Russian Electronics: Semiconductor and Processors

MCST is relying on lithography shrinking to get more out of the design. This is only part of the proper approach. The other one
is optimizing the transistor layout. I am assuming something like this is being done, but originally the TSMC production line was
using off the shelf transistor block layouts. Having a CPU design is not the same as having an optimized rendering in silicon.

We have seen 32 core CPUs from AMD that are using 12 nm "FinFet":

https://www.globalfoundries.com/sites/default/files/product-briefs/pb-12lp-11-web.pdf

The Elbrus had a substantially smaller number of transistors compared to Intel x86 designs when it was first mass produced. That
is why it had ultra low power output. A huge part of the transistor count goes to L1, L2 and even L3 caches in modern
CPUs, but the Elbrus still has a some sort of advantage in terms of lower number of transistors. It most likely has something
to do with its VLIW design and the use of the compiler to do what x86 does in hardware. The Intel Itanium tried this but
was a failed effort to "rip off" the Elbrus since it made no sense of Intel to dabble with such a non sequitur branch to its
hardware focus since the 1970s. Russians are really good at coding and software technology is not trivial and secondary
to hardware.

To wit, the Elbrus 32 core can be rendered on 14 nm. So they are kicking the can down the road in terms of properly laying
out this processor. And 14 nm is actually achievable for Russian IC manufacturers.

https://www.zelenograd.ru/hitech/v-zelenograde-namereny-postroit-fabriku-chipov-28-nanometrov/

We have 28 nm production capacity being deployed. This easily allows 8 core Elbrus processors to be manufactured in
Russia. Relying on TSMC is a bad idea.

miketheterrible wrote:Russia may switch to Chinese nodes since China is working closer to 7nm upcoming.

But while I agree it should be done in Russia, I dunno what overall plans are.

They may wait till they move more to 3d chip design.

China is safer than Taiwan as an offshore production point, but it is the same problem of having your jewelry in someone else's
pocket.

Even if Russia goes down the photonics pathway, there is a clear need for conventional silicon IC production inside Russia.
Aiming for short term transition to 14 nm would be ideal. The 7 nm target is more hype than substance. In fact, 7 nm
is too close to the 3 nm molecular cluster transition limit where continuum mechanics does not apply. So there is no
"material" to have predictable properties and which can be molded into functional components. There is talk of going to
5 nm but I think that will be entering the excessive error rate regime. It is already an achievement to be using 7 nm but
the end of the line is fast approaching.

Hence there is a need for photonics. Photonics will allow 3D in 2D hardware. Different frequencies of photon can be
use in the same "wires" but only one type of electrical current. So a chromatically selective photonic transistor (likely
some sort of complex not as simple as any of the current conventional IC transistors) would enable multiple IC layers
to coexist on a single plane. The different photon frequencies can be routed into different domains of a generic transistor
grid (sort of like FPGAs) to deploy the 3D aspect.

PhSt wrote:
Isn't this going to encounter possible "complications" since TSMC is based in Taiwan?  we know Taiwan is a US lackey.. It would be best if Russia develops its own semiconductor industry so they wouldn't have to rely on foreign fab foundries to manufacture chips for them.

True. But consider if word gets out that Taiwan tampers with its customer's orders: The damage to their reputation is going to be measured in tens of billions of future business alone, and people would be incensed to actually spend on their own chip fabs. I don't think they would be willing to risk it, even for such a paltry contract such as this. But who knows?

https://www.stalkerzone.org/domestic-16-core-processor-elbrus-16s-has-russia-really-caught-up/

Hole wrote:https://www.stalkerzone.org/domestic-16-core-processor-elbrus-16s-has-russia-really-caught-up/

Still no real push to manufacture it in Russia as opposed to Taiwan.

The author also does not understand CPU clock frequencies are not some intrinsic measure of speed. They merely reflect the
number of pipeline stages. So the Pentium 4 with its high clock rate back around the year 2000 achieved this since it had
long pipelines with small stages. It was also crap that was prone to pipeline bubbles and was replaced by the "Core" model
that had much less pipeline stages. Of course lithography resolution increases maintained the clock around 3 GHz. But
if you not there has not really been much increase in Intel CPU clock rates over the last 15 years compared to the 15 years
previous to 2005.

The VLIW principles behind the Elbrus automatically make it a short pipeline design with complex and large pipeline stages.
This "wide" format allows it to execute more instructions per stage so it has a lower clock but a higher productivity per
clock cycle compared to something like the Pentium 4. The 300 MHz E2000 was not "absurd". It was effectively a
1 to 1.5 GHz "Intel" analogue. But it also was not transistor optimized. Intel spends a lot of time and effort and money
to hand tune its transistor layouts. If MCST was able to do this then the E2000 would have been close to 3 GHz equivalent.

Regarding manufacture in Russia, the 28 nm process is good enough to produce the 16 core version on Russian soil. The
hype 7nm and 14nm fabrication processes are nothing of the sort (see my post in the general physics thread). The 28 nm
process without FinFet likely involves actual transistor sizes below 50 nm. The magic 7 nm transistor sizes are in reality
over 20 nm in size. So the silicon foundry gap is closing.

PhSt wrote:I have been a regular visitor to another forum (sinodefence) and from the discussions there concerning Chinese developments in the semiconductor industry it seems domestic Chinese companies like SMIC are having difficulties manufacturing chips with 7nm and 5nm process to compete with competitors like TSMC. It seems China still relies on the US and other suppliers for machinery and essential tools for production. How sophisticated is this field of science? and since China is heavily investing in this area, perhaps it is something that can be considered an essential strategic industry (just like aircraft engine production) that Russia needs to focus on.

The so-called 7 nm finFet from TSMC is actually a 10 nm process. This was demonstrated by a video where an AMD Ryzen CPU had
a piece of it extracted with a highly precise instrument (see the physics thread). Anything under 10 nm right now is marketing BS.
So I am not sure how much of a real problem is actually involved in China. Dependence on US companies is a real weakness just
as it is for Russia.

But my impression is that China can actually manufacture 10+ nm class ICs.

Tests of various games running on the Elbrus 801-PC workstation.

ABOUT THE RUSSIAN ELBRUS PROCESSOR
February 21, 2021

I have to say that specific advertising policy of the former head of MCST Babayan in the form of periodic high-profile claims that the new processor Elbrus E2K is about to “break” Intel, permanently ending zilch, being zero from the beginning, has raised in me distrust of this company since the beginning of the 2000s. Moreover, when Babayan with a team of leading employees left for Intel in 2004, I thought that this was definitely all.

And yes, the company went quiet, and all the following years just quietly profited from something for the defence industry, almost without communicating with the outside world. It withdrew into itself. With long interruptions, it dryly posted new versions of processors on its website:

2005 – Elbrus-2000, which turned out to be a very simplified version of the E2K. Processor technology 130 nm, frequency 300 MHz, 1 core.
2010 and 2011 – Elbrus-S and Elbrus-2C+. Processor technology is 90 nm, the frequency is 500 MHz, 1 and 2 cores, respectively.
2014 – Elbrus-4C. 65 nm processor technology, frequency 800 MHz, 4 cores.

All of this was produced and financed for the defence industry, and so everything would have been sluggish and continued further, if not for the events of 2014. The fall in the ruble exchange rate and the sanctions of the western world have led to the need to have their own computing equipment not only in the military, but also in the civilian segment. But, I must say, the Medvedev government was particularly slow to do so.

The movement around the Elbrus processors intensified in an avalanche in 2020, exactly with the arrival of the Mishustin government. Probably just a coincidence… In general, it became clear that the government is planning something big, which was confirmed by the unprecedented government resolutions of the end of last year 2020. Why in the media it even slipped out, somehow sluggishly – it is not clear. Media leaders don’t understand the scale of what is happening?

I will start with the fact that from January 1st 2022, in order to obtain the status of Russian equipment and supply products to state structures, it is required to use a Russian central processor or controller in its composition (paragraph 2458 of 31.12.2020). Previously, this requirement was only applicable to data storage systems (paragraph 719 of 17.07.2015).

Nevertheless, the first and successful attempt to introduce Elbrus into civilian state structures nevertheless began in 2016, and since May 2017 the first data processing centre was introduced at the passport and visa service based on Elbrus-4C processors (back then about 130 servers, now already about 200 servers) instead of the IBM mainframes, which have been there since 2006, the cost of technical and service maintenance of which has suddenly increased. Thus, the cost of 1 year of maintenance of IBM mainframes has actually become comparable to the purchase of a data centre based on Elbrus.

From January 1st 2022, the requirement for a domestic processor or controller applies to almost any electronics, including ATMs, cash registers (processor or ad valorem part), I/O devices, monitors, peripherals (for example, printers) and “semiconductor storage devices” (controller in solid-state drives). There is also a requirement for quotas (paragraph 2013) for the purchase of radio-electronic equipment, and several other standards and proposals that are likely to become standards in the coming months.

Since the transition to a different architecture of the central processor in the equipment produced in Russia is often quite a complex process, measures of state support for companies developing products based on Russian processors (paragraph 109 and paragraph 529) have also been worked out. For example, the Ministry of Industry and Trade gives grants and subsidies for the creation of equipment with a Russian processor. The Ministry of Finance may also give similar grants, and there is a discussion about this. In the same way, the development of automatic design systems (CAD) and the development of electronic components and chips are supported. There is a program to subsidise consumers. And add to this the tax benefits for IT companies!

The scale of the measures taken indicates the seriousness of the state’s intentions to implement a massive transition of industry and state structures to domestic equipment. This can be treated in different ways (there will always be opponents), but at the moment it becomes obvious that a clear practical course has finally been taken on import substitution in radio electronics, and it is happening!

From the above, it follows that radio electronics engineers and programmers of system and application software have a new potentially very promising direction, which already needs to be actively studied, especially since several large successful projects have already been implemented on the basis of Elbrus processors:

In addition, a very high-load system for the Ministry of Internal Affairs (Center for the Automated Recording of Administrative Offences) is currently being deployed in Elbrus, where video analytics will be conducted in real time and there will be large databases. The system is expected to be completed by the end of 2021.

I think that now more and more new projects of this kind will actively appear.

Now companies that used to develop motherboards and build servers and other computing equipment based on x86 processors are actively expanding their solutions with lines based on the Elbrus processor. There are already about a dozen of them, and in the future there will be hundreds of them.

In connection with such activity of everything and everything, I think that Elbrus processors will soon catch up with Intel and AMD in performance, and maybe even approach them in cost due to the volume in the batch and taking into account the unprecedented measures of state support.

I think that our own factory is not far off. After all, in one of the drafts of the “strategy for the development of the electronic industry of the Russian Federation for the period up to 2030” for 2022-2025 there is the development of microelectronic products at the technological level of 28 nm and below, and development at the level of 14 nm and below, although the final document does not contain such specifics. There, with more general terms until 2030, the development and industrial development of silicon technologies for the production of electronic components with topological standards of 65-45 nm, 28 nm, 14-12 nm, 7-5 nm and subsequent production of products based on them, as well as the development of products based on silicon technology with a topological norm of 5 nm, followed by the production of products based on them at foreign factories and the transfer of production to the Russian Federation, are recorded. It is planned to create silicon factories operating in a “foundry” mode for the production of digital integrated circuits with topological standards of 28 nm, 14-12 nm, 7-5 nm.

I wonder if the norms of 5 nm will be relevant in 10 years? Perhaps, since science has almost reached the physical limit of microminiaturisation, and further reduction in size will affect the durability of such products. Perhaps productivity growth will take a different path.

Elektromozg

https://www.stalkerzone.org/about-the-russian-elbrus-processor/

How can somebody write an article about the Elbrus today and totally omit any reference to the 8 core version? Then
why bother listing any of the previous models?

http://mcst.ru/elbrus-8c
http://www.ineum.ru/elbrus_801-pc

The rest of the article reads like an internet forum post with basically no real information about either Russian IC fabrication
or development plans for the Elbrus (e.g. native 64 bit floating point).

miketheterrible wrote:I heard Elbrus 16C which is next and apparently already seen, is more than simple core increase. Apparently more of an IPC improvement too.

I hope they make more effort with cost-free 64 bit floating point. They should also add more FPUs for vectorization.

In the article we have the author harping about how the Elbrus needs to catch up to Intel and x86. The Elbrus should
never be subjected to this "race". It is rigged from the start. Recall back in the 1990s we had Cyrix CPUs "failing to deliver"
with Quake supposedly because they had inferior FPUs. It turns out that Quake used the FPU for integer manipulation
and not as an actual floating point unit. Cyrix did not fail when code did not use Pentium specific optimization hacks.
Using write combining would achieve similar performance but without the Pentium specific trickery.

The Elbrus should be used to create a Russian software ecosystem instead of trying to import western rubbish (yes
rubbish). Russia should ban Micr$haft software from any institutional use. Ridiculous bloat with built in back doors
for NSA infiltration. No thanks. In terms of real productivity, the test is in HPC environments. Here is where the
improvement effort should be (i.e. expanded floating point power).

Actually the current Elbrus implementations all take a big hit from 64 bit floating point compared to 32 bit. This
has nothing to do with emulation. The Elbrus simply does not have 64 bit wide FPUs.

The 8 core variant has the following official numbers:

Пиковая производительность:

25 операций в такт в каждом ядре (8 цел., 12 веществ.)
250 GFLOPS одинарной точности, 125 GFLOPS двойной точности

There is a factor of two hit from using 64 bit (double precision) floating point.

And this is not an academic problem. I know from personal experience that single precision is crap for physical process
codes. If Russia wants an HPC processor, it needs to have 64 bit x 4-8 units for floating point. If the need is for some
military specific format, then have a separate model for it. Trying to satisfy contradictory constraints is a losing proposition.

No. You are wrong. The Intel and AMD x64 is wired at the CPU level for 64 bit floating point. The Elbrus is clearly using
32 bit floating point and cannot do 64 bit natively. It can do 24 floating point operations in a single cycle for 32 bits
and 12 for 64 bits. I am quoting MCST itself so you can send them an letter telling them that they do not know what they
are talking about. Originally it was claimed that the Elbrus would do 24 floating point operations per cycle for double
precision (64 bits).

Your are also engaging in obfuscation with the vector MMX unit. The Elbrus does not even have an MMX type multiple FPU
capability. What relevance does 4 x 32 have for the Elbrus? The problem is that it cannot do 1 x 64 in a single effective cycle
and requires two, thus the exact factor of two GFLOPs reduction.

https://wiki.tuflow.com/index.php?title=Hardware_Benchmarking_Topic_Single_Precision_VS_Double_Precision

The primary difference between the SP and DP numbers in the above tests is due to main memory hits. It costs more
to push 64 bit words between RAM and the caches than 32 bit words. There are also hits from FDIV and FMUL total
cycle differences between single and double precision.