Martyn Williams/IDG

Former chief engineer of Intel told what is wrong with the company, and how to fix it


In a very straightforward video posted recently, former Intel chief engineer François Piednoël gave his advice on how to “fix” Intel processors, criticized the current management for not being engineers, said the AVX512 was a mistake. and said that the company is very lucky that AMD has not yet captured even more market share.

Martyn Williams / IDG

First, Intel really not focused He said in an almost hour-long video presentation. – Today’s leaders Intel – not engineers, they do not understand what needs to be designed for the market “… He noted that Intel’s tech solutions since 2016 have been pretty much bad.

François Piednoel left Intel in 2017 after spending 20 years as a chief engineer and performance architect. He has a hand in processors ranging from the Pentium III to the 6th generation Core i7. The engineer often gave technical presentations and press demos on hardware, passionately arguing why certain design decisions Intel made were correct. He is currently employed by Mercedes-Benz Research and Development North America as Chief Architect.

Gordon Mah Ung / IDG

Mr. Piednoel admitted that his information about Intel’s internals was essentially “outdated” for years. At the same time, everything he says is not related to non-disclosure agreements. Most of the analysis concerns public information circulating around Intel. Those interested can watch the video in full, but we will highlight the most interesting statements.

AVX512 is an error

The AVX512 instruction set is the foundation for the DL Boost AI machine learning acceleration that Intel uses in its Xeon server processors. And this technology has been pushed into consumer processors, including 10th Gen Ice Lake laptop chips. François Piednoel categorically considers the inclusion of the AVX512 in consumer chips a mistake.

Intel weren’t just Skylake and Skylake X, – he noted. – Instructions AVX512 are designed for the bandwidth competition that graphics accelerators have already lost. There are two ways to increase your bandwidth. One of them is to expand the length of vectors processed by the kernel (in AVX-512 – up to 512 bits), and the other – by increasing the number of cores. “

After Intel’s misadventures with the Pentium 4 (the architecture required too many optimizations), Mr. Piednoel once said that the company concluded that it was impossible to recompile the whole world, seemingly bearing in mind that Intel is not very strong in software.

“Today, the state of software is not conducive to expanding the length of vectors, – says the engineer. – In fact, you can clearly see, for example, in Cinebench (can’t be called my favorite test, especially for laptops where it doesn’t make any sense, but nonetheless) that AMD wins the battle for performance. This is because they have more cores and they can afford to have more cores. “… For Mr. Piednoel, who has been competing with AMD for two decades, this recognition must be especially painful.

Francois piednoel

Dadi Perlmutter Pearlmutter, head of the Intel Architecture Group) understood that long vectors in consumer electronics like laptops are bad: 1) increased power consumption if implemented correctly; 2) almost complete absence of software using AVX512 despite the fact that it is necessary to create larger cores; 3) suitable for bandwidth tests – who needs all this on laptops? “, – noted the specialist.

The decision to use the AVX512 in consumer processors resulted in larger die sizes and higher power consumption. For those who don’t know, Intel CPUs drastically drop clock speeds under AVX512 workloads. This is probably not news to those who have heard what famous Linux creator Linus Torvalds said last month regarding Intel’s approach to promoting AVX512: “I hope that AVX512 will die in agony, and Intel will begin to fix real problems instead of trying to create magic instructions that will then allow their processors to perform favorably in specially designed tests. “

Intel lost concentration

Another mistake that Intel made, according to the engineer, is that it was distracted from the main business of manufacturing fast processors. Over the last half of the decade, Intel continued to actively diversify its activities, with the result that the company was unable to focus on its core processor business. This allowed the competitor AMD to catch up with it, and today the only thing that saves Intel from losing a more significant market share is AMD’s limited ability to produce its popular processors.

Intel very lucky that AMD cannot obtain more significant production volumes in order to be able to compete at full strength, – says Mr. Piednoel. – If they had increased the volume, the difference in price would definitely be worth Intel a much larger market share than it is today. Intel is fortunate that AMD has production capacity constraints so they can’t grab market share fast enough. We had the same thing when AMD had the Athlon 64, and we were mainly trying to catch up with them between Pentium 4 and Conroe. “… According to Mercury Research, AMD has already captured 20% of the laptop market – a record in the company’s history.

Francois piednoel

Indeed, AMD surpassed Intel in performance when the Pentium 4 with Netburst architecture fell behind a competitor. For several years, AMD chips retained their leadership crown. It was only with the release of Intel Core 2 processors with Conroe architecture in 2006 that Intel regained the crown of performance and did not lose it until AMD’s resurgence with the advent of Ryzen in 2017. “Upcoming Intel Tiger Lake chips will delay the problem, but won’t solve it,” – said Piednoel.

Ryzen’s multi-threaded mode looked good due to low single-threaded performance

During his talk, François Piednoel touched on the technical details of the basics of the Skylake-based Intel core, noting that the architecture was essentially designed for high single-threaded performance and then improved for multi-core and multi-threaded processors.

He praised AMD for its excellent throughput, but Zen 1 cores have their issues. For example, the original Ryzen processor offered much larger performance gains when multithreading was enabled than Intel did. “People don’t understand that the performance gains in multithreading mode are inversely proportional to out-of-order performance,” – noted the engineer: if Zen 1’s performance in this area were as efficient as Intel’s, multithreading would not give such a serious return.

Xeon should get rid of unused blocks

Intel Xeon chip designs now use the same cores as other applications, from compute-intensive supercomputers to plain old web servers and virtual machines. According to François Piednoel, such completely different areas of application lead to the fact that about 10% of the chip area is simply not used. This valuable die area could be used for additional cores to better compete in markets that do not require, for example, AVX512 instructions.

However, this requires flexibility from Intel, which the company currently does not have: it strives to use the same core everywhere. Mr. Piednoel believes that dedicated teams should be allocated to develop specialized, niche versions of Xeon for narrow applications, rather than making the most universal kernels.

Francois piednoel

Engineers should play first fiddle, not managers

François Piednoel spared no words on criticism of Intel culture, which, according to him, has changed dramatically and now puts economic managers at the forefront, not technical specialists. This has led to a lack of innovation, aggressive plans and an exodus of engineers frustrated that only managers have growth prospects. Rather than reacting to challenges on the fly to counter AMD products, today’s Intel simply isn’t willing to fight back. Instead, he said, product launch plans are drawn up by executives with a master’s degree in economic management who are unable to adapt. Intel’s factories now continue to produce enough processors to support profitability, but brand credibility is slowly declining.

“I remember when the Pentium II was released, Andy Grove Grove) said something to the team. He said: “The Lord gave us only one stamp. Never harm her. ” I think the company is spoiling the brand right now. You have to get down to business and work hard to make the Extreme Edition chips truly exceptional and win. “– said the former chief engineer of Intel.

If you notice an error, select it with the mouse and press CTRL + ENTER.

Leave a Comment