We have previously explored the importance of memory scaling in Ryzen CPU AMDs: today's answered question is how much effect the performance of memory frequencies has when Zen is paired with AMD's own Vega graphics core. We run a full suite of tests on AMD Ryzen 3 2200G ($ 99) and Ryzen 5 2400G ($ 169) APU with memory speed from DDR4-2133 to DDR4-3466 using the G.Skill Ripjaws V kit.
Memory Scaling on AMD Ryzen APUs
Although adding Vega to Zen may be a new concept, the premise of APU that combines computing and graphics on the same chip remains the same. Graphs are often memory-bound operations - the speed at which data can be accessed by graphics is directly related to the frame rate, and we have seen on chips in the past that memory speed (or temporary cache) can greatly help speed up graphics performance. Graphics usually focus here, because faster memory only helps the limited workload of the CPU's memory.
One of the main problems with today's memory is price. With DDR4 prices increasing during 2017 and without signs of a slowdown in 2018, building new desktop systems has been looking more expensive over the past few years: GPU price inflation is also certainly contributing to their misery. While the general view on the current DDR4 DRAM market is that for users who want extra speed, more money should be spent is true, how it's equivalent to the actual performance becomes more relevant than ever before. At a price, for example, this is a Corsair Vengeance LPX 2x8 DDR4-2666 memory device in Amazon:
The price of this memory when it was launched was $ 142, which dropped to as low as $ 57 on sales but averaged $ 75 during early 2016. During 2017 and 2018, this very popular memory kit is now trading at $ 179, having reached $ 200. To put it in perspective, the kit was launched at a cost of $ 8.88 per GB, down as low as $ 3.56 per GB, and now at $ 11.19 per GB. It's almost certainly a seller's market, not a buyer's market. People often spend money on capacity exceeding speed. The purpose of this article is to determine how much the actual speed, especially when we see low cost processors like AMD Ryzen APU.
For users who want to build a budget system without overly focusing on high-performance applications such as CAD or content creation, Ryzen 3 2200 and Ryzen 5 2400 APU have a lot to offer, especially when money is a very limiting factor in purchasing decisions. As in our Ryzen 5 2400G review, we conclude that the Ryzen 2000 AMD series pair offers the best value and performance compared to what is currently offered on both sides of the APU / CPU market (Intel or AMD) when an iGPU is displayed on the chip.
Memory Scaling on APUs: More Cloth Infinity
Most of the following analysis in this section is taken from previous Memory Scale article on our Ryzen 7.
While we already know because of previous testing we did with Ryzen with what effect memory frequency on the Zen core, and AMD added this new element when equipped with Ryzen 3 2200G and Ryzen 5 2400G with Vega. In keeping with the rest of the Ryzen processor range from AMD, each chip incorporates some of the most relevant, but relatively speaking, technology that has the ability to affect memory performance in the 2000 Ryzen series called Infinity Fabric.
The Infinity Fabric (hereinafter abbreviated to IF) consists of two fabric fields: Scalable Control Fabric (SCF) and Scalable Data Fabric (SDF). SCF is all about control: power management, remote management and security and IO. Basically when the data must flow to different elements of the processor other than the main memory, SCF is in control. SDF is where the main memory access comes into play. There's still management here - being able to set buffers and queues in order of priority helps with latency, and the organization also relies on quick implementation. The slides below are more intended for IF implementations in AMD server products, such as power controls on individual memory channels, but are still relevant to speed up the consumer workflow.
AMD's goal with IF is to develop interconnects that can go beyond the CPU, CPU, and GPU scales. In the EPYC server product line, IF connects not only the core in the same silicon part, but silicon in the same processor as well as the processor to the processor. Two important factors come into design here: power (usually measured in energy per bit being transferred) and bandwidth.
Bandwidth IF is designed to adjust the bandwidth of each major memory channel, creating potentially unified solutions without the use of buffers or large delays. Discussing IF in the server context is somewhat beyond the scope of what we tested in this article, but the point we are trying to get across is that the IF is built with a wide product coverage in mind. On the consumer platform, while the IF is not necessarily used for such a large degree on the server, the potential speed of IF to affect performance is just as high.
As per our testing policy, we take the appropriate premium category motherboard for the socket, and complete the system with the appropriate amount of memory. By setting this test, we use BIOS to set the CPU core frequency using the straps available on the MSI B350I Pro AC motherboard. The memory is set to the speed range provided for our test.
In order to obtain a consistent set of results, the G.Skill Ripjaws V DDR4-3600 kit is set to the 17-18-18-38 latency as long as each of the different strings is tested. Due to the inability to support 100 MHz straps on our motherboards, the XMP profile is enabled in the BIOS on the Ripjaws V kit memory and adjust the latency time adjusted to 17-18-18-38 manually on the MSI B350I Pro AC motherboard. Each rope is set for purpose in continuity testing for the frequency scale.
Additional notes about our previous experience with memory scale. In the past we introduced the concept of Performance Index (PI) for each memory device, to provide rough performance comparison metrics between memory brands. This PI is defined as data rate (such as DDR4-2400) divided by CAS Latency (like 17 in 17-18-18) rounded down to the nearest integer. In previous articles like this, typically the highest PI memory gets the best overall score, especially in games, although in combination with similar PIs, those with the highest frequencies are often ahead. We will review this concept later in the review.
In this review, we will test the following combinations of data rates and latencies:
At the Ryzen launch, a number of industry sources personally revealed to us that the platform side of the product line is in a hurry. There is little time to do a complete list of DRAM compatibility, even with standard memory kits on the market, and this causes some problems for early users to try and get a suitable kit that works well without some tweaking. In a few weeks it's ironed when memory vendors and motherboard vendors have the time to test and customize their firmware.
Replacing this is a lower than expected DRAM frequency support level. During the launch, AMD has promised that Ryzen will be compatible with high-speed memory, but reviewers and customers are having problems with high speed memory kits (DDR4-3200 and above). These problems have been solved through a motherboard BIOS update wave built on the latest version of AGESA (AMD Generic Encapsulated Software Architecture), specifically up to version 1.0.0.6, but now preceded by 1.1.0.1 (thanks to the usual AMD system numbering).
While the maturity of the Ryzen platform is no longer a problem encountered, the AGESA 1.1.0.1 microcode is specifically focused on supporting Raven Ridge Ryzen 3 2200G and Ryzen 5 2400G APU announced before launch; we reviewed this BIOS update for AMD's Ryzen APU in February at launch.
The whole purpose of the current test is to evaluate the scalability of the Zen AMD architecture and to see if its performance can be affected by increasing DRAM frequencies. It would be foolhardy not to define the effect on game performance and to see if the memory frequency has a direct impact on the frame rate given that previous AMD APU generations have been reported to do so.
The Infinity Fabric (hereinafter abbreviated to IF) consists of two fabric fields: Scalable Control Fabric (SCF) and Scalable Data Fabric (SDF). SCF is all about control: power management, remote management and security and IO. Basically when the data must flow to different elements of the processor other than the main memory, SCF is in control. SDF is where the main memory access comes into play. There's still management here - being able to set buffers and queues in order of priority helps with latency, and the organization also relies on quick implementation. The slides below are more intended for IF implementations in AMD server products, such as power controls on individual memory channels, but are still relevant to speed up the consumer workflow.
AMD's goal with IF is to develop interconnects that can go beyond the CPU, CPU, and GPU scales. In the EPYC server product line, IF connects not only the core in the same silicon part, but silicon in the same processor as well as the processor to the processor. Two important factors come into design here: power (usually measured in energy per bit being transferred) and bandwidth.
Bandwidth IF is designed to adjust the bandwidth of each major memory channel, creating potentially unified solutions without the use of buffers or large delays. Discussing IF in the server context is somewhat beyond the scope of what we tested in this article, but the point we are trying to get across is that the IF is built with a wide product coverage in mind. On the consumer platform, while the IF is not necessarily used for such a large degree on the server, the potential speed of IF to affect performance is just as high.
Test Bed and Hardware
As per our testing policy, we take the appropriate premium category motherboard for the socket, and complete the system with the appropriate amount of memory. By setting this test, we use BIOS to set the CPU core frequency using the straps available on the MSI B350I Pro AC motherboard. The memory is set to the speed range provided for our test.
In order to obtain a consistent set of results, the G.Skill Ripjaws V DDR4-3600 kit is set to the 17-18-18-38 latency as long as each of the different strings is tested. Due to the inability to support 100 MHz straps on our motherboards, the XMP profile is enabled in the BIOS on the Ripjaws V kit memory and adjust the latency time adjusted to 17-18-18-38 manually on the MSI B350I Pro AC motherboard. Each rope is set for purpose in continuity testing for the frequency scale.
Additional notes about our previous experience with memory scale. In the past we introduced the concept of Performance Index (PI) for each memory device, to provide rough performance comparison metrics between memory brands. This PI is defined as data rate (such as DDR4-2400) divided by CAS Latency (like 17 in 17-18-18) rounded down to the nearest integer. In previous articles like this, typically the highest PI memory gets the best overall score, especially in games, although in combination with similar PIs, those with the highest frequencies are often ahead. We will review this concept later in the review.
In this review, we will test the following combinations of data rates and latencies:
AGESA and Memory support
At the Ryzen launch, a number of industry sources personally revealed to us that the platform side of the product line is in a hurry. There is little time to do a complete list of DRAM compatibility, even with standard memory kits on the market, and this causes some problems for early users to try and get a suitable kit that works well without some tweaking. In a few weeks it's ironed when memory vendors and motherboard vendors have the time to test and customize their firmware.
Replacing this is a lower than expected DRAM frequency support level. During the launch, AMD has promised that Ryzen will be compatible with high-speed memory, but reviewers and customers are having problems with high speed memory kits (DDR4-3200 and above). These problems have been solved through a motherboard BIOS update wave built on the latest version of AGESA (AMD Generic Encapsulated Software Architecture), specifically up to version 1.0.0.6, but now preceded by 1.1.0.1 (thanks to the usual AMD system numbering).
While the maturity of the Ryzen platform is no longer a problem encountered, the AGESA 1.1.0.1 microcode is specifically focused on supporting Raven Ridge Ryzen 3 2200G and Ryzen 5 2400G APU announced before launch; we reviewed this BIOS update for AMD's Ryzen APU in February at launch.
The whole purpose of the current test is to evaluate the scalability of the Zen AMD architecture and to see if its performance can be affected by increasing DRAM frequencies. It would be foolhardy not to define the effect on game performance and to see if the memory frequency has a direct impact on the frame rate given that previous AMD APU generations have been reported to do so.