AMD Ryzen 7 3700X and Ryzen 9 3900X
AMD’S RYZEN 3: THE REAL DEAL
AMD roared back to competitiveness when the first generation Ryzen processors launched in early 2017. The first Zen processors received a lot of praise for their strong multi-threaded performance, though they still lacked that bit of single and lightly threaded performance compared to Intel that kept the blue company on top. Second generation Ryzen saw some evolutionary improvements with a 12nm process and some latency tweaks. Fast forward to now, and we have the first real Ryzen overhaul.
We’re not exaggerating when we say that Zen 2 is as big a deal as 1st generation Ryzen was back in early 2017, if not more so. It’s not just that there’s an architectural overhaul, there’s also the not insignificant engineering challenges of moving to a new 7nm process node as well as a chiplet based design. Any one of these things alone would be news worthy, but having all 3 together at the same time, without even changing the socket, is something quite extraordinary. Yeah, throw in PCI Express 4.0 for good measure.
Here’s a look at the Ryzen 3 line-up, including Australian pricing for the first time. The range appears to be particularly well balanced. Perhaps the pick of the bunch based on spec alone is the Ryzen 7 3700X. It’s an 8c/16t part that’s notable for having a healthy 4.4GHz boost clock and a very impressive 65W TDP, at $519. This one could end up being an enthusiasts’ favourite.
We’d like to see something like a 45W quad core at $250 or less but that’s likely to come later. Let’s not forget the 3200G and 3400G APUs, though they are Zen + 12nm based. We’ll likely see the range fleshed out down the line with Ryzen 3 models towards the end of the year.
Moving up the range, the Ryzen 9 3900X and 3950X manage to pack 12 and 16 cores into a 105W TDP. This means no need for mega cooling. Note that the 3950X is expected to come on sale in September with pricing to be announced later. Compare this to the likes of the HEDT Threadripper 2950X at 180W, along with Intel’s 165W Kaby Lake-X models and it’s clear that AMD’s power efficiency has taken a massive step forward.
All Ryzen 2 CPUs come with a bundled cooler. The Wraith Prism RGB that came with our 3700X and 3900X samples is a capable cooler, though it won’t be able to compete with the larger surface area air coolers or decent AiO’s. It certainly looks good with its dash of RGB lighting.
INTRODUCING ZEN 2
Ryzen 3 desktop processors are also known by their codename, Matisse. We’ll start with the move to a chiplet based design. Ryzen 3 CPUs contain three dies, consisting of two core complexes built with the 7nm process, and an I/O die built on 12nm. This means better yields, higher wafer density and lower costs while increasing design flexibility and allowing for easier scalability. Consider though, that all AM4 processors must adhere to the same pinout featured on the first 28nm monolithic die quad core processors. Add it all up: Moving from 28nm to 7nm, monolithic to multi die, four cores to sixteen plus PCIe 4.0. All with the same pin layout with backward compatibility. That’s really amazing.
The move to 7nm allowed AMD to really shrink transistor size and hence the core complexes (CCX). A single CCX is just 74mm2, with a big chunk of this being taken up by the L3 cache. If the I/O die was to move to 7nm in the future, we could see AMD choosing to further increase the L3 cache size.
Take a look at the routing diagram below. AMD really deserve credit for sticking to the AM4 socket with its pin restrictions while rearranging almost the entire package. Don’t forget that higher memory clocking and PCIe 4.0 have tighter tolerances. It’s a real feat of engineering.
Perhaps the biggest architectural gain comes from a redesigned cache hierarchy. A larger micro-op cache, 512k L2 cache and a doubling of the L3 cache to 32Mb per die all add up to a big chunk of the latency improvements between Ryzen 2nd and 3rd generations, which particularly helps to increase game performance. AMD goes as far as to call this cache ‘Gamecache’, an interesting nomenclature! The doubled L3 cache size results in less requests to main system memory which helps with latency sensitive applications such as high FPS gaming.
Additionally there’s an all new TAGE branch predictor, doubled floating point capability and a bunch of other things that take a PHD in engineering to understand, leading AMD to claim that IPC has been uplifted by up to an impressive 15%, and that’s before taking into account any clock speed improvements.
All of this means that math performance of Ryzen 3 has taken a big step forward and this is manifested in the very strong multi-threaded and content creation performance of the new CPU’s, as we’ll see in the results section below.
One of the interesting features of Ryzen 3 is the updated Infinity Fabric interconnect. AMD claims that overall efficiency has improved by 27% leading to lower power consumption.
Zen and Zen+ processors faced issues with memory speeds because the memory clock and Infinity Fabric clock (fclk) were locked at 1:1. With Zen 2, the fclk and DRAM clocks are decoupled. Ostensibly this is a good thing, but it will require a detailed analysis as to how it affects bandwidth once 2:1 is enabled. It should be possible to run well over DDR4-4000 but the trade-off loss of bandwidth from a 2:1 reduced fclk could mean that DDR4-3600 (or a bit more if your CPU can handle it) tied 1:1 with the fclk will be the best performing setting. This is something we’d like to test, but with limited hands-on time and also a new GPU series launching at the same time, we’ll have to revisit this in the near future. Luckily, a setting of DDR4-3600 C16 required zero effort beyond enabling XMP. This kind of RAM speed with optimized timings is a good, high performing sweet spot for most users.
Security is a big deal and AMD has taken further steps to harden its products from exploits. Fortunately for AMD, many of the publicized side channel vulnerabilities don’t affect AMD processors, but they’re not immune from all of them. The Spectre exploit has been fixed in hardware. Spectre V4 has also been patched, which in conjunction with OS level protections, should keep AMD processors secure from current known attacks. Other issues such as the more recent zombieload do not affect AMD processors.
Another welcome issue that’s been addressed is Windows scheduling. This has been a problem for high core count AMD CPU’s and is something we noted with Threadripper. The May 10th 1903 update to Windows introduces some features in order to get the most out of Ryzen 3. The first improvement is thread allocation. Windows now knows that it should allocate threads to a single core complex and fill it before accessing another. AMD says this will manifest itself with improved game performance, for example threads will not have to go via the I/O die to talk to other threads.
The other significant improvement comes from faster clock ramping. AMD claims just 1-2ms to shift between different frequencies which will help for burst type workloads and transitions between various load and idle states. Pre build 1903 Windows supposedly takes up to 30ms to shift between clocks.
There’s been a lot of talk about PCIe 4.0. While Ryzen 3 is backwards compatible with most 300 and 400 series motherboards (depending on the manufacturer), only X570 motherboards will support PCIe 4.0 at this time. PCIe 4.0 won’t actually benefit a GPU for now, unless under very specific workloads. It does however, allow for faster NVMe SSD’s and diverse platform I/O configurations. The latter is key as it allows board manufacturers to allocate lanes towards a bigger variety of controllers. 10G LAN? Wi-Fi 6? USB 3.2? Even single PCIe 4.0 lanes are sufficient for a lot of auxiliary functionality.
Page 1: Introduction, Architecture Details
Page 2: AM4, Test Setup, Benchmarks & Performance
Page 3: Temperatures, Power Consumption, Conclusions
All original content copyright James Rolfe.
All rights reserved. No reproduction allowed without written permission.
Interested in advertising on OCAU? Contact us for info.