs which can be made with Chinese equipment like discrete electronics, LEDs, or other semiconductors for automotive and power electronics like inverters and the like. Production of energy intensive or petrochemical related materials for semiconductor production should also be incentivized. For example the manufacture of polysilicon for own use and export or that of electronic-grade ammonia.

Towards the middle or latter part of this decade once either the Chinese tools become available or Russia makes its own tools then it will be time to think of creating new fabs with more advanced equipment.

The consumer sector will have to rely on second sourcing of products in the meantime.

Once again lancelot, you are ignoring realities.

Russia cannot purchase chip production from China due to sanctions.  They aren't wanting to produce chips for Russia. If it was that simple, they would be doing just that instead of looking internally.  All chips Russia will get from China will be at best Chinese design or foreign chips and sold via black/Grey market like Alababa.  Russia has no choice now but to make chips internally and then possibly send then to Malaysia to package them since they offered to do it.  So lowest Russia will have is 65nm from Mikron and or if the 45nm is real or not from GS.

Also, Russia is working on their own production for equipment to do 28nm. Why are you constantly ignoring this fact that the government under ministry of industry and trade has contracted out already?

Also, you can't compare Elbrus 2C+ and 2SM. 2C+ uses elvees DSP cores on top.  2SM doesn't. One is produced in Russia since 2C+ is for radar systems and anything that utilizes Digital signal processing.

https://www.russiadefence.net/t3439p650-russian-electronics-semiconductor-and-processors#369506

This is about the news of Russia developing their own lithography equipment that is blatantly being ignored for some reason instead promoting a fictional agreement with China which China isn't interested in doing to avoid sanctioning themselves too.

Governments talk a lot but business matters and even China isn't really keen on doing business with Russia in a lot of fields but at least using sketchy third parties to do trade with each other.

The sanctions only apply to things with Western content in them. In a couple of years China will have commercial fabs with 100% Chinese production tools and equipment in them.

Even today the Chinese government already has sanctions proof semiconductor production facilities using equipment made and operated by CETC which can make chips at 90nm or better. Those Chinese production facilities do not operate in the commercial market but are part of the Chinese MIC. I told that here before.

China has had sanctions on purchasing military grade equipment and components from the West since the Tienanmen Square incident in 1989. How do you think they make the chips for AESA radars and the avionics in their fighter aircraft? Russia's MIC only got similar sanctions after annexation of Crimea in 2014. So how come the Chinese have AESA radar on their aircraft and thermal sights on their tanks? They surely are not importing them.

The Chinese commercial sector is currently the market leader in security cameras and they are quickly moving towards dominance of the commercial camera sensor sector as well. Why do you think the US slapped sanctions on Hikvision, a commercial CETC subsidiary, and yet made a carveout so their own MIC can continue purchasing Hikvision equipment?

And sorry, but a more complex chip tends to be slower, not faster. So how come the Elbrus-2C+ has the same CPU cores, with extra DSP cores, and is faster? 500 MHz vs 300 MHz. It is quite simple. The 90nm process at TSMC is better than the one Mikron uses which they got from STMicro.

As for Russia making its own lithography equipment, it takes more than lithography equipment to make chips. The Chinese have been at this for over a decade already. Probably close to two. And you think the Russian industry will solve all those problems in less than a decade? I think you are being overly optimistic. The current plan only seems to be focused on lithography to begin with. What about making masks, and all that other crap? You need like a dozen different types of machine tools just on the main production line of a semiconductor fab. That is ignoring all the equipment you will need to make the materials to operate the fab.