News Russian-Made Elbrus CPUs Fail Trials, 'A Completely Unacceptable Platform'

[Admin]

Russia's Sber finds the domestic Elbrus-8C CPU platform to be unacceptable for modern banking workloads.

Russian-Made Elbrus CPUs Fail Trials, 'A Completely Unacceptable Platform' : Read more

 

[Alvar "Miles" Udell]

It only took AMD 5 years and a shoestring budget to design the Zen architecture from scratch, all it takes is the right people.

 

[Co BIY]

I think the size of the potential market is key to justify the investment.

This is why China's efforts for silicon independence are a lot more likely to succeed. A lot larger market and greater global market penetration.

Russia can probably find more "bang" for their strategic "buck" elsewhere.

 

[Endymio]

In an attempt perhaps to justify the misleading headline, the author left out one of Zhbankov's most salient conclusions: that the issues largely didn't appear due to the CPU themselves, but rather in the assembly of the particular servers involved in the testing.

 

[VforV]

In an attempt perhaps to justify the misleading headline, the author left out one of Zhbankov's most salient conclusions: that the issues largely didn't appear due to the CPU themselves, but rather in the assembly of the particular servers involved in the testing.

If that is true, it would only follow the trend of unprofessional tech press we've seen for some years now...

 

[pavel.mateja]

They should throw away their proprietary VLIW-like architecture and go RISC-V way.
They might even be able to sell some outside of Russia and SW support would be on different level.

 

[DavidC1]

Surprise, surprise.

The original talk of Elbrus came during the time when Intel was still selling various Itanium chips back in 2004! And they are still on 28nm. 3x slower for 1/3rd the cores, 1/6th threads and 2/3rds the Base Xeon clocks and lacking Turbo is really not bad at all.

Newsflash: CPU from 2018 outperforms CPU from 2004 significantly! Pentium 4 beats Pentium II! Bulldozer beats Core 2!

It only took AMD 5 years and a shoestring budget to design the Zen architecture from scratch, all it takes is the right people.

The architecture doesn't seem that terrible based on core count/clocks and how it performs.

The real issue is not getting on at least the 14nm process and having more cores than latest Intel/AMD consumer desktop chips.

 

[ZorBG]

Surprise, surprise.

The original talk of Elbrus came during the time when Intel was still selling various Itanium chips back in 2004! And they are still on 28nm. 3x slower for 1/3rd the cores, 1/6th threads and 2/3rds the Base Xeon clocks and lacking Turbo is really not bad at all.

Newsflash: CPU from 2018 outperforms CPU from 2004 significantly! Pentium 4 beats Pentium II! Bulldozer beats Core 2!

Tom's Hardware was very good site back in the day. Very professional, accurate, and above all didn't mix politics with hardware.

I am sad to see some 8-core trial processor is compared with Intel 2K USD 40-core madness... No wonder its slower...

 

[husker]

Tom's Hardware was very good site back in the day. Very professional, accurate, and above all didn't mix politics with hardware.

I am sad to see some 8-core trial processor is compared with Intel 2K USD 40-core madness... No wonder its slower...

You seemed to miss the point of the article. This wasn't a Tom's Hardware review, but it simply a report on what happened. In Russia. By Russians. If the Russians* are comparing mismatched processors, well then darn those pesky Russians*!
The mention of politics (Crimea, Ukraine) in the article are not introduced by Tom's as some kind of statement, but merely serve to inform the reader as to why the Russians* are developing a new processor in the first place.

* Note: No disrespect is meant against Russia, Russian citizens, or Russian companies. Some of my favorite hockey players are Russian.

 

[Icho]

They should throw away their proprietary VLIW-like architecture and go RISC-V way.
They might even be able to sell some outside of Russia and SW support would be on different level.

Changing, why? Finding that a 70W processor is twice slower than a 125W one means parity in "performance/Watt". And when one achives parity with 28nm process vs Intel's 14 nm one, that is telling big things about the archtecture. Only the Java results are disapointing here ... So VLIW seems sound approach here.

 

[blazed]

After Intel and AMD have stopped selling to Russia due to the Ukraine invasion, I'm curious Sberbank won't have a lot of options to purchase CPUs for enterprise work. Will they reconsider? Even if they considered black market for those chips, they wouldn't be able to meet the demand they require, these companies even stopped selling PC/server CPU/APU to Russia. Its interesting to see if these companies preventing the tech sale to Russia will give new life to Elbrus?

 

[upiehu]

This is a strange article.

The Russian-language reporting doesn't contradict what's written here, but the positive aspects of the test are missing from Tom's coverage. This includes the fact that the bank was surprised to receive a working chip, and not just a design proposal, which means that progress is far ahead of schedule.

This article also says that chips like Elbrus can be used to build large cluster systems where "scalability and energy efficiency are not a concern" but that the chip may be "good enough for office workloads". First of all, if this is a slow chip, its worst application would be to office workloads since they only benefit from single-system performance. Meanwhile clusters where "scalability and energy efficiency are not a concern" simply do not exist, but low-performing chips often excel in these applications.

The result makes this feel like a Russophobic hit piece, or ill informed at best.

I admire every government that's ensuring its own sovereignty by decoupling critical infrastructure from the semiconductor industries of its (potential) adversaries.