Ryzen 9 3900X [in 15 benchmarks]
AMD
Ryzen 9 3900X
Buy
- Interface
- Core clock speed
- Max video memory
- Memory type
- Memory clock speed
- Maximum resolution
Summary
AMD started AMD Ryzen 9 3900X sales 7 July 2019 at a recommended price of $499. Based on a Matisse (Zen 2) architecture, this desktop processor is primarily aimed at home systems. It has 12 cores and 24 threads, and is based on 7 nm manufacturing technology, with a max turbo frequency of 4600 MHz and an unlocked multiplier.
Compatibility-wise, this is
AMD Socket AM4
processor with a TDP of 125 Watt. It supports
DDR4
Dual-channel memory.
It provides good benchmark performance at
26. 38%
of a leader’s which is AMD EPYC 9654.
EPYC
9654
Compare
General info
Ryzen 9 3900X processor market type (desktop or notebook), architecture, sales start time and pricing.
| Place in performance rating | 131 | |
| Value for money | 35.47 | |
| Market segment | Desktop processor | |
| Series | AMD Ryzen 9 | |
| Architecture codename | Matisse (Zen 2) (2019−2020) | |
| Release date | 7 July 2019 (3 years old) |
|
| Launch price (MSRP) | $499 | of 305 (Core i7-870) |
| Current price | $480 (1x MSRP) | of 25332 (Xeon Platinum 8276L) |
Value for money
To get the index we compare the characteristics of the processors and their cost, taking into account the cost of other processors.
- 0
- 50
- 100
Technical specs
Basic microprocessor parameters such as number of cores, number of threads, base frequency and turbo boost clock, lithography, cache size and multiplier lock state. These parameters can generally indicate CPU performance, but to be more precise you have to review its test results.
| Physical cores | 12 (Dodeca-Core) | |
| Threads | 24 | |
| Base clock speed | 3.8 GHz | of 4.7 (FX-9590) |
| Boost clock speed | 4.6 GHz | of 7.3 (Ryzen 9 7900) |
| L1 cache | 96K (per core) | of 1536 (EPYC Embedded 3401) |
| L2 cache | 512K (per core) | of 12288 (Core 2 Quad Q9550) |
| L3 cache | 64 MB | of 32 (Ryzen Threadripper 1998) |
| Chip lithography | 7 nm | of 4 (Ryzen 9 7940HS) |
| Maximum core temperature | 95 °C | of 110 (Atom x7-E3950) |
| Number of transistors | 19,200 million | of 9900000 (Ryzen 5 7645HX) |
| 64 bit support | + | |
| Windows 11 compatibility | + | |
| Unlocked multiplier | + |
Compatibility
Information on Ryzen 9 3900X compatibility with other computer components and devices: motherboard (look for socket type), power supply unit (look for power consumption) etc.
Useful when planning a future computer configuration or upgrading an existing one.
Note that power consumption of some processors can well exceed their nominal TDP, even without overclocking. Some can even double their declared thermals given that the motherboard allows to tune the CPU power parameters.
| Number of CPUs in a configuration | 1 (Uniprocessor) | of 8 (Opteron 842) |
| Socket | AM4 | |
| Thermal design power (TDP) | 125 Watt | of 400 (Xeon Platinum 9282) |
Technologies and extensions
Technological capabilities and additional instructions supported by Ryzen 9 3900X. You’ll probably need this information if you require some particular technology.
| AES-NI | + | |
| AVX | + |
Virtualization technologies
Supported virtual machine optimization technologies.
Some are specific to Intel only, some to AMD.
| AMD-V | + |
Memory specs
Types, maximum amount and channel number of RAM supported by Ryzen 9 3900X’s memory controller. Depending on the motherboard, higher memory frequency may be supported.
| Supported memory types | DDR4 Dual-channel |
of 5600 (Ryzen 9 7940HS) |
| Maximum memory size | 128 GB | of 786 (Xeon E5-2670 v3) |
| Max memory channels | 2 | of 12 (Xeon Platinum 9221) |
| Maximum memory bandwidth | 51.196 GB/s | of 281.6 (Xeon Platinum 9221) |
| ECC memory support | — |
Graphics specifications
General parameters of GPU integrated into Ryzen 9 3900X.
| Integrated graphics card | — |
Benchmark performance
Single-core and multi-core benchmark results of Ryzen 9 3900X.
Overall benchmark performance is measured in points in 0-100 range, higher is better.
Overall score
This is our combined benchmark performance rating. We are regularly improving our combining algorithms, but if you find some perceived inconsistencies, feel free to speak up in comments section, we usually fix problems quickly.
Ryzen 9 3900X
26.38
- Passmark
- GeekBench 5 Single-Core
- GeekBench 5 Multi-Core
- Cinebench 10 32-bit single-core
- Cinebench 10 32-bit multi-core
- 3DMark06 CPU
- Cinebench 11.5 64-bit multi-core
- Cinebench 15 64-bit multi-core
- Cinebench 15 64-bit single-core
- Cinebench 11.5 64-bit single-core
- TrueCrypt AES
- WinRAR 4.
0 - x264 encoding pass 2
- x264 encoding pass 1
Passmark
Passmark CPU Mark is a widespread benchmark, consisting of 8 different types of workload, including integer and floating point math, extended instructions, compression, encryption and physics calculation. There is also one separate single-threaded scenario measuring single-core performance.
Benchmark coverage: 67%
Ryzen 9 3900X
32741
GeekBench 5 Single-Core
GeekBench 5 Single-Core is a cross-platform application developed in the form of CPU tests that independently recreate certain real-world tasks with which to accurately measure performance. This version uses only a single CPU core.
Benchmark coverage: 37%
Ryzen 9 3900X
1276
GeekBench 5 Multi-Core
GeekBench 5 Multi-Core is a cross-platform application developed in the form of CPU tests that independently recreate certain real-world tasks with which to accurately measure performance.
This version uses all available CPU cores.
Benchmark coverage: 37%
Ryzen 9 3900X
11667
Cinebench 10 32-bit single-core
Cinebench R10 is an ancient ray tracing benchmark for processors by Maxon, authors of Cinema 4D. Its single core version uses just one CPU thread to render a futuristic looking motorcycle.
Benchmark coverage: 20%
Ryzen 9 3900X
6019
Cinebench 10 32-bit multi-core
Cinebench Release 10 Multi Core is a variant of Cinebench R10 using all the processor threads. Possible number of threads is limited by 16 in this version.
Benchmark coverage: 19%
Ryzen 9 3900X
45539
3DMark06 CPU
3DMark06 is a discontinued DirectX 9 benchmark suite from Futuremark. Its CPU part contains two scenarios, one dedicated to artificial intelligence pathfinding, another to game physics using PhysX package.
Benchmark coverage: 19%
Ryzen 9 3900X
14889
Cinebench 11.5 64-bit multi-core
Cinebench Release 11.5 Multi Core is a variant of Cinebench R11.5 which uses all the processor threads. A maximum of 64 threads is supported in this version.
Benchmark coverage: 17%
Ryzen 9 3900X
34
Cinebench 15 64-bit multi-core
Cinebench Release 15 Multi Core (sometimes called Multi-Thread) is a variant of Cinebench R15 which uses all the processor threads.
Benchmark coverage: 14%
Ryzen 9 3900X
3049
Cinebench 15 64-bit single-core
Cinebench R15 (standing for Release 15) is a benchmark made by Maxon, authors of Cinema 4D. It was superseded by later versions of Cinebench, which use more modern variants of Cinema 4D engine. The Single Core version (sometimes called Single-Thread) only uses a single processor thread to render a room full of reflective spheres and light sources.
Benchmark coverage: 14%
Ryzen 9 3900X
207
Cinebench 11.5 64-bit single-core
Cinebench R11.5 is an old benchmark by Maxon, authors of Cinema 4D. It was superseded by later versions of Cinebench, which use more modern variants of Cinema 4D engine. The Single Core version loads a single thread with ray tracing to render a glossy room full of crystal spheres and light sources.
Benchmark coverage: 14%
Ryzen 9 3900X
2.4
TrueCrypt AES
TrueCrypt is a discontinued piece of software that was widely used for on-the-fly-encryption of disk partitions, now superseded by VeraCrypt. It contains several embedded performance tests, one of them being TrueCrypt AES, which measures data encryption speed using AES algorithm. Result is encryption speed in gigabytes per second.
Benchmark coverage: 13%
Ryzen 9 3900X
11
WinRAR 4.
0
WinRAR 4.0 is an outdated version of a popular file archiver. It contains an internal speed test, using ‘Best’ setting of RAR compression on large chunks of randomly generated data. Its results are measured in kilobytes per second.
Benchmark coverage: 12%
Ryzen 9 3900X
7534
x264 encoding pass 2
x264 Pass 2 is a slower variant of x264 video compression that produces a variable bit rate output file, which results in better quality since the higher bit rate is used when it is needed more. Benchmark result is still measured in frames per second.
Benchmark coverage: 12%
Ryzen 9 3900X
147
x264 encoding pass 1
x264 version 4.0 is a video encoding benchmark uses MPEG 4 x264 compression method to compress a sample HD (720p) video. Pass 1 is a faster variant that produces a constant bit rate output file. Its result is measured in frames per second, which means how many frames of the source video file were encoded per second.
Benchmark coverage: 12%
Ryzen 9 3900X
268
Relative perfomance
Overall Ryzen 9 3900X performance compared to nearest competitors among desktop CPUs.
AMD Ryzen 7 7700
105.23
Intel Core i9-10980XE
101.18
AMD Ryzen 9 3900XT
100.27
AMD Ryzen 9 3900X
100
Intel Core i5-13500
99.28
AMD Ryzen Threadripper 2990WX
98.03
Intel Core i9-12900T
96.47
Ryzen 9 3900X Intel equivalent
We believe that the nearest equivalent to Ryzen 9 3900X by Intel is Core i5-13500, which is slower by 1% and lower by 4 positions in our rating.
Core i5
13500
Compare
Here are some closest Intel rivals to Ryzen 9 3900X:
Intel Core i7-12700K
106.14
Intel Core i7-12700KF
105.46
Intel Core i9-10980XE
101.18
AMD Ryzen 9 3900X
100
Intel Core i5-13500
99.28
Intel Core i9-12900T
96.47
Intel Core i9-12900E
96.29
Similar processors
Here is our recommendation of several processors that are more or less close in performance to the one reviewed.
Ryzen 9
3900XT
Compare
Core i9
10980XE
Compare
Ryzen Threadripper
2990WX
Compare
Core i9
12900T
Compare
Core i9
12900E
Compare
Ryzen 9
PRO 3900
Compare
Best GPU for Ryzen 9 3900X
People consider these graphics cards to be good for Ryzen 9 3900X, according to our PC configuration statistics.
GeForce RTX
3080
6.3%
GeForce RTX
3070
6.2%
GeForce RTX
3090
6.1%
GeForce RTX
2070 Super
5.2%
GeForce RTX
2080 Super
5%
GeForce RTX
3060
4.9%
GeForce RTX
2080 Ti
4.7%
GeForce RTX
2060
4%
Radeon RX
5700 XT
3.
7%
GeForce RTX
3080 Ti
3.7%
These are the fastest graphics cards for Ryzen 9 3900X in our user configuration statistics.
There is a total of 2209 configurations based on Ryzen 9 3900X in our database.
GeForce RTX
4090
0.3% (7/2209)
GeForce RTX
4070 Ti
0.05% (1/2209)
GeForce RTX
3090 Ti
1.7% (38/2209)
Radeon RX
6950 XT
0.05% (1/2209)
GeForce RTX
3080 Ti
3.
7% (81/2209)
GeForce RTX
3090
6.1% (134/2209)
GeForce RTX
3080 12 GB
0.3% (7/2209)
Radeon RX
6900 XT
1.3% (29/2209)
GeForce RTX
3080
6.3% (140/2209)
Radeon RX
6800 XT
0.9% (20/2209)
User rating
Here is the rating given to the reviewed processor by our users. Let others know your opinion by rating it yourself.
Questions and comments
Here you can ask a question about Ryzen 9 3900X, agree or disagree with our judgements, or report an error or mismatch.
Please enable JavaScript to view the comments powered by Disqus.
AMD Ryzen 9 3900X RandomX Benchmarks
?
Below is a table showing all submitted benchmarks for the AMD Ryzen 9 3900X CPU.
- Cores
- 12
- L2 Cache
- 6144 KB
- L3 Cache
- 65536 KB
- Avg. H/s
- 12951
- Avg. Watts
- 167
- Avg. H/Watt
- 77.55
Averages above are per CPU? CPU specifications
| CPU Count ? | H/s ? | Watts ? | RAM ? | OS ? | Submitted ? | Miner ? | Parameters ? | Notes |
|---|---|---|---|---|---|---|---|---|
| 1 | 11835 | Linux |
2019-07-10 Source |
RandomX benchmark | —init 24 —threads 24 —nonces 1000 —largePages | 3200 MHz CL16 RAM | ||
| 1 | 12327 | 32 GB DDR4 |
2019-07-15 Source |
RandomX benchmark | —mine —threads 30 —init 24 —nonces 800000 —largePages —jit | 2*16Gb 3200mhz CL14, 4Ghz on all 12 cores | ||
| 1 | 10965 | 16 GB DDR4 | Windows 10 | 2019-07-27 | RandomX benchmark | 24 threads and init (1:1 infinity fabric to ram results in 200 h/s gain) | ||
| 1 | 11500 | 130 |
2019-07-29 Source |
RandomX benchmark | ||||
| 1 | 12200 | 180 |
2019-07-29 Source |
RandomX benchmark | ||||
| 1 | 13064 | 287 | 16 GB DDR4 | Windows 10 | 2019-07-31 | RandomX benchmark | —mine —jit —largePages —threads 24 —init 24 —nonces 800000 |
slight overclock to 4000 with 1. 25 voltage, Ram timings based upon Ryzen Dram Calculator fast timings settings for Ballistix Elite 3600 CL16
|
| 1 | 13500 | 180 | 16 GB DDR4 | 2019-09-22 | RandomX benchmark | —mine —jit —largePages —init 24 —threads 24 —nonces 800000 | All cores overclocked to 4.2Ghz with 1.28v on Asrock Phantom Gaming 4 x570 motherboard. Memory tuned with Ryzen Dram Calculator with fast settings | |
| 1 | 14238 | 253 | Windows 10 |
2019-11-20 Source |
XMRig | RAM: 3200@C14 | ||
| 1 | 13860 | 16 GB DDR4 | Linux | 2019-12-05 | XMRig | |||
| 1 | 13950 | 132 | 16 GB DDR4 | Linux | 2019-12-07 | XMRig | c14 3200 ram dram calculator | |
| 1 | 13500 | 64 GB DDR4 | Windows 10 | 2019-12-08 | XMRig | |||
| 1 | 16170 | 16 GB DDR4 | Linux | 2019-12-18 | XMRig |
c14 ram ddr4 3200mhz base ram, dram fast timing tweak, OC 3466 mhz @1. 4volt
CPU at 4.375 ghz @ 1.375volt enable huge pages xmrig 5.3.0 |
||
| 1 | 14900 | 220 | 16 GB DDR4 | Windows 10 | 2019-12-22 | XMRig | 2x16GB 3733@CL16 + all subtimings optimized | |
| 1 | 13607 | 16 GB DDR4 | Windows 10 | 2021-06-11 | XMRig |
VID=1295
CCX1=4100,CCX2=3950,CCX3=3950, CCX4=3800 DDR4@3600 2X8GB (ddr4 iridium pro 3200). 
CPU power 142 W @74 deg. C (air cooler Silentium Spartan 4 Max with Delta 1 … |
||
| 1 | 12600 | 75 | 16 GB DDR4 | Linux | 2021-06-28 | XMRig | 1gb Pages, Huge Pages | DDR4 Corsair LPX CMK16GX4M2D3600C18 @933Mhz CL10/11/10 1,13V. CPU @3.8GHZ 1V, FCLK=UCLK=MEMCLK. OS Linux Ubuntu Hirsute Hippo @21.04, XMRig Advance Builder. Mobo MSI X570 … |
| 1 | 9000 | 50 | 2021-10-12 | XMRig | 0 |
AMD Ryzen 9 3900X/3900XT review.
Characteristics and tests. Everything you need to know before buying!
Author Anatoly Vorontsov Reading 5 min Views 333 Published
Updated
Good to see you. Today in the review we will look at a pair of «regional» Ryzen9 processors from AMD: 3900X and 3900XT, and also test their performance in modern games. nine0036
Even the tenth generation of the Intelcore desktop processors could not induce very successful overhead amp to this day they are the best buy for their price. However, in games they are still slightly inferior to competitor solutions, and not only top ones.
The announcement of updated models, which received the XT index, was intended to correct this.
Such unification in terms of naming with video cards looks quite appropriate. Moreover, here the addition of the letter T can also be safely deciphered as “turbo”. The fact is that the only difference in the characteristics of the new models from the old ones is the increased overclocking ceiling in turbo mode by 100–200 MHz. nine0007
While the Ryzen 9 3900X comes with a WraithPrism cooler in the box, the new Ryzen 9 3900XT lacks any stock cooler. With it, AMD offers to use powerful liquid cooling systems purchased separately. Due to the slightly higher frequency achievable in turbo mode, the Ryzen 9 3900XT, ceteris paribus, operates at higher voltages and as a result heats up several degrees more than its predecessor. Therefore, a high-performance cooling system really does not hurt here. nine0007
The Ryzen 9 3900XT is clocked between 3.8 and 4.7 GHz, which formally means that the new processor can overclock 100 MHz more than the Ryzen 9 3900X. At the same time, the Ryzen 9 3900XT retains a TDP of 105W and a 142W power consumption limit.
In general, the Ryzen 9 3900XT did not set overclocking records, but processors capable of operating at a frequency of 4.2 GHz were rare among the Ryzen 9 3900X. For example, according to the statistics of the SiliconLottery store of overclocking processors, the share of such processors in their sample was about 6%. And in the case of Ryzen 93900XT to work at 4.2 GHz was able to get the first copy. So, the Ryzen 9 3900XT confirms the general rule: processors with the XT index overclock better than their predecessors.
Testing
The assembly for the tests will be as follows: an inexpensive ASRock X570 Taichi motherboard, 2 sticks of 8 GB G.SkillTrident Z RGBDDR4-3600 RAM, an NVIDIA GeForce RTX 2080 Ti video card, a Samsung 970 EVOPlus 2TB SSD, and a ThermaltakeToughpower 1000W power supply.
In synthetic tests, programs, processors showed themselves in two ways:
| 3900X | 3900XT | |
| PCMark 10, Essential | 10998 | 11002 |
| 3DMark Time Spy Extreme CPU | 6705 | 6756 |
Corona 1. 3 |
6763 | 6750 |
| Blender 2.82a | 175.7 | 176.5 |
| 7-ZIP 19.00 | 168.6 |
The Ryzen 9 3900XT outperforms the Ryzen 9 3900X more often than it outperforms it. But even where the Ryzen 9 3900XT is ahead, its level of superiority is tenths of a percent. All AMD’s promises about the increase in performance of the Ryzen 3000XT series processors, which allegedly can reach 10%, are nothing more than marketing idle talk. Indeed, there are no such results even close, and it is clearly not necessary to think that with the help of the updated Matisse the company has made some noticeable step forward. nine0007
Let’s see how processors behave in games.
| Ultra, FullHD | 3900X | 3900XT | |
| Assassins Creed: Odyssey | 81.2 | 81.4 | |
| Civilization IV: Gatering Storm | 120. 2 |
119.6 | |
| Far Cry 5 | 129.6 | 129.3 | |
| Gears 50034 | Shadow of the Tomb Raider | 141.5 | 140.2 |
If the Ryzen 9 3900XT couldn’t distinguish itself in resource-intensive applications, then it makes no sense to expect any noticeable success from it in games. In general, new processors with the XT index give out about the same frame rate as their predecessors. It’s better not to go into details, otherwise you can see that the Ryzen 9 3900XT in more cases loses to the old Ryzen 9 3900X. nine0007
Conclusion
The R9 3900X is not very good for esports records. Where low resolutions, outrageous framerates, and minimal graphics settings rule, Intel still does better (thanks to higher clock speeds). And the AMD Ryzen 9 3900XT is a cosmetic upgrade that provides minimal benefits. But, fortunately, even before this very successful processor. And most importantly — at the same price.
What it will be in reality, the near future will show. But the manufacturer, who cares so much about the relevance of the current line, can only be praised. nine0007
But in any case, the R9 3900X/XT is a good system with support for all modern interfaces and standards, outstanding performance in everyday tasks, excellent results in AAA hits, no problems with streaming and upgrades. 4K, high-definition content, video editing and 3D, complex computing.
new Zen2 microarchitecture and already 12 cores on the familiar AM4 platform
Methodology for testing computer systems of the 2017 model
Some three years ago, many of our readers stopped taking AMD seriously as a developer of high-performance processors. In this market, you need to run fast even in order to just stay in place, and AMD “ran” for several years exclusively in laboratories — secretly from prying eyes. The stores continued to receive processors from the FX family, which raised questions even at the time of their appearance in 2011, and the AM3 + platform has not evolved much since the last decade.
In general, the only argument in favor of these decisions from the surviving few fans was «but cheap!». nine0007
AMD Ryzen 7 1800X Processor:
Things changed radically in 2017 with the debut of the Zen microarchitecture. It cannot be said that technically it turned out to be ideal — in fact, the consumer characteristics of the first models were very reminiscent of Haswell of the 2013 model, and by that time Intel had mastered Skylake for a year and a half. Tellingly, the AM4 platform had some of its features more in common with the Intel LGA1150, and not with the latest (at that time) LGA1151. Thus, while maintaining the same approach to the layout of processors, it would be difficult to talk about parity between products. Therefore, AMD did not keep it — the company «simply» increased the number of cores in mass processors, which Intel has not done since the end of the «zero». As a result, the older Ryzen 7 began to compete in performance with the Core i7, but not for LGA115x, but for LGA2011-3.
This is also Haswell — only more expensive. And here — similarly, but much cheaper. Against this background, LGA1151 already looked pale: four cores from Intel turned out to be either much more expensive than four from AMD, or slower than six or eight from AMD. nine0007
Testing eight-core processors Intel Core i7-9700K and i9-9900K (Coffee Lake Refresh)
However, at that time both Skylake and the 14 nm process technology still had upgrade reserves. Intel took advantage of them in the fall of 2017, releasing a new LGA1151 modification and six-core Coffee Lake processors for it. On the one hand, six is not eight. On the other hand, these six turned out to be no worse than those eight in multi-threaded code, since they were more efficient when comparing “core to core”. A certain price correction was also carried out: the prices of each Core line were kept the same, but the cores (and performance) were thrown for this money. As a result, a precarious balance has developed in the mass segment, which neither Ryzen «2000 series» nor Intel’s Coffee Lake Refresh changed last year.
The latter could, fortunately, eight cores appeared there. But the prices of older solutions have increased — with all the consequences. So AMD did not need to come up with something new. Moreover, the company already had an inexpensive HEDT platform in stock — not without flaws, but inexpensive. Accordingly, for consumers who “missed” the older Ryzen 7, AMD could offer Ryzen Threadripper — also not without rough edges (since it is practically a two-processor system in one package), but for the sake of 12 or 16 processor cores, many were ready to turn a blind eye to them. nine0007
However, as mentioned above, in this market needs to run fast even in order to simply stay in place of . AMD did not forget about this — and prepared Zen2. Today, the novelty has finally and officially debuted in final products, which you must definitely get acquainted with, which we will do.
Brief summary of changes
It should be noted right away that although Zen2 is based on Zen developments, there are a lot of changes in both the microarchitecture of cores and the architecture of processors.
In fact, you can even talk about something new — and to some extent revolutionary. Therefore, this topic deserves a separate detailed material, and today we will just go over some points — which will be needed in the future, so as not to be surprised by the results. nine0007
So, first and foremost, in Zen2, the company is seriously concerned about the efficiency of the execution of each stream of commands, and not their number. Many have already seen slides with an increase in IPC (the number of instructions executed per clock) by 15%, but we must remember that this is only an average value, depending on the specific code. In particular, the rate of execution of AVX / AVX2 commands has doubled — and the new Ryzen at the same frequency should cope with it no worse than Core. In addition, the frequencies have also increased: if the first generation rarely broke through the 4 GHz border even in the hands of overclockers, and the second (Zen +) did not go far from it, now the once maximum frequencies have simply become starting points for many processors.
The maximum, in favorable conditions, are in the region of 4.5 GHz, and without manual overclocking. Thus, to 15% of the increase in IPC, it is worth adding another 15% -20%: this is how much “C” has increased (strictly speaking, decreased) in this abbreviation. And all this at the same time is based on a significantly redesigned memory controller, increased speed of working with caches, and a simple increase in the capacity of the L3 cache too: it has been tritely doubled. In general, in short: what was worse than in Core (AVX or the performance of the cache memory of micro-ops), they did the same, and what was better, they strengthened it a little more 🙂
All of these changes required a finer process technology, which the company received — now it is 7 nm. But it is quite difficult to produce large crystals on it, so the company decided to go for another risky experiment. All the same blocks of two CCX (as in Zen / Zen +), i.e. 8 processor cores with three levels of cache memory and Infinity Fabric controllers, are produced in the form of chiplets using the 7 nm process technology.
But for the first time since the appearance of the first Athlon 64, the memory controller is absent from the processor chip: together with the PCIe 4.0 controller and peripheral controllers (such as USB, etc.), it is placed on a separate chip manufactured according to 12 nm standards. What is the risk of such a maneuver? It was not for nothing that the memory controller was «dragged» closer to the processor: this allows you to reduce delays. However, this is in theory — but in practice it turned out differently in different processor models. Apparently, AMD decided that the potential decrease in performance is small — and is fully offset by the price reduction in the production of such an assembly using two factories. In addition, in the new architecture of processor chiplets on the same substrate, there may be not one, but two — which is implemented in the Ryzen 9 family. Moreover, unlike Threadripper, all 12 or 16 cores work with one memory controller, so access to the latter is symmetrical.
Of course, the memory controller remains dual-channel — this is necessary for compatibility within the same platform. For the same reason, although the PCIe controller has grown in speed (PCIe 4.0 is twice as fast as 3.0), it still has the same 24 lines — four of which are needed to communicate with the chipset, and four more usually serve the «main» SSD. Thus, AM4 cannot serve as a complete replacement for HEDT platforms, but it can satisfy the needs of those users who need more than eight processor cores, but not rich peripheral capabilities. And inexpensive. And almost in a Mini-ITX case, since the same Ryzen 9compatible with a large number of motherboards already on sale, up to the most compact ones. For Ryzen 5 and Ryzen 7, this performance is simply the result of savings. Whether it interferes somewhere in practice — tests will show. On the other hand, given the potential increase in productivity by a third on average (somewhere — and more), this is unlikely to become noticeable.
Testing the Intel Core i7-8705G and i7-8809G (Kaby Lake-G) processors and studying related issues
Along the way, we also note that this performance opens the way for the creation of multi-core APUs. Previously, AMD had to make a special die for their production, where one CCX was replaced by a GPU, which limited the number of cores to four. The current version allows you to put side by side a «full» eight-core chiplet, and a powerful (relatively) GPU. There (albeit without maintaining compatibility) it is theoretically possible to «pack» several gigabytes of HBM2 memory, i.e. get an expensive, but very powerful solution for high integration — for example, for top laptops. And here you can also remember that Intel does not plan to update Kaby Lake-G in the near future, and AMD can develop a faster assembly (especially since Intel still uses AMD GPUs in the mentioned processors). nine0007
But these are all questions for the future. Currently, six models of «clean» processors for desktops with from 6 to 12 cores have been announced and are starting to ship.
Each core is at least a quarter faster than before, which means that the new processors will be just as faster than the old ones similar in positioning. At the same time, they are compatible with the old ones: this is the same AM4 platform. Nuances with ensuring compatibility, however, are possible — as well as the fact that on some old motherboards new processors «will not give out» everything they are capable of. However, only time and practice of use will give exact answers to all questions. For now, let’s just see what we can expect. nine0007
Test stand configuration
| Processor | AMD Ryzen 7 3700X | AMD Ryzen 9 3900X |
|---|---|---|
| Core name | Matisse | Matisse |
| Production technology | 7/12 nm | 7/12 nm |
| Core frequency, GHz | 3.6/4.4 | 3.8/4.6 |
| Number of cores/threads | 8/16 | 12/24 |
| L1 cache (total), I/D, KB | 256/256 | 384/384 |
| L2 cache, KB | 8×512 | 12×512 |
| L3 cache, MiB | 32 | 64 |
| RAM | 2×DDR4-3200 | 2×DDR4-3200 |
| TDP, W | 65 | 105 |
Number of PCIe 4. 0 lanes |
20 | 20 |
The main heroes of today’s testing are two AMD processors, which are formally the junior models in the Ryzen 7 and Ryzen 9 lines. not yet officially announced. As for the 3700X, in the new line it just corresponds in price to the former «flagship» with the 2700X index, as well as Intel’s Core i7, so it is the one that is most interesting. The Ryzen 7 3800X is, after all, just a slightly more expensive «steroid» modification, so we think it will be in much less demand. nine0007
| Processor | AMD Ryzen 7 1800X | AMD Ryzen 7 2700X |
|---|---|---|
| Core name | Summit Ridge | Pinnacle Ridge |
| Production technology | 14 nm | 12 nm |
| Core frequency, GHz | 3.6/4.0 | 3.7/4.3 |
| Number of cores/threads | 8/16 | |
| L1 cache (total), I/D, KB | 512/256 | 512/256 |
| L2 cache, KB | 8×512 | 8×512 |
| L3 cache, MiB | 16 | 16 |
| RAM | 2×DDR4-2666 | 2×DDR4-2993 |
| TDP, W | 95 | 105 |
Number of PCIe 3. 0 lanes |
20 | 20 |
Since we mentioned the first generation of Ryzen above, it would be logical to take for comparison not only the Ryzen 7 2700X, but also the 1800X — from which AMD’s triumphant return to the market began the year before last.
| Processor | AMD Ryzen Threadripper 2920X | AMD Ryzen Threadripper 1950X | AMD Ryzen Threadripper 2950X |
|---|---|---|---|
| Kernel name | Colfax | Summit Ridge | Colfax |
| Production technology | 12 nm | 14 nm | 12 nm |
| Core frequency, GHz | 3.5/4.3 | 3.4/4.0 | 3.5/4.4 |
| Number of cores/threads | 12/24 | 16/32 | 16/32 |
| L1 cache (total), I/D, KB | 768/384 | 1024/512 | 1024/512 |
| L2 cache, KB | 12×512 | 16×512 | 16×512 |
| L3 cache, MiB | 32 | 32 | 32 |
| RAM | 4×DDR4-2933 | 4×DDR4-2666 | 4×DDR4-2933 |
| TDP, W | 180 | 180 | 180 |
Number of PCIe 3. 0 lanes |
60 | 60 | 60 |
Well, since we have a «multi-core» Ryzen 9 (at least one), let’s add a trio of Ryzen Threadripper: 2920X with the same 12 cores and a couple of 16-cores, the first (older) of which has fallen in price by this year almost to the «popular» level, and the second is simply the fastest of the unconditionally applicable in the PC (we are still not inclined to perceive the WX-series as such — and not only because of the price). nine0007
| Processor | Intel Core i7-8700K | Intel Core i7-9700K | Intel Core i9-9900K |
|---|---|---|---|
| Core name | Coffee Lake | Coffee Lake Refresh | Coffee Lake Refresh |
| Production technology | 14 nm | 14 nm | 14 nm |
| Core frequency, GHz | 3.7/4.7 | 3.6/5.0 | |
| Number of cores/threads | 6/12 | 8/8 | 8/16 |
| L1 cache (total), I/D, KB | 192/192 | 256/256 | 256/256 |
| L2 cache, KB | 6×256 | 8×256 | 8×256 |
| L3 cache, MiB | 12 | 12 | 16 |
| RAM | 2×DDR4-2666 | 2×DDR4-2666 | 2×DDR4-2666 |
| TDP, W | 95 | 95 | 95 |
Number of PCIe 3. 0 lanes |
16 | 16 | 16 |
But, perhaps, we will not consider Intel HEDT processors today — they have become too “thing in themselves”. It is clear that the LGA2066 platform has its advantages, but from the point of view of lovers of the notorious price / performance ratio, the LGA2066 processors themselves are not even competitors to Threadripper. But today we definitely need a trio of models for LGA1151: the coincidence of the indexes of AMD and Intel processors is far from accidental (although officially AMD may claim that this is just a coincidence — but you can’t deceive the recommended retail prices :)). nine0007
Testing methodology
In general, we have a new testing methodology ready, but too few processors have been tested using it, so there is nothing to compare the new products with. But given that AMD is particularly focused on the increased performance of each core, it will be useful to evaluate what this gives just in not the newest applications — where there are clearly no special optimizations.
Therefore, today we will use the iXBT.com performance measurement method based on real applications of the 2017 sample and the related power consumption measurement method when testing processors, and we will conduct one of the first articles based on the new method with the same (or close participants) — at the same time and see , what will change. nine0007
Detailed results for all tests are available as a complete results spreadsheet (in Microsoft Excel 97-2003 format). Directly in the articles we use the processed data. This is especially true for application tests, where everything is normalized relative to the reference system (AMD FX-8350 with 16 GB of memory, GeForce GTX 1070 graphics card and Corsair Force LE 960 GB SSD) and grouped by computer application areas.
7 Games on 2 GPUs, Part 3: AMD and Intel 6-Core vs. Other Solutions
And also separately later we will conduct a study of performance in games. Current developments on this topic (which can be found in the relevant materials) show that games, as a rule, do not fully utilize more than six computational threads, firstly, and secondly, they are still strongly tied just to “single-threaded performance”.
Thus, in principle, the assessment of the latter will immediately give an answer to the question of how the new Ryzen are suitable for games. Moreover, it is interesting (taking into account the «first» above) for Ryzen 5, but not for Ryzen 7 and Ryzen 9- where the number of cores is clearly redundant .
iXBT Application Benchmark 2017
We were initially mentally prepared for the Ryzen 7 3700X to overtake the Core i9-9900K, but Ryzen 9 turned out to be a real discovery. No, that it will bypass a pair of 1920X / 2920X, it could also be assumed with great confidence, but such a “hello” to the buyers of all Ryzen Threadripper, which turned them into a pumpkin … But there will be 3950X in the fall! However, it is quite clear that the scalability of the TR4 platform is still a little, but better. So what? The kernels are much worse now. And after the update, their number may already seem redundant to many buyers. Well, Intel can’t oppose anything to this at all.
nine0007
These applications are better optimized for modern instruction sets, so here the 3700X is still slower than the i9-9900K. Well, it’s okay: formally, this processor is a competitor for the i7-8700K / i7-9700K, and the old Ryzen 7 more or less coped with this. The new ones are a quarter faster. Which, however, no longer causes reverence, since Ryzen 9 has become the top family. And its first representative continues to amaze: in fact, the increase in the efficiency of each core is enough for it to stay at the level of 16-core Threadripper (albeit without overtaking them in this case). nine0007
Working with video is a traditional area of application for multi-core processors. However, the emergence of mass models of such quickly showed that from a certain point on, the “race” for the number of cores and computation threads begins to lose its meaning — “quality” is more important. This was especially well demonstrated by a pair of Ryzen Threadripper 1950X / 2920X: the second has fewer cores, but it is faster because the cores are better.
And in the «3000 family» the quality is still much better — with a natural result: already the 3700X easily invades the holy of holies of the HEDT segment. And only the existence of faster models with the same architecture prevents him from being considered the best friend of a video editor. However, they are also more expensive — but those who wish to pay extra, as it seems to us, will be found anyway. But pay for Core i7 or Core i9, it seems that there is simply no point in such conditions.
These programs, moreover, always gave preference to the quality of the cores, rather than their quantity, which Ryzen did not leave much chance before: most of them did not catch up with the Core i7-8700K, not to mention the Core i9-9900K. And now everything has magically changed: it seems that even many Ryzen 5 will already be faster than all Intel processors. But we will check this later — when they get to us. In the meantime, we state the fact that the Ryzen 7 3700X can be considered the best solution considering the price.
Ryzen 9there is no need to buy for such an application, but if you suddenly do, they are simply the fastest here.
A simple integer code where the quality of the cores always mattered less than the quantity. But with an equal number of cores, the quality still decides. And we have noted more than once that the task, in fact, is very “cash-loving”, so we remember the huge L3 of the new Ryzen and stop being surprised at the results.
As mentioned above, memory handling in Zen 2 is drastically accelerated. However, this does not always “save”, “massive” Core is still more efficient — in the sense that they allow you to “squeeze” more out of slower memory. In addition, archivers are very susceptible not to memory bandwidth, but to delays — and here in the new generation of Ryzen one can even talk about deterioration. As a result, purely technically, the Ryzen 7 3700X is still inferior to the Core i9-9900K with the same number of cores, with a higher memory frequency (we used, recall, DDR4-3200 and DDR4-2666, respectively), with a larger cache capacity .
.. But this is “purely technically”, since these processors are very expensive differently. The increase relative to 2700X (that is, AMD’s earlier best processor of the same class) is the usual ≈20%. And this is very serious. As well as the superiority of this model over the Ryzen Threadripper. And now there is Ryzen 9 for records. The individual preferences of the programs in the group, however, are different: SolidWorks, for example, began to work only 5% faster (however, it did slow down on Threadripper, so this is just an example of a program that is generally very difficult to speed up). But the fast Fourier transform 3700X performs about 60 percent faster than 2700X — earlier Ryzen had difficulties with this program due to the low rate of work with “doubles” (and the fastest, by the way, for once continue to be Skylake -X — even the smallest models of this architecture). The cumulative effect is almost killer : if earlier it was possible to talk about some kind of parity between LGA1151 and AM4, now this is not even close.
In scientific computing, the second platform now lives at the same level as the HEDT solutions from AMD and Intel. Moreover, even the number of cores does not always allow the latter to steadily dominate, since there are already 12 cores on AM4, and in the fall there will be 16. And far from all programs can still normally digest a larger number (and this is still very mildly speaking).
The overall result is natural. And here it is worth paying attention to the next point. Since the first FX, not a single update to the AMD range has been complete without reasoning that processors are good, but the software does not know how to use them correctly . And the operating system somehow “wrongly” distributes processes among the cores, and application programs “wrongly” use them, and even energy-saving schemes “do not coincide” with hardware capabilities, etc., etc. So, they say, you must definitely wait: miracle patches will come out, and everything will magically get better.
And patches did come out — but suddenly it turned out that in reality nothing had changed significantly. And this was true even for ordinary Ryzen — and there was always even more noise around Threadripper with their “very large” number of cores and cunning memory schemes. nine0007
In the case of the Ryzen 3000, the magic patches… are simply not needed. AMD has significantly redesigned both the microarchitecture and its practical implementation — and this immediately had an effect in terms of performance. In the same operating systems and the same (not even new) application programs, we calmly and without additional gestures get up to 30% acceleration: half of them are microarchitectural changes, and the same amount — all the rest. And all of this scales well to more cores — only slightly worse than when «globally doubling» all the blocks in the Ryzen Threadripper. Moreover, it is much cheaper, and the result is compatible with the same motherboards that run «regular» Ryzen. nine0007
Actually, this is how an intensive rather than an extensive approach to development should look like.
Intel showed us something similar in 2011, when Sandy Bridge processors entered the market. True, the overall effect then looked a little more modest, since the company did not increase the number of cores in the mass platform, and even neglected other extensive methods — the same cache memory capacity, for example, which doubled in the new Ryzen. Therefore, then the increase was more modest — but also impressive. However, all this has become overgrown with past, and now, as we see, there is simply nothing to answer the new AMD products from Intel. Even if we ignore the prices, then in terms of the pure performance of the Core i9The -9900K is no faster than the Ryzen 7 3700X, and then there’s the faster Ryzen 7 3800X… And then there’s the Ryzen 9 — which is a whole other story than .
Energy Consumption and Energy Efficiency
Mastering the new process technology allowed AMD to «fit» into the same «energy limits» as before. In fact, the Ryzen 7 3700X in this regard is an analogue of the Ryzen 7 2700, and not a somewhat “overclocked” 2700X — and not from scratch has the same TDP of 65 W (it often climbs beyond which, if possible, but this has long been common practice).
But Ryzen 9 is already comparable to the 2700X3900X, and it «eats» a little more than older models of processors for LGA1151. But its consumption is radically lower than that of processors for TR4 — with which it is just more correct to compare it in terms of performance.
As a result — new records for «energy efficiency». We are used to those made by Intel, but this company has not pleased with them for a long time: the “ninth” generation core is worse than the “eighth” in this regard, and HEDT solutions after the introduction of Skylake there have become quite voracious. AMD, on the other hand, first entered the Haswell level, firmly established itself on it («successful» instances of processors we also came across in the «2000 series») — and now they have made a new big jump . Even the Ryzen 9 3900X is comparable to the best Intel products — despite the fact that it is clearly not the best chips that need to be stocked for the release of the 3950X. Well, the 3950X itself will certainly be even better in this regard.
Total
In principle, everything is clear with the technical details, and for once the results speak for themselves. So let’s talk about other issues. For example, that improving “single-threaded” performance is a difficult task. But if it works out, the result will be cumulative. And Zen2 once again brilliantly confirms this: new cores, even while maintaining their number, are faster always and everywhere. No need to think about how well the software used can utilize their amount: even a purely single-threaded program on the Ryzen 3000 will work faster than on representatives of previous families or Core. So the main problem of AM4 is solved, and this makes this AMD platform the best choice for a mainstream computer. And taking into account the fact that the number of cores has once again grown — the choice is practically uncontested. Yes, of course, HEDT solutions still have advantages in terms of the number of RAM modules and / or full-speed PCIe slots, but this will be significant in scenarios that go far beyond the needs of an individual user.
But 12/16 cores Ryzen 9many of them may already come in handy — especially since the company does not require special money for this. For games, the new Ryzen 5s may turn out to be excellent offers: too many cores in this segment are still not required, so the six «new» ones can easily cope with loading almost any video card. On the other hand, the Ryzen 7 3700X is not so expensive against the backdrop of the cost of top-end video cards, so it is perfect for the role of a universal solution — both to play and work. In any case, you no longer need to be torn between smart and beautiful, choosing either «fast cores» or «many cores inexpensively.» There was an offer «a lot of fast cores inexpensively», which is extremely difficult to resist. nine0007
And AMD has chosen an excellent moment for this: the rival is not in a position to unleash price wars. The 14 nm process technology used by Intel for almost five years now, coupled with the doubling of the number of processor cores since then (and, in fact, the same ones — inside it is a Skylake sample of 2015) and a corresponding increase in the die area, has led to a shortage of processors, despite their not the lowest prices.
It looks like AMD will help to cope with this deficit 🙂 Let it be only in some segments so far, but even that is bread. And looking at the energy efficiency of new solutions, we can assume that their adaptation for laptops will also be very successful. In any case, for gaming laptops, where you still have to use a discrete graphics card. Well, there will be updated APUs, and, as mentioned above, a new approach to building a finished product can lead to revolutionary changes here too. nine0007
In general, the bets are made. More precisely, rate — from AMD. We will see a full-fledged response from Intel only next year — when the company plans to radically update the microarchitecture and master a new technical process. What happens in the end — we’ll see. But not this year. And in this we can see, as a last resort, a price reduction, and even then it is hardly radical. Thus, at least in the desktop processor segment, at least the second half of 2019 (including the «fat» sales by September 1 and December 25 — the time for the release of new processors was chosen correctly in this regard) will pass under the AMD sign.