The Radeon RX 5700 XT and RX 5700 presented

– AMD wants to attack Nvidia again with the Radeon RX 5700 XT and the Radeon RX 5700. With the new RDNA architecture and a significantly increased performance per cycle and energy efficiency, the middle class should have a new winner from July 7th.

With the Radeon RX 5700 XT and the Radeon RX 5700 AMD has presented the first graphics cards of the Navi generation. The graphics cards will beat the GeForce RTX 2070 and RTX 2060 and will cost $ 449 and $ 379 respectively. The market launch is scheduled for July 7th, and thus at the same time as the Ryzen 3000 CPUs. The technology used is the new RDNA architecture, which replaces GCN.

The key data of the Radeon RX 5700 XT and Radeon RX 5700

First of all, the rough basic data. The Radeon RX 5700 XT offers 40 compute units and thus the maximum expansion of the Navi 10 GPU. Also at RDNA, a CU consists of 64 individual ALUs, which means that the fastest Navi expansion level has 2,560 shader and 160 texture units. The maximum boost clock is 1.905 MHz, the base clock is 1.605 MHz. There is also a new "game clock", which is specified at 1.755 MHz – more on that later. An 8 GB large and 7,000 MHz clocked GDDR6 memory ensure a memory bandwidth of 448 GB / s. The board power is 225 watts. Since RDNA, according to AMD, has 25 percent more power per cycle, the Radeon RX 5700 XT is supposed to work faster than a Radeon RX Vega 64, which actually has a much higher raw power with around 13 TFLOPs. The Radeon RX 5700 XT has only 9.8 TFLOPs in comparison. According to AMD, the Navi graphics card is supposed to beat a GeForce RTX 2070. If this is confirmed in our own benchmarks, RDNA is only slightly behind Turing in terms of speed per cycle. The Radeon RX 5700 XT's full retail price is $ 449.

The Radeon RX 5700 is the second new graphics card that uses a partially deactivated Navi 10. 36 compute units are active, which amounts to 2,304 ALUs and 144 TMUs. The clock rates are significantly lower at 1,465 MHz (base), 1,625 MHz (game) and 1,725 ​​MHz (boost). As compensation, it remains with the full 64 ROPs and the full 256 bit interface. The 8 GB GDDR6 memory also works with the maximum 7,000 MHz. The board power is 180 watts. According to AMD, the graphics card should work a good deal faster than a GeForce RTX 2060. There is no performance comparison to the Radeon RX 5700 XT or an older Radeon graphics card. The Radeon RX 5700's full retail price is $ 379.

PCIe 4.0, DSC compression, a new video engine and more

AMD has also made changes to the navigation system that go beyond the actual architecture. So Navi supports DisplayPort 1.4 including DSC compression, with the help of which a cable is sufficient for, for example, 8K resolution at 60 Hz. Navi is thus leveling with Turing, because the predecessor Vega only had DisplayPort 1.4 without DSC. With the HDMI standard, however, version 2.0b remains. According to AMD, Navi was too far in development to integrate HDMI 2.1 when the final specifications were announced.

Even if the Navi reference design does not have a VirtualLink, AMD has inadvertently revealed that Navi 10 is capable of controlling a VirtualLink connection with USB Type-C. It is still unclear whether onboard partners are free to block them.

PCIe 4.0 is activated

The Vega 20 GPU of the Radeon VII can handle PCIe 4.0, but only PCIe 3.0 is activated on the consumer graphics card. This is different on the Radeon RX 5700 XT and the Radeon RX 5700, because AMD advertises PCIe 4.0 for both 3D accelerators. This makes AMD the only manufacturer that can offer a complete package consisting of a processor (Ryzen 3000 with Zen 2) and a graphics card that supports PCIe 4.0. For gamers, the interface on the graphics card will initially bring only minimal or no benefits. Unless you use a CrossFire network, since most mainboards throttle the PCIe speed on x8 lanes and PCIe 4.0 with twice the bandwidth is then an advantage.

The video engine is significantly faster

AMD has significantly expanded the video engine, which is called "Radeon Media Engine" on Navi. It can now decode the H.264 codec with up to 4K150 (4K resolution at 150 FPS), 1,080p / 600 or 8K30 and encode 4K90 or 1,080p / 360. H.265 videos can be decoded at 1,080p / 360, 4K90 or 8K / 24 and encoded at 1,080p / 360 or 4K60. The VP9 codec that YouTube uses is supported with 4K60 and 8K24 when decoding. In contrast, AMD does not speak of the new AV1 codec to which YouTube will switch, so it is still only calculated by the CPU.

The reference design from Navi

Both the Radeon RX 5700 XT and the Radeon RX 5700 will initially only be available as reference designs. This should also be available in Germany at AMD.com or as usual from the on-board partners in retail trade. AMD has explicitly mentioned that you do not want to compete with the AIBs with your own sales. Partner cards with their own cooling design and higher clock rates are to appear in the course of August in the current state.

The same cooler, but one with a dent

The most important thing about the reference design first: Unlike the Radeon VII, AMD uses an almost identical blower cooler with a radial fan for the Radeon RX 5700 XT and the Radeon RX 5700. Although AMD is aware that this is potentially louder than an axial fan, you want to offer a comparable experience regardless of the case ventilation. BitcoinMinersHashrate could hear the graphics card briefly during load. And so much can be said: The Radeon RX 5700 XT in reference design will not be quiet. Despite a fairly high noise level in the area, there was a lot of noise.

The cooler on both navigation models is identical. AMD relies on a vapor chamber and an aluminum heat sink for heat transfer. The connection between the GPU and the cooler is made like on the Radeon VII with a graphite pad, which transfers the heat well, but has to be replaced after removing the cooler once. The radial fan on the graphics card, which is also always in operation on the Windows desktop, is relatively large.

Even though the cooler itself is identical on both models, the Radeon RX 5700 XT shows a dent in the aluminum casing on the upper side just before the radial fan. AMD openly admits that this is primarily for optical reasons. However, this should also result in the volume at low fan speeds and thus, especially on the desktop, being minimally lower than the Radeon RX 5700 without dent. Both graphics cards come with a backplate.

OC should be possible with both navigation reference cards

With the reference design, AMD also wants to enable a good overclocking experience. The identical PCB on the Radeon RX 5700 XT and the Radeon RX 5700 requires a 7-phase power supply. In addition, there is an eight-pin and a six-pin connector on both 3D accelerators, so that in theory a power consumption of around 300 watts is possible.

Visually, the Radeon RX 5700 XT makes a good impression. The reference design also looks high quality when touched due to the materials. The Radeon RX 5700 could only be inspected by the editors.

AMD's own benchmarks for the Radeon RX 5700 XT and Radeon RX 5700

AMD has also made some benchmarks for both Navi graphics cards – so they should be enjoyed with a good dose of caution. According to AMD, both navigation graphics cards are primarily intended for gaming in 2,560 × 1,440 or a comparable resolution. This shows that these are mid-range models, otherwise AMD would advertise with 3,840 × 2,160. The Radeon RX 5700 XT is intended for WQHD with maximum graphics details, the Radeon RX 5700 for WQHD with (partially) reduced graphics details.

AMD has produced three different series of measurements. The Radeon RX 5700 XT is compared against the GeForce RTX 2070. In ten games, the Radeon wins eight games with a lead of two to 22 percent – usually the plus is around five percent. In contrast, two games are lost with a gap of one to three percent.

The benchmarks of the Radeon RX 5700 against the GeForce RTX 2060 are clearer. Because all games are won by the AMD graphics card with a lead of five to 23 percent. The increase is usually around ten percent.

And then there is a benchmark of the Radeon RX 5700 XT against the Radeon RX Vega 56. The new graphics card wins all tests with a plus of 24 to 41 percent. The smallest victory is achieved in Far Cry New Dawn, the largest against it in both Dirt Rally 2.0 and Metro Exodus. Mostly, the plus is around 30 percent.

An attempt to classify the benchmarks

If you compare the AMD benchmarks of the two AMD graphics cards with the new BitcoinMinersHashrate test course in WQHD, the Radeon RX 5700 XT would be around 15 percent faster than a Radeon RX Vega 64 and just under ten percent faster than the GeForce RTX 2070. Compared to Radeon VII, the backlog in WQHD would not be quite ten percent. The Radeon RX 5700, on the other hand, should be about the same speed as the Radeon RX Vega 64. The editorial team feels that the classification is rather optimistic. Only your own benchmarks will show the correct order here.

The new game clock

In addition to the already known basic and turbo clock, there is now a game clock on Navi. The first two work on the Radeon RX 5700 XT and Radeon RX 5700 as well as on older Vega graphics cards. The game clock, however, is just like Nvidia's "average turbo" a fictitious value that says the same thing. The game clock should simply show the frequency with which the graphics card usually works when playing.

AMD wants to have chosen the game clock carefully. In most games, the frequency should be a bit higher. In general, the clock should fluctuate between the slightly higher game clock and a value close to the boost clock. The base clock, on the other hand, should only apply to "power viruses" such as the Furmark and the maximum turbo should only be applied at low loads and low temperatures.

No ray tracing for satnav

A big unknown was whether and if so, how Navi can handle ray tracing in games. In a nutshell: The Radeon RX 5700 XT and the Radeon RX 5700 do not master ray tracing in games. Not even about the shader units.

According to AMD, ray tracing should only be available on "RDNA 2" or "Next Gen RDNA". Then there should be hardware support for "Selected lighting effects for real time“Give on a radeon. How exactly this is then implemented, whether comparable or different from Nvidia on Turing, is still unclear.

It can be assumed that the future high-end models from AMD will be equipped with RDNA 2. At Computex it was heard that the corresponding graphics cards will probably open in the middle of next year.

What Navi does not support are special units for accelerating AI algorithms. There is no such thing as tensor cores.

First details on the RDNA architecture

The Radeon RX 5700 XT and the Radeon RX 5700 are the first graphics cards to come with the new RDNA architecture. Even if something else is often said in various forums, this should consistently be a new development, even if well-functioning parts have of course been more or less taken over by GCN. RDNA was developed differently than GCN for games, whereas GCN is rather intended for a high data throughput and thus for compute applications. That is why GCN should live on in the professional segment, new products are also being planned. RDNA should nevertheless also offer high compute performance and replace GCN at a later date for the professional models.

Both new AMD graphics cards rely on the Navi 10 GPU. This is manufactured in a 7 nm process at TSMC and has a size of 251 mm² for 10.3 billion transistors. This makes Navi 10 slightly larger than its mid-range predecessor Polaris 10, but also offers almost twice the number of transistors. Vega 10, on the other hand, is twice the size, but has only two billion additional circuits.

The biggest differences from RDNA to GCN are to be found in the compute units (better efficiency for gaming requirements), the cache structure (lower latency, better utilization of bandwidth, lower power consumption) and the graphics pipeline (higher performance, higher clock rates) his.

The basic structure is not that different

The basic structure of Navi 10 with RDNA is similar to Polaris 10 and Vega 10 with GCN, but differs in some aspects. For example, there is only one Geometry Processor at RDNA compared to the four on Vega, but it has become significantly more powerful. In addition, this works in conjunction with four new "Prim Units" (one per shader block), which have comparable tasks, although all shader blocks can fetch data from the Geometry Processor without having to calculate them themselves in the Prim Units.

There are also four ACE and one HWS unit (instead of two), which is responsible, among other things, for the compute utilization of the shader blocks. However, these have been drilled out and can now put tasks on hold in order to get values ​​of already started tasks first.

New CUs for higher utilization

The RDNA Compute Units are completely new, of which there are ten per shader block. These are summarized on RDNA as a "Dual Compute Unit", so there are five per block. This has the advantage that if the CUs have to perform the same computing tasks, this relieves the workload and there is less idle time. But the individual CUs also work differently.

Instead of one scalar unit (for scalar arithmetic operations) there are now two for each CU on RDNA. And the schedulers for task allocation within the CUs are doubled. In addition, an RDNA-CU is no longer constructed as a 4 SIMD16 vector unit (one instruction for multiple data) like GCN, but as a 2 SIMD32 vector unit. While GCN can distribute a work task to the ALUs every four cycles, RDNA can do this every single cycle, since each SIMD on RDNA can also start a "computing job" – at GCN there was only one starter for all four SIMD units. Although this does not increase the maximum data throughput of a CU, it does significantly increase the load. Especially for games, this should bring great advantages.

A new cache for more bandwidth usage

The third major innovation in RDNA is a new cache hierarchy. Every CU now has a new L1 cache. There is generally 512 KB, so 128 KB are available per shader block. At 4,096 MB, the L2 cache is twice as large as on Polaris and has remained the same as on Vega. However, the cache has been optimized so that less data movement (saves energy) is required and the latency of each cache has been reduced.

In addition, the cache has been optimized so that the ALUs can now write and read compressed data throughout. Therefore, DCC memory compression on RDNA should work much better. So far, this has not been really effective on Vega either.

Further details on the architecture of Navi and RDNA will follow in a later article.

BitcoinMinersHashrate has received information about this report from AMD at NDA. The only stipulation was the date of publication, which was set by AMD. The provision of these materials was preceded by participation in a press event in mid-June in Los Angeles, USA. The costs of arrival and departure and accommodation were borne by AMD.

This article was interesting, helpful, or both? The editors appreciate any support in the form of deactivated ad blockers or a subscription to BitcoinMinersHashrate. More on the topic of ads on BitcoinMinersHashrate.