Review: AMD Radeon VII graphics card

Table of Contents
We analyze the AMD Radeon VII graphics card, which has problems with consumption, temperature, performance and price.
The AMD Radeon VII graphics card has recently been released. This new graphics card is based on the Vega 20 architecture that derives from the GCN design and with the new 7nm lithograph. It has been launched as an exclusive solution for the high end and it may be the only one that will arrive until the Navi architecture is ready. The Radeon VII arrives with many problems, including an out-of-market price, lower than expected performance, higher consumption than the Vega 10 and serious temperature problems.
This Radeon VII graphics card is basically a Radeon Instinct MI 50 launched for the gaming segment. A solution to get out of the way by the company that needed to bring to the market something that 'competed' with the NVIDIA RTX. We will talk in the review of several problems that we have found in this graph.
SPECIFICATIONS
- Architecture: Vega 20
- Lithography: 7nm
- DIE size: 331mm square
- Cores: 3840 Stream Processors
- FP32 Computing: 13.8TFLOPS
- Base frequency: 1400MHz
- Boost Frequency: 1750MHz
- Memory: 16GB HBM2
- Memory bus: 4096bits
- Bandwidth: 1TB / s
- L2 cache: 4MB
- TDP: 300W
TEMPERATURE MEASUREMENT
This new graphics card implements a total of 64 thermistors which have been strategically placed all over the GPU silicon. Thermistors now offer a large amount of data, offering protection for shutdown in case of exceeding the maximum threshold. In addition to being used for this, it now allows you to regulate and control ventilation.
AMD Wattman software It will indicate two types of temperature. The first will be the average temperature of the silicon and the second will be the most unfavorable point (Tº Junction). We have verified that the graph without load or at rest is at 40ºC, while at the minimum load it passes 80ºC and the Tº Junction indicates 110ºC at the most unfavorable silicon point. This causes the fans to jump to 2.900RPM, which generates a high level of noise, since the series heatsink has three fans.
This is a problem, since the graph has serious problems keeping the frequencies stable by fluctuating depending on the temperature of the silicon. This creates a significant imbalance in performance, as it must be kept within safe values.
THE GRAPHITE
Something quite interesting is that this graphics card does not use thermal paste as you would expect, but instead uses a graphite base sheet. This is first of all due to the thickness of this material, greater than that of thermal paste and allows that in areas where otherwise the contact would be worse, it improves. Therefore, heat transfer with this material is better.
We must bear in mind that the thermal conductivity of the typical thermal paste is 0.7-0.9W / m K, while the case of the graphite sheet is 23W / mK (It can reach 460W / mK, depending on the purity). So we can say that it is a very wise decision.
So much so that we have made a change between the graphite sheet and the thermal paste. By putting thermal paste between the DIE and the heatsink what we have achieved is to make things worse. We have tried to do several benchmarks, among them one of Final Fantasy XV at a resolution of 720p and graphic quality at a minimum and just five seconds after starting the test, the system has been turned off.
This is a serious problem. The graphite sheet prevents any maintenance, since it is not a product that can be easily found on the market. We must understand that not all sheets are valid, since this depends to a large extent on the purity and the combination of materials used for its development.
DURABILITY?
The temperature problems that we have detected (like all those who have analyzed the graph) lead us to wonder what is the durability of this graph. The thermal stress to which the silicon is subjected and the adaptability of the BGA solder tin balls on the PCB, since they will have to withstand thermal variation.
We must bear in mind that to eliminate the height difference and the separation of the DIE and the memories, in addition to in theory, contribute to improve the thermal conductivity. A host of factors that can end up generating problems with this silicon.
Given all these factors, it is possible that the durability of this silicon is limited and although we play an average of three or four hours a day, this thermal stress is bad for this graph.
HOW TO CONSUME IT
Without a doubt this is the most critical point. Despite the fact that Vega architecture has gone from 14nm lithography in Vega 10 to 7nm lithography in Vega 20 silicon, we have a significant problem in terms of energy consumption.
The Radeon VII has a average power consumption of 370Wbut in the Furmark this graphics card has reached 420W. We are talking about a graph that has a large energy consumption despite the reduction of the transistor node, which makes us think that something is not working correctly in this graph.
The consumption is much greater than that of the RTX 2080, a graphics card against which it should compete. According to the benchmarks that we have made to an RTX 2080, it never reaches 300W, but in the case of the Radeon VII it is much more inefficient, which is a highly negative aspect.
COMPONENTS FOR POWER SUPPLY
Powering this graphics card is not easy and requires a configuration and a combination of key components for its correct operation. The first element is a VDDCR_GFX five-phase controller controlled by an IR35217 International Rectifier. It has an IR3599 multiplier per phase that is combined with a system of 10 total voltage conversion circuits.
It makes use of 10 onboard Infineon TDA21472 power stages, each with a synchronous Buck gate driver, synchronous and control MOSFETs, and Schottky diode. The IR35217 is a multiphase controller that allows a 6 + 2 phase system to be offered with a second circuit supplying partial voltage for the VDDCI_Mem. A voltage controller has been implemented for this case IR35401M with a simpler PowIRstage.
The memory part is somewhat more complex. It also has an IR35217 for the generation of two real phases to power the four HBM2 memory stacks within the VDDCR_HBM circuit. It has two TDA21472s addressed by IR35217. The controller also offers the VDDCI_SoC. Instead it only generates one phase, which is then duplicated by the IR3599 using two phase shift voltage converters, each with its own TDA21472.
Circuits with voltages of 1.8V, 0.85V and 0.75V have been designed, which are generated by simpler Buck drivers. It is interesting that AMD implements on all 12V rails with suitable LC elements, something that should help smooth out the spikes.
SUPPLEMENTARY DATA
This graphics card has dimensions of 268mm long, 115mm of something and a thickness of 35mm and a total weight of 1.282Kg. The aluminum fin block, the case with the fans and the backplate weigh almost 1Kg.
The radiator block has vertically oriented fins and is traversed by five flat heatpipes that start from a central vapor chamber. To aid in heat dissipation, three 75mm fans have been arranged, which have an 82mm opening.
BENCHMARK
CONCLUSION
The AMD Radeon VII can not be classified as a gaming graphics card because it is a Radeon Instinct MI 50 with the name changed for the gaming market. It is an absurdly expensive graphic because it uses 16GB of HBM2 memory that we will never get performance for gaming. It has very serious temperature problems due to the fact that it is a DIE composed of five parts and this makes it difficult to develop a flat surface that allows good contact with the heatsink and causes that at some points 110ºC is reached, which means subjecting the silicon and the solders to great stress.
If we run a game, we stop having a graphics card and go to have an airplane turbine, since the three fans are put to work at 100% and it generates an awful noise.
The main problem is consumption. This power supply in peaks passes 400W of street consumption, which is a really worrying figure. Above all, it is worrying when compared to the Radeon Vega 64 based on 14nm lithography. This new graphic is based on the 7nm lithography and is still even more inefficient than the previous ones based on the Vega architecture. We see how the GCN design is more than obsolete and highly inefficient in terms of consumption, so AMD has to consider what to do at this time.
Radeon was a good division, but betting on the processor division and taking 60% off the budget has affected the development of graphics cards that can compete with NVIDIA solutions. Right now Radeon is its own entity that needs at least two years (as happened with the development of Ryzen) to return to being competitive at all levels.
To close, we do not recommend under any circumstances to buy this graphics card, since it is the worst on the market by far. Although it has positive aspects such as a large memory bandwidth, for gaming it is useless, neither the bandwidth nor the 16GB HBM2, since no game will take advantage of it.