We covered AMD ads on Ryzen and Navi during E3 throughout the week. We still have an aspect of the situation to discuss. We discussed Navi and its RDNA architecture, but we did not discuss any of the software enhancements that AMD plans to offer with its next GPUs. Some of these gains will also be available for GCN cards.
Let's talk about some features and improvements.
First, there are the quality of life gains generated by AMD's Radeon software. With Navi, the system automatically switches your TV to low latency game mode, if the display supports one. You will be able to save the settings to separate the files and reimport them if you need to install the driver completely from scratch or if you reinstall your entire operating system. Some improvements have also been made to the way WattMan reports its results.
The AMD Link streaming application now supports streaming on TVs, including Apple and Android TV. VR wireless streaming is now supported as well. These enhancements are not associated with any specific GPUs.
Radeon Chill is AMD technology to reduce the power consumption of GPUs during games. The software can now set frame rate limits on 60Hz screens to reduce the number of images rendered when you do not actively control your character due to your afk.
The AMD footnote on Radeon Chill deserves to be read. Under the right circumstances, this can significantly reduce the power consumption of the graphics processor, although this has an impact on the rate, and the total size of the gain varies from one title to the other. Any graphics processor that previously used Radeon Chill can take advantage of these enhancements.
Then, Radeon anti-lag. According to AMD, the company has invented a method to reduce the time between the moment you press a button in a game and the one where you see the results. To do this, some CPU jobs are delayed to ensure that they occur simultaneously alongside the GPU rather than being completed in advance.
Honestly, I can not say that I have observed a difference between the activation of Radeon Anti-Lag and its deactivation. AMD demonstrated that the effect worked with custom-built latency monitors attached to displays, and I think the company on which the monitor I tested had slightly higher latency. I am at an age when motor reflexes have already begun to decline, and if I am honest, I have never been a very good twitch player.
In the best case, this feature reduces your total latency by a few milliseconds. If you are good enough to compete in these spaces, it could be worth something. This is not something I feel able to comment on.
The anti-lag is supported in DX11 on all AMD GPUs. Support for DX9 games is a Navi-only feature. DX12 games are currently not supported because of the extremely different implementation requirements in this API.
Radeon Image Sharpening is a feature that combines adaptive contrast sharpness with the use of GPU resizing techniques to improve the quality of the base image without the need for a native 4K rendering penalty. The following slides compare RIS enabled or disabled.
RIS is disabled in the slide above.
RIS is enabled in this slide. The effect is very subtle. You may want to open the two images above in separate tabs, zoom in carefully, and then compare the final product. Although there is a clear improvement in the IQ in the picture "ON", it is a small one.
Nevertheless, small improvements to IQ are generally welcome. RIS was also designed by Timothy Lottes, who worked on FXAA at Nvidia. The use of this feature should not affect performance (the impact on performance is estimated at 1% or less). RIS is a Navi-only feature and is only supported by DX12 and DX9.
Finally, there is FidelityFX.
FidelityFX is AMD's new addition to GPUOpen. It is offered to any developer wishing to take advantage of it. Adaptive Contrast Sharpness can be used on any GPU if developers want it.
Some extra hardware details on Navi that were not in the previous articles, but probably should have (a frenetic briefing schedule and some scrambled note taking):
AMD plans to maintain GCN GPUs on the market to manage HPC workloads. The AMD engineer we spoke to compared GCN to an extremely efficient sword was it balanced properly, but that it was relatively tedious to use, while RDNA was more of a lightsaber in terms of concentration on elegance and the economy of movement. GPUs such as the MI50 and MI60 also offer much larger memory bandwidth and larger memory pools than any of the Navi cards coming on the market.
The RDNA should eventually replace the GCN in this space and correct some of the slow path anomalies that the GCN suffers. Irregular performance with some texture formats has been corrected, for example, and the RDNA has larger caches to prevent bubbles in the pipeline. Overall performance should be more predictable with rDNA-derived GPUs than with GCN.
There is nothing new in these details, but I thought to include them for the sake of completeness. This concludes our E3 coverage.
AMD announced last night its new Radeon GPUs at E3. The new Radeon RX 5700 and 5700XT are positioned as responses to the Nvidia RTX 2060 and 2070 family, rather than a frontal assault on the RTX 2080. As previously announced, the Radeon VII will remain on the market over the RTX 2080 .
Navi, however, seems to be a significant step forward for AMD on several fronts. We will have a deep architectural dive in the near future, but for now, let's look at speeds, flows and competitive positioning. You can click all the slides to open a larger version in a new window.
The Radeon 5700XT is a 40-unit design, with 2,560 stream processors, 9.75 TFLOP floating point performance, and clocks far superior to anything we've seen before. AMD's new RDNA architecture, which finally GCN replaces, is significantly more efficient than its predecessor, with a projected increase of 1.25 times in performance per clock. The TDP on the 5700XT is 225W, compared to 295W on the Vega 64. The power supply is via an 8-pin connector and a six-pin connector.
The Radeon 5700 is a 36 CPU design with 2,304 stream processors and the same 8GB RAM pool as the 5700 XT. As expected, there is no sign of HBM on these products. The 5700 features a 180 W TDP and the same 1x 8-pin + 1x 6-pin power distribution system.
According to AMD, its new RDNA architecture is a major improvement over GCN, with substantial improvements in clock performance, raw clock speed, and watt performance.
The performance improvement per watt of 1.25x does not take into account clock speed gains but adds to them. It's a good time to talk about the newly defined synchronization scheme by AMD, so let's talk about it.
The basic clock of these cards is equivalent to what you will see if you run a type of workload type of power virus like Furmark. The "Game Clock" is a conservative estimate of the clock that you will see when running current titles over long periods of time. According to AMD, this is not the GPU's median clock rate over time during gaming – it's actually a bit lower than expected to allow for variations in silicon and cooling from one system to another. AMD derived its Game Clock values by measuring the average speed of the game's graphics processors on 25 different games.
The Boost Clock is an opportunistic clock that the GPU will try to hit when possible. Even this value does not represent the maximum potential speed (AMD has described it as "close to the maximum"). With improvements to the underlying architecture and overall design, GPU clocks are much higher than anything we've seen before from AMD. (We will talk about this in more detail in the coming days.)
According to AMD, the power improvements to give RDNA 1.5 times better performance in limited power environments than an equivalent GCN configuration. The change to 7nm represents just over 20% of the total gain, with improvements in design and power of about 15-18%. Most improvements result from improving the performance of the core of the graphics processor. We will write as teaser for deep diving: RDNA can execute instructions at each cycle, compared to GCN, which takes at least four cycles. The net result of these enhancements is a significantly better GPU than the cards it replaces, dramatically improving performance and simultaneously reducing power consumption.
AMD expects the Radeon 5700XT to deliver performance around 1.14 times faster than Vega 64, while consuming 23% less energy. TDPs rated on the 5700XT and 5700 are still superior to their Nvidia counterparts, but TDP is not a substitute for power; we will have to test the hardware to see how the GPUs compare. The performance improvement by area is substantial. Vega 64 was 495mm2 part, while Navi is at 251mm2 part. The RTX 2060 and 2070, on the other hand, are 445mm2.
These gains should put the 5700XT slightly higher than the GTX 1080 and the RTX 2070 in terms of overall performance. AMD also told us that we have read our reactions in terms of colder reference noise.
The 5700 and 5700XT models not only use wrapped fans that vent heat from the system, but AMD also promises to lock them to a volume of 43dbA. (It is not known whether it is absolute maximum volume or maximum volume provided you do not manually set the fan to 100%). This should respond to one of the recurring complaints about AMD's reference cards that they are often much more powerful than the competition. New buyers will receive a redeemable card for a three-month subscription to the Microsoft Xbox Game Pass for PC service.
We will have much more to say about Navi and its underlying architecture in the days to come. The GPU seems to have taken a significant step forward in reducing AMD's power over Nvidia and appears to be reaching a higher performance / watt target than Radeon VII.
The price of both cards is mobile, depending on the performance shown. The Radeon RX 5700XT will be available for $ 449, while the RX 5700 is $ 379. The price of the RX 5700 is significantly higher than the current RTX 2060 cards, which sell for $ 335 at the bottom of the market. The RX 5700XT is a $ 450 card, compared to the $ 500 price of the RTX 2070. A 50th birthday of $ 500 from the card will also be available.