Intel has released its fastest consumer desktop processor ever: the Core i9-9900KS, which has all eight cores clocked at 5,0GHz. There is a lot of noise around the new processor, but not everyone knows that the company already has a processor with a clock speed of 5,0 GHz, moreover, with 14 cores: Core i9-9990XE. This extremely rare item is not available to ordinary consumers: Intel only sells it to selected partners, and only through an auction, once a quarter, and without any guarantees from its side. How much would you pay for such luxury? Well, we managed to get hold of one of these monsters to see how good it is.
Build it and they will come (one of AFI's 100 famous American movie quotes in 100 years)
Core i9-9990XE is the pinnacle of Intel's 14nm process technology, a certain limit of possibilities. By the way, Intel can neither guarantee how many processors it will be able to produce, nor provide support in any way. Unlike other mainstream processors, there is no such thing as "EOL". If you win a processor at auction, you will pay prohibitively expensive for it, because that's the whole point of bidding. Getting 14 cores running at 5,0 GHz is a lot worth joining the "financial race" for.
This processor is part of the high-end family and runs on some X299 motherboards. It's also a Core i9, not a Xeon, which means only four memory channels and no ECC support. Technically, it supports overclocking. This is a processor for only one market, and that market is willing to spend a lot of money to take advantage of lower millisecond latency: high-frequency trading.
At the first auction, we initially knew about the three companies that were supposed to take part in it. For those who decided to participate in the auction, the closed auction remained a mystery: it was known only what equipment was offered by Intel, and not the number of units. Of the three companies we spoke to, only one was present without placing any bids, the second got three processors, and the third got the rest. The number of lots and the amounts spent on them are unknown.
High-frequency trading systems do not shy away from exotic arrangements. I've heard stories of companies spending tens of millions to implement line-of-sight lines of microwave transmitters to reduce latency by 3 milliseconds. All major financial traders have servers located as close to the exchange as possible, because the speed of light through optical cable is still not fast enough for them. These companies not only pay for the hardware, but also pay experts and technicians to set up these systems with low latency. This means tweaking the memory, overclocking the CPU, and even implementing custom cooling to get a completely stable yet as fast system as possible.
So how much are these people going to pay for a pre-overclocked 14-core 5,0GHz processor? Some of them may run above this level, as the stock Core i9-9980XE off the shelf can potentially run at this speed. We ended up getting a response from CaseKing, the recipient of the majority of the Core i9-9990XE: $2800. In fact, the price has gone up to $2850 since then. Not much compared to the Core i9-9980XE ($1979) or the recently announced Core i9-10980XE ($999) and of course traders will easily spend $1000-2000 more on the lowest latency x86 processor on the market.
So where to start? We have a sample processor. Technically, we have a whole system from International Computer Concepts, or ICC. They are server experts. We first met them at Supercomputing 2015 where they presented a crazy tower system with 8 different servers. ICC works closely with Intel to provide specific solutions for various market verticals: Oil & Gas, Medical, Computing. And, very importantly, for the financial segment, where they can sell a system overclocked to the limit.
Unfortunately, due to some proprietary technology, we are unable to show you the inside of the server that was sent to us. This is a standard 1U design with an ASUS X299 motherboard inside and 32GB of custom memory. To keep the hot Core i9-9990XE under control, a liquid cooling system (all copper) is used. For most processors, such cooling is absolutely unnecessary. It's a 1U form factor system, which means 1,75 inches tall (4,45 cm), meaning the need to house this monstrous processor requires top-class cooling, and ICC doesn't skimp here. An important point: the system is very noisy. It is unlikely that she will become a comfortable neighbor because of too loud work. More details later in the current review.
In addition to the standard specifications, ICC has made additional changes to the BIOS to ensure minimal latency and stability. Again, we can't reveal all the details, but we didn't update the BIOS for our testing. The 1U server has space for two graphics cards, two M.2 drives, four SATA drives, and comes with a 1200W power supply. We have paid some attention to the power consumption overview below.
Be careful not to drop
At first glance, the Core i9-9990XE is a standard LGA2066 chip. It uses the usual 18-core "HCC" Intel Skylake silicon, but is targeted at the "consumer" platform as part of Intel's product segmentation strategy. The processor does not support error correction memory and is therefore limited to 128 GB of standard DDR4 memory, although you can be sure that any HFT system using this processor will work with high speed memory. The chip has 44 PCIe 3.0 lanes like the other LGA2066s, and since it's not a Xeon, it doesn't support RAS or vPro management features.
This is where one of the problems with this chip comes in: the price tends to be of interest to professional users who need internal management features and other security features to keep their expensive equipment safe and manageable. By labeling the processor as Core i9 rather than Xeon W, Intel is discounting these offers: OEMs who buy and resell the part to end users will have to explain that this rare chip has some limitations.
At the moment, we don't know how many chips Intel is going to bring to market. Intel Corporation holds quarterly auctions where it offers chips that have passed all the tests. Assuming that all OEMs want to buy them for their customers, we can talk about no more than 100 units per year. Because of these product-or-not-product nuances, the Core i9-9990XE has not received its own page in the Intel processor database, and it will never fall under the end-of-life program, as it is not subject to standard ordering and delivery process. All long-term support for the processor ends up in the hands of the company or OEM that buys them.
Chip and our tests
Essentially, the Core i9-9990XE is a 14-core processor with a base frequency of 4,0GHz and a TDP of 255W at that frequency. The turbo frequency of this processor is 5,0 GHz on all cores. But this poses a bit of a problem when classifying a processor as "all cores clocked at 5,0 GHz".
In our interviews with Intel representatives, we talked about how the turbo should be turned on: how the system turns on the turbo mode depends on the instructions used and on the motherboard manufacturer. Turbo is determined by the higher level power limit (PL2) and the turbo budget time (Tau). Typically, Intel "suggests" turbo consumption 25% higher than the stated TDP (i.e. for a TDP of 255W, the consumption will be 319W), and from 8 to 200 seconds of turbo depending on the platform.
On the 1U server we were given for testing, ICC enabled the turbo in "unlimited power for unlimited time" mode (technically up to 4096 seconds, I believe) as they want the CPU to run at 5,0 GHz at all times on all cores. This, as mentioned above, requires very efficient cooling. A star challenge for ICC, given the 1U form factor, a proprietary cooling technology has been developed to solve it.
Technically, this chip supports Turbo Max 3.0, thanks to which Intel determines the most powerful cores for applying even higher turbo frequencies. In our case, out of 14, the 10th core was determined to be the best. On Windows, the ACPI interface detects the key software (or identifies it by the active window) and will attempt to run it on these "better" cores with an additional boost (+100 MHz or so). As far as our system is concerned, since the TBM3 and ACPI interface fixed the software on certain cores, we did not see an increase in frequency due to this way of configuring the system. One of the key features for users of ICC systems is consistent low latency. In order not to change this consistency, TBM 3.0 does not affect processor frequencies in our testing.
Other features of the chip are support for quad-channel DDR4-2666 memory in single rank mode. ICC shipped our system with custom memory modules and matching heatsinks, and the system ran DDR4-3600 CL16. This chip also has 44 PCIe 3.0 lanes, like other Intel HEDT 9-series processors.
The Core i9-9990XE is simply surrounded by competition.
The first of these is the upcoming Core i9-9900KS, an octa-core processor that supports all eight cores at 5,0GHz. This chip uses standard consumer silicon and therefore only has two memory channels and 16 PCIe 3.0 lanes.
The other competitor is Cascade Lake-X's new 18-core flagship, the Core i9-10980XE, which is priced at $999. This is the latest high performance desktop processor with (we believe) the latest security patches from Intel, as well as some clock speed improvements over the Core i9-9980XE. It ends up having four more cores than the 9990XE, but lower frequencies and is cheaper. A user lucky enough to get a good sample can overclock it to the 9990XE. The Core i9-10980XE has four more PCIe 3.0 lanes and the same number of memory channels.
From AMD's side, the 16-core Ryzen 9 3950X coming out in November is just the first of the competition. Being built at 7nm it's certainly more power efficient and the Zen 2 microarchitecture has a higher IPC than Intel chips, but the CPU won't be able to reach the same high frequencies. It is designed for home PCs, and is equipped with 24 PCIe 4.0 lanes and two memory channels. Given an MSRP of $749, it will certainly cost a lot less than an Intel processor.
It is necessary to pay attention to the launch of AMD's new generation of Threadripper, based on the same Zen 2 and 7 nm. We don't have many details at the moment other than that AMD has announced that the line will be coming out in November and the line will start with a 24-core processor. It is expected to have four memory channels, 64 PCIe lanes and be able to run at around 4,0GHz. It still won't be able to reach as high clock speeds as Intel's, and its price and power consumption are still unknown.
In addition, AMD has released the EPYC 2 series Zen 7002 server hardware. Instead of looking at a high-frequency 14-core CPU, users can consider a 32-core CPU with eight memory channels, high IPC, and 128 PCIe 4.0 lanes. Again, they will face a frequency deficit, which is a very important parameter for HPC traders. The EPYC 7502P sells for around $3400 and on the right server, this could be an option if an HPC trader needs to scale.
Whatever we compare it to, it's undeniable that the Core i9-9990XE pushes the boundaries of Intel's 14nm process. That's why it doesn't have an MSRP, and why Intel can't predict how many will be released next quarter. It's not for nothing that CaseKing put it up for sale (with a one-year OEM warranty) for €2849, because it's way above any other Intel desktop processor, and for good reason.
Our test bench
It should be noted right away that the latest Intel updates regarding Specter, Meltdown and ZombieLoad may affect performance. Based on the data we got from Intel, security mitigation measures are the least damaging to the latest hardware (compared to, say, Broadwell). The system provided by ICC does not have built-in protections in the firmware, however, we used an OS version that had some software security patches applied. ICC has made it clear that some of its clients, while concerned about these issues, often just want the fastest system possible, depending on how they use those systems.
As a result, our results do not match our previous reviews. Due to the fact that it uses a custom BIOS with overclocking options locked, the benchmark data will not necessarily reflect the performance of a freshly purchased processor on your home PC, but rather will demonstrate its performance on a system built specifically for it, which is ultimately the expected use for these chips. As a result, we put an asterisk next to our results to indicate that the test environment for this chip was different.
- CPU: Intel Core i9-9990XE, 14 Cores, 4.0 GHz Base, 5.0 GHz Turbo, 255W TDP, $Auction
- DRAM: 4×8 GB Custom ICC Modules, DDR4-3600 CL16
- Motherboard: ASUS X299
- GPU: Sapphire Radeon RX460 2GB
- Cooling: ICC Proprietary Liquid Cooling
- Power Supply: Dual 1200W 1U Redundant Supplies
- Storage: Micron MX500 1TB SSD
- Chassis: 1U Rack Server
In our reviews, we usually test outdoors, with powerful cooling, a high-end motherboard, DRAM at manufacturer-supported frequencies, and the latest public BIOS version for this motherboard.
For tests, we used our standard set of processors. Due to the 1U form factor, and the nature of the application of this chip, we did not use a large graphics card for gaming tests. Users who want to get this system and link it to a large CUDA map for financial modeling will most likely have to work hard. And for games, it's best to wait for the release of the Core i9-9900KS.
Contents of this review:
- Analysis and competition
- Core i9-9990XE: Compilation Champion
- CPU Performance: Rendering Benchmarks
- Processor Performance: Encoding Tests
- Processor Performance: System Benchmarks
- Processor performance: office benchmarks
- CPU Performance: Web Tests and Legacy Tests
- Power consumption and thermal properties
- Conclusions and closing words
Compilation Champion, Windows VC++ and Chrome
Many of AnandTech's readers are software developers who are interested in hardware performance on the tasks they are used to. While compiling the Linux kernel is the "standard" for reviewers who compile frequently, our test is a bit more variant - we use the Windows instructions to compile Chrome, more specifically the March 56 build of Chrome 2017, as it has been since this test was first used. Google gives pretty detailed instructions on how to compile under Windows, along with the ability to download 400k files for the repository. This is by far one of our most popular benchmarks and is a good measure of core performance, multithreading performance, and memory access speed.
In this test, following Google's instructions, we use the MSVC compiler and Ninja developer tools to manage compilation. As you might expect, this test is multi-threaded, with some DRAM dependency, where faster caches provide an advantage. The data obtained as a result of testing is the compilation time, which we convert into the number of compilations per day. The test takes from one hour on a fast, high-performance desktop processor to several hours on the slowest PCs.
In this test, two processors fought for supremacy almost on par: the 16-core Ryzen Threadripper 2950X and the 8-core i9-9900K. Armed with six more cores, a much higher frequency, and an additional two memory channels, the Core i9-9990XE easily passes this test, compiling in 42 minutes and 10 seconds, and is the only processor to break the 50-minute mark (not to mention 45-minutes). minute).
Render Tests
In a professional environment, rendering is often the main CPU workload. It is used in a variety of formats, from 3D rendering to rasterization, in applications such as games or ray tracing, and uses the software's ability to manage meshes, textures, collisions, aliases, and physics (in animation). Most renderers offer code for the CPU, while some of them use GPUs and choose environments that use FPGAs or specialized ASICs. However, for large studios, processors are still the main hardware.
: 3D Creation Suite
A high-end rendering tool, Blender is an open source product with many settings and configurations used by many high-end animation studios around the world. The organization recently released a Blender test suite, a couple of weeks after we decided to reduce the use of the Blender test in our new suite, however the new test can take over an hour. To get our results, we run one of the subtests in this package via the command line, a standard "bmw27" scene in CPU-only mode, and measure the render completion time.
Blender knows how to take advantage of more cores, and while the 9990XE's frequency wins over the 7940X, it's not enough to outperform 18-core hardware.
LuxMark v3.1: LuxRender via various code paths
As stated above, there are many different ways to handle render data: CPU, GPU, Accelerator, and more. In addition, there are many frameworks and APIs in which one can program, depending on how the software will be used. LuxMark, a benchmark developed using the LuxRender engine, offers several different scenes and APIs.
In our test, we run a simple "Ball" scene on C++ and OpenCL code, but in CPU mode. This scene starts with a rough render and slowly improves over the course of two minutes, resulting in a final score that can be called "an average of thousands of rays per second".
We see a slight performance lag compared to the 7940X, which is interesting. I wonder if the 2,4GHz fixed mesh is the limiting factor?
: ray tracing
The Persistence of Vision ray tracing engine is another well-known benchmarking tool that was dormant for a while until AMD released its Zen processors, when all of a sudden both Intel and AMD started pushing code to the main open source branch of the project. For our test, we use the built-in test for all cores, called from the command line.
Encoding Tests
With the growth of streams, vlogs and video content in general, encoding and transcoding tests are becoming increasingly important. Not only are there more and more home users and gamers involved in converting video files and video streams, but servers processing data streams need on-the-fly encryption, as well as log compression and decompression. Our coding tests target such scenarios and take into account community feedback to ensure the most up-to-date results.
Handbrake 1.1.0: streaming and archival video transcoding
A popular open source tool, Handbrake is a software for converting videos in any way possible that is, in a sense, a benchmark. The danger here lies in version numbers and optimizations. For example, recent versions of software may take advantage of AVX-512 and OpenCL to speed up certain types of transcoding and certain algorithms. The version we are using is pure CPU work, with standard transcoding options.
We divided Handbrake into several tests using recording from a Logitech C920 1080p60 native webcam (essentially a stream recording). The recording will be converted to two types of streaming formats and one for archiving. Used output options:
- 720p60 at 6000 kbps constant bit rate, fast setting, high profile
- 1080p60 at 3500 kbps constant bit rate, faster setting, main profile
- 1080p60 HEVC at 3500 kbps variable bit rate, fast setting, main profile
Our encoding tests require a good balance of cores and frequency, with 14-core 5,0GHz hardware easily outperforming the 7940X and showing that having 28 cores isn't always a winner.
7-zip v1805: popular open source archiver
Of all our archive/unzip tests, 7-zip is the most requested, and has a built-in benchmark. We added the latest version of this software to our test suite, and we run the benchmark from the command line. The results of archiving and unzipping are displayed as a single total score.
This test clearly shows that modern multi-die processors have a big difference in performance between compression and decompression: they perform well in one, and poorly in the other. In addition, we have active discussions about how the Windows Scheduler implements each thread. When we get more results, we will be happy to share our thoughts on this.
Please note, if you plan to publish the compression data anywhere, please include the results of the decompression as well. Otherwise, you will provide only half of the result.
The presence of 28 cores shows its strength here, and the additional frequency cannot tip the scales.
WinRAR 5.60b3: Archiver
When I need a compression tool, I usually choose WinRAR. Many users of my generation used it over two decades ago. The interface has not changed much, although the integration with Windows right-click commands is a nice plus. It doesn't have a built-in benchmark, so we run compression on a directory containing over thirty 60-second video files and 2000 small web files at normal compression speed.
WinRAR has variable multithreading and is caching intensive, so in our test we run it 10 times and average the last five runs to test only CPU performance.
WinRAR is one of the tests with variable multithreading, so the combination of number of cores and frequency is important here. Interestingly, the 9990XE, despite its higher frequency, is slightly slower than the 7940X. It is possible that the extra power required to overclock the cores to peak frequency may cause additional delays when working with a large number of small files.
AES Encryption: File Protection
Some platforms, especially mobile devices, encrypt file systems by default to protect content. Windows-based devices often use BitLocker or third-party software for encryption. In the AES encryption test, we used the discontinued TrueCrypt benchmark, which tests several encryption algorithms directly in memory.
The data generated from this test is the combined performance of AES for encryption/decryption, measured in gigabytes per second. The software uses AES instructions if the processor allows it, but does not use AVX-512.
Some ads 🙂
Thank you for staying with us. Do you like our articles? Want to see more interesting content? Support us by placing an order or recommending to friends, , 30% discount for Habr users on a unique analogue of entry-level servers, which was invented by us for you: (available with RAID1 and RAID10, up to 24 cores and up to 40GB DDR4).
Dell R730xd 2 times cheaper? Only here in the Netherlands! Dell R420 - 2x E5-2430 2.2Ghz 6C 128GB DDR3 2x960GB SSD 1Gbps 100TB - from $99! Read about
Source: habr.com