What is the hierarchy of Intel processors

New processors from AMD and Intel

The third generation of AMD Ryzen processors scratches Intel's Core-i dominance several times: For the same money, there are twelve instead of eight cores. And higher clock frequencies than before, a level 3 cache twice as large and modernized Zen 2 arithmetic units with doubled floating point performance should raise the single-threading performance to the level of the fastest Intel Core i9-9900K chip to date. In addition, AMD gave the entire AM4 platform a makeover. The Ryzen 3000 processors and the new X570 chipset together provide up to 36 PCIe 4.0 lanes for graphics cards and fast SSDs.

From 7.7. According to AMD CEO Lisa Su, there will be five Ryzen 3000 CPUs to buy. The date refers to the 7 nm technology TSMC uses to manufacture the chips. AMD promises significantly higher efficiency: The 65-watt eight-core Ryzen 7 3700X therefore completes the Cinebench R20 rendering benchmark with a single thread by 15 percent and with all 16 threads by 18 percent faster than the previous AM4 top model Ryzen 7 2700X with 12 -nm technology and 105 watt thermal design power. For the time being, AMD is keeping further benchmark values ​​of the third Ryzen generation under lock and key.

Heated PCI Express 4.0

The Ryzen 3000 will run on many current mainboards with chipsets such as X470 and B450, provided that the respective manufacturer provides a BIOS update. The novelty PCI Express 4.0 (PCIe 4.0) will not be usable with older boards, but only with new ones with the X570 chipset. It also has PCIe 4.0 lanes, the predecessors are still stuck with PCIe 2.0.

The manufacturers Asrock, Asus, Gigabyte and MSI already showed numerous X570 mainboards at Computex. With the exception of expensive luxury boards with gigantic passive or water coolers, all are equipped with chipset fans. The culprit is the PCIe 4.0 controller in the X570, which brings the throughput to 2 GB / s per lane, but unfortunately also increases the power consumption from 5 to 11 watts compared to the X470. MSI explained that the fan on their boards only runs when needed.

In addition to PCIe 4.0, AMD has also heavily bored out the USB ports on the X570 and Ryzen 3000. Both the four ports on the CPU and eight of the twelve on the chipset now work in USB 3.2 Gen 2 mode at 10 Gbit / s. For USB 3.2 Gen 2x2 with 20 GBit / s you need additional chips like the Asmedia ASM3242, but mass production will not start until the end of the year. With comparable equipment, the X570 boards will cost around 50 euros more than the X470 boards. The latter will therefore still be available to buy as well as cheaper ones with B450. Beta firmware updates will be available for some, but not all, X370 and B350 motherboards. Older A320 boards probably don't get BIOS updates for Ryzen 3000 because their BIOS flash chips don't have enough storage capacity.

Intel announced the "Special Edition" Core i9-9900KS. In contrast to the Core i9-9900K, the eight-core turbo can run on all cores with 5 GHz. When asked, Intel is silent about Thermal Design Power as well as about the exact start of sales, which is only planned for the fourth quarter. But Intel has announced a plethora of other processors, including the Xeon E-2200 series workstation chips. The LGA1151v2 processors are closely related to the Core i-9000, but also control ECC RAM and only run on motherboards with a C240 ​​chipset. For business computers, in turn, Intel has also released 14 desktop and mobile CPUs of the ninth Core i generation for systems with the remote maintenance and management functions vPro. In the fall, Intel promises faster Core X CPUs with a higher clock rate and improved turbo for the high-end platform LGA2066. These processors are probably descendants of the current Xeon SP server chips Cascade Lake.

The surprising criticism of some benchmarks also shows that Intel's nerves are a bit bare: The popular Cinebench R20, which comes from the Cinema 4D rendering program and uses many cores and threads, allegedly does not reflect the performance of much more frequently used ones, according to Intel "Real-World" applications. In addition, Intel accuses AMD of not using the optimal compiler settings for the Intel system when comparing Epyc "Rome" with Xeon SP.

10 nm in the notebook

After a delay of several years, Intel has finally finished the 10 nm Ice Lake-U mobile processors. For the quad-cores with 9, 15 and 28 watt TDP, the chip manufacturer has made major changes to the architecture of the computing units for the first time since Skylake. The "Sunny Cove" cores can process more commands in parallel per clock step and receive the command set extension AVX-512, which was previously reserved for server processors such as Xeon SP and the expensive Core X. The AVX-512 DL Boost function accelerates AI algorithms.

The computing power should increase by 18 percent per cycle. In practice, however, the performance increase should often fizzle out, since Ice Lake reached a maximum of 4.1 GHz in turbo, while the predecessor Whiskey Lake (Core i-8000U) managed up to 4.8 GHz. Intel hides benchmark values ​​for the CPU as well as product names and clock frequencies.

The first product announcements with Ice Lake-U have already been made: Dell is installing the quad-core in the XPS 13 2-in-1 notebook, the display of which can be folded 360 degrees to turn it into a tablet. It will be available from July for 1000 euros. The XPS 13 2-in-1 also meets all the requirements of Project Athena. In order to participate in this marketing program from Intel, the notebooks have to comply with a few requirements: These include at least 16 hours of runtime when playing video at 150 cd / m2 Display brightness and nine hours of productive work with 250 cd / m22 as well as permanent internet access with WLAN or LTE. In addition, refueling with electricity for four hours of running time must not take longer than 30 minutes. The minimum hardware requirements include a Core i5 or i7, 8 GB of RAM, Thunderbolt 3 and an SSD with 256 GB or more. The display size must not exceed 15.9 inches.

Fast PCIe 4.0 SSDs

In line with the revised AM4 platform with Ryzen 3000, SSD manufacturers showed prototypes of upcoming products with PCIe 4.0 controllers, whereby the time frame was consistently "end of the year". Depending on the contact person, this meant everything from “from August / September” to “hopefully in December”. The Phison controller PS5016-E16 is used on the Corsair Force MP600, the PNY CS4040 and at Klevv / Essencore. At the booth, Adata showed a pre-production model with Silicon Motion SM2267 and a maximum capacity of 8 Tbytes.

For the pre-production samples mentioned, all manufacturers name transfer rates between 4 and 5 GByte / s - significantly less than just under 8 GByte / s, which would theoretically be possible with PCIe 4.0 x4. Phison is already working on the faster PS5019-E19, which should achieve more than 6 GB / s. Micron (Crucial) is also working on its own PCIe 4.0 controller, which will appear on the first SSDs at the turn of the year. Significantly more than 5 GB / s can currently only be achieved by bundling several SSDs, as demonstrated by Gigabyte on a plug-in card for an x16 slot: Four Gen 4 SSDs achieved in a Ryzen 3000 system with an X570 chipset impressive 15 GB / s.

Navi GPUs with a new architecture

At the beginning of the Computex opening speech, AMD boss Lisa Su finally presented more information about the upcoming Navi GPUs: Officially, the new graphics card series will be called the Radeon RX 5000, based on the company's 50th anniversary. It is based on a new architecture called Radeon DNA (RDNA). This should become the gaming engine for the next decade, for example in the Playstation 5.

Compared to the previous “Graphics Core Next” (GCN) design, AMD has streamlined the graphics pipeline in order to achieve higher clock rates, introduced a new multi-level cache hierarchy and revised the compute units. Like the Radeon VII, the new chips are manufactured with a structure width of 7 nm. In addition, they will use GDDR6 memory and already master PCIe 4.0.

Compared to an unspecified Vega card, computing power per cycle increases by 25 percent and energy efficiency by 50 percent. A Radeon RX 5700 is said to be 10 percent faster than a GeForce RTX 2070 in the game Strange Brigade. That gives hope for real competition.

Su did not mention specific data such as the number of shader cores, clock frequency, memory expansion, power requirements and price - they will only be revealed on June 10th at the E3 computer game fair. The actual launch should not take place until July. At the Asrock press conference, however, the product manager slipped out August as the delivery date. Presumably he was referring to in-house card designs. This process also fits what we have heard from other manufacturers.

In the absence of specific data - even the card manufacturers do not yet have their own GPU samples - the rumor mill was seething. One of our contacts said that the first generation called Navi 10 is not yet a complete RDNA design, but still contains a lot of GCN, not least to ensure compatibility with older games. Only Navi 20 will completely rely on RDNA, but then deliver a powerful performance boost. Expectations regarding the price structure ranged from “the same price as the 2070 with a slightly higher performance” to “must not cost more than 399 dollars”.

It was agreed that Nvidia is already preparing the counter blow in the form of the GeForce RTX 2060 Ti and RTX 2070 Ti. In addition, the company will revise its existing portfolio: A new GPU called TU104-450 will drive a card that fits between the GeForce RTX 2080 and 2080 Ti. In addition, the TU104-400 and -400A are being replaced by the TU104-410. There are no restrictions for the manufacturer with regard to the factory overclocking, which Nvidia previously reserved for the more expensive TU104-400A.

It was more specific with the notebooks, because Nvidia has presented mobile workstation GPUs for this. Quadro RTX 5000, RTX 4000 and RTX 3000 are the counterparts to the consumer models GeForce RTX 2080, RTX 2070 and RTX 2060, while the Quadro T2000 and T1000 are close relatives of the GTX 16 series without RT and tensor cores. (chh)

This article is from c't 13/2019.