What could go wrong? Review – from FPGA to dating sites
I read an article with an inspiring slogan
«With the introduction of government initiatives, Russian developers are confident that by 2030, more than 70–80% of companies will reorient themselves to using local equipment.” and thought – really, what could go wrong? What a great plan, reliable as a Swiss watch.
For LL and experts with yesterday's registration: a regular article from a regular neural network, well, you know.
Part 1. Russian hardware
For some reason, the authors of such articles often believe that technologies cannot be lost. What can I say? The USSR had technologies that allowed it to land devices on the Moon, Mars and even Venus, move Lunokhods 1 and 2 on the Moon, and return lunar soil. They did.
Similarly, at Mikron and similar production facilities there were, or were promised to be, separate production elements, or separate production of elements using 250-180-130 nm technology, and there was talk of switching to 90-65.
2022, quote:
At the Microelectronics-2022 forum, residents of the Technopolis Moscow SEZ, Mikron JSC and NM-Tech LLC, signed a cooperation agreement to set up technologies and launch production at NM-Tech facilities.
Micron will assist in the development of such products as:
dual-interface SBIS for use in MIR payment cards;
integrated circuits for passport and visa documents with biometrics and for compulsory medical insurance policies.
Poles of Power in the World of Chips
Well, two years have passed? Have you mastered it?
Now, as an example of successful success, a joint venture with Planar is being presented for either production, or development, or localization, of 250, or 300, or 350 nm steppers, described in the press in the style of “It will be great, but how and when is none of your business, This is a lecture for collective farmers“.
You can also look at interesting facts, for example, what a neural network accelerator is made of Mustang-F100-A10 from Nieshants-automatics
Or ask where the FPGA went in the transition from Continent 3.9 IPC-3000FC to the Continent 4 line.
It seems like there is a project, One of the projects that our “Continent Hardware” department is working on is a module (expansion board) for a crypto accelerator based on FPGA Xilinx Kintex Ultrascale+. – but they are not looking for anyone to fill this position, and for a similar vacancy in Moscow, they were looking for someone with a courier’s salary (FPGA (RTL/FPGA) development engineer from 173,000 ₽ before taxes).
Kintex™ 7 FPGA – called AMD Kintex since February 14, 2022. 28nm process technology.
Usergate is similar – They promised a year ago to roll out FPGA, but here's the problem – in the UG blog on Habr, the penultimate article is 21 Aug 2023 at 10:34, the last one is on August 9 at 12:00.
1 article per year. And that's all.
You can compare the declared characteristics of the Continents and UG with a simple Fortinet FortiGate 6500F:
IPS Throughput 2 170 Gbps
NGFW Throughput 2, 4 150 Gbps
Threat Protection Throughput 2.5 100 Gbps
Not to mention something more complicated, like the Palo Alto PA-7050
Tests – TEST RESULTS OF NGFW/UTM CLASS SOLUTIONS (v1.0 FROM 19.02.24)
What's the problem with FPGA and nanometers, you ask, because your grandfather had that same Wolf catching eggs (Electronics 24-01 “Well, Just You Wait!”) – that is Nintendo EG-26 Egg Game & Watch series.
There is no problem, everything was discussed 2 years ago in the article Harmful misconceptions about nanometers. Or why having access to only 90 nm in Russia is a disaster
In the same way, there are no problems for either civilian communications specialists or the military “in general”, and the report by the Royal United Services Institute (RUSI) (Silicon Lifeline: Western Electronics at the Heart of Russia’s War Machine) is slander, and that’s all.
Nothing new in six months since the articles Russian microelectronics – two years later and The long path of the microprocessor to the market or learning to read news about Russian lithographers
That's all I can say about the Vietnam War. and about pleasant dreams about Russian hardware.
Part 2. Transition price in the first approximation
At first glance, it may seem that switching to an import-substituting server is simple and easy. Schwarzenegger can't, but I canand this is true if you have 1 (one) server. And even 2 (two). And even 2 (two) processors in each. You can even install Linux with KVM on such a server. And it will even work.
What's not to work there?
Intel Xeon processor, motherboard with track layout (and layers) from a Chinese vendor of the 2-3 echelon, on which in the Russian Federation they somehow performed surface mounting of Chinese power elements, on a Chinese machine, with Chinese solder. Made in our country (C).
The cardboard for the boxes is definitely from the Russian Federation. I'm not sure about the cardboard machine and paints.
The cow, of course, state Chinese, and everything that it gives – milk, calves – is already ours (C).
Memory – choice of 3, Micron, Samsung and Hynix.
SSD – choose from InnoGrit, Marvell, Maxio, Phison, Silicon Motion (SM), SandForce, Samsung.
IPMI to choose from ASPEED AST2500 and AST2600 (I haven’t seen 2700 and 2750 yet).
All this is great, as long as there are 1-2 servers. As soon as the slightest automation is required, not even ironicat least redfish – oh, how disgusting everything is there. How well the BIOS settings and boot selector are saved during the update, you will not know until you start to encounter this.
Linux with KVM also works, and works very well, until you start to clarify the nuances. For example, what and why is written in the article The Linux Scheduler: A Decade of Wasted Cores, or where to look for something similar in terms of session content VMware Explore US 2023 Breakout Session – CEIT1319LVR Extreme Performance Series: Performance Best Practices by Mark Achtemichuk, Valentin Bondziovideo CEIT1319LVRor Extreme Performance Series 2022: SAP HANA 8-Socket Performance or Extreme Performance Series 2024: Enabling and Optimizing GenAI Workloads with LLMs ?
But everything works. In the demo stand for 3 servers. Without load and tests. Why only three servers in the demo, you ask? Because it is impossible to connect more servers to the import-substituting external SAS shelf, and there is no money to buy a Russian storage system for the demo stand – there is no money for a storage system, especially for a Broadcom Brocade G720, as well as Mellanox MSN2410-CB2FO (48-port 25GbE + 8-port 100GbE), and two of them are needed for the stand, and both of them – some businesses in the Russian Federation are very conservative and prefer FCP. And looking at some so-called network engineers, I understand perfectly why.
Part 3. Quality of the offered product
As soon as we start discussing the package contents:
– iron,
– plus hardware support,
– plus a hardware warranty (with delivery time and penalties – SLA,
– plus the ability to automate hardware settings”,
– plus the quality of work of any import-substituted fork of Openstack under the brand “this is not a fork of a 3-5 year old branch at all”,
plus calculation of the cost of hardware for Ceph or (by the way, not bad) vitastorit turns out that there are not many people on the market who are ready to work with all this.
More precisely, ready-made receive there is a lot of money, but those who can do anything other than cry “it doesn't work” – very few. Because many of those who are ready and able to work – were taken to Europe, and there are simply few of the remaining “ready”. They are not at all interested in working with these complexes, especially in small volumes.
At the same time, internal solutions of Yandex and VK are not replicable, that is, there is simply no “ready-made box like Windows + S2d” or “VSAN ready nodes” or “Nutanix”. All import-substituting basic products (OS + virtualization) are not ready for use in production at all, even with 12*6 requirements, not to mention 24*7*365.
The level reached is 8*5*300 – when the server can lie overnight. The configuration of libraries should be cemented, otherwise you can get a pumpkin from the cluster with one awkward upgrade –y.
Forget about updates.
And all this somehow works until the moment when your import-substituted imitation security does not even offer, but demands to install an import-substituted product, containing antivirus software, not identical to the natural one, inside the hypervisor. Even the developer's support does not know how this product works, but how it affects the work is evident from the fact that, according to experts, the level of smoke and the average temperature in the office after installation increase by 4%.
Part 4. Transition price in the second approximation
Hardware. It costs 30% more. 20% because production must pay for itself, and 10% because for every 5..20 servers purchased, you need to immediately buy another one – for your warehouse as a reserve. Not to mention storage systems. And not to mention SDS. There are simply no SDSs “that just work, and not all this.” Those that really work, and are ready to show the PMI and bug tracker.
People who are ready to support and configure these designs “so that it works normally”, and not “so that it starts up and is ok” are expensive, capricious, and… and.
And during recent interviews (we wanted to hire a couple of people inside the Russian Federation, so that we could possibly take some of them to the West (or East) later), an unpleasant surprise came to light – there are a lot of people who do not understand fundamental things, but are trying to replace them with faith and templates. This is no longer news – the Russian IT community has long been reborn from a community into some kind of mixture of a workshop structure from the late 19th century and a miners' club. In all the chats of the “Russian mining federation” they cry once a month, like “I thought this would be about mining, but you're showing me pictures of Aquarius on a rocket.” Like enough; you wouldn't look to find a bishop here, I reckonone-legged John Silver would say in such cases, it's also called ham.
Finding people is extremely complicated by the fact that more and more people in IT in the Russian Federation can retell the standard opinion (Windows is bad, Linux is good), but as soon as the conversation starts to get a little more serious, “reboot your computer”, when you already need to get into the tcpdump + strace bundles, or act in a hybrid infrastructure – that’s it, dry your oars sir. If you ask such an enthusiastic candidate to calculate how much it will cost to implement CEPH on at least 100 TB of useful capacity with N+2 availability, and at least 150,000 IOPS 4k block per write, then records of such slowdown have not been published since the Luna-25 flight on August 19, 2023. Why 150k, you ask? This is how much the junior storage system of one company gives – both according to the calculator and in fact, without dedupe and correction.
If you need a person who can get into the KVM source code and see what's there, and at the same time tell you what's in the rabbit, and not execute
git clone https://git.openstack.org/openstack-dev/devstack && ./stack.sh
they are not even in the so-called developers of distributions. There were some in a couple of integrators.
Part 5. Russian clouds
There could have been a video of “it's just some kind of holiday” with Vladimir Etush, but it seems it's not on Rutube.
In Russian clouds, everything is probably great. Wherever you poke, it hurts. To lie for a week a year, that is, to have not the declared 9.99, but the actual 98% – that's still very good. IOPS limits are everywhere, there is no functionality to “squeeze a lot of IO in the Russian cloud, except to actually give the customer a separate server with NVME”.
But NVME raid is not the same as even VROC. Given that VROC in terms of quality is almost the most disgusting thing I've seen, right after Basis, and slightly better than Broadcom VMware before Update 3 and Windows before Service Pack 3.
Storage vmotion between disk types is cut so much that even moving tens of TB of loaded database is almost out of the question. There is no flavor update, and none is expected. Of course, I can do it in DSC, and I can do it in Ansible, I can do it manually, but how is that possible?
Part 6. Why does the employee need all this? And the management too
And now we come to the softest part.
The continuation, of course, is on my Telegram channel and on Pikabu.
