AMD Radeon AI R9700 aka RX 9070XT in the big workstation graphics card test – What are the benefits of double the memory when working?

1 - Introduction and technical data 2 - Test system and equipment 3 - Autodesk AutoCAD 4 - Autodesk Inventor Pro 5 - PTC Creo 6 - Dassault Systèmes Solidworks 7 - Autodesk Maya 8 - SPECviewperf 15 (2025) 9 - Adobe Photoshop 26.10 10 - Adobe After Effects 2025 11 - Adobe Premiere Pro 25.41 12 - KI Benchmarks (AI Vision, Image, Text) 13 - Rendering 14 - Temperatures, clock rate, fans, noise and power draw 15 - Summary and conclusion

Temperatures

The temperature development of the Radeon AI R9700 shows very clearly that the thermal design was not fully utilized in several places, especially in the area of memory cooling. The recorded heating curve illustrates a rapid rise in temperatures immediately after the start of the load phase, followed by a comparatively early achievement of thermal equilibrium. This behaviour is generally typical for GPUs in this performance class, but the distribution of the individual temperature zones reveals a clear need for optimization.

While the GPU edge temperature of around 65 °C remains at an uncritical and perfectly acceptable level for air-cooled cards, the hotspot temperature of around 84 to 85 °C already shows a noticeably higher internal load distribution. However, a look at the memory becomes critical, as its temperature reaches the 90 °C mark several times. Although this level is still functionally within the specifications for GDDR6, it is clearly too high for a product that is marketed as a professional solution and should therefore meet higher demands in terms of thermal stability and component service life.

Particularly striking in this context is the almost complete lack of thermal coupling between the backplate and the memory modules. The thermography shows a broadly increased temperature on the rear side, which, however, remains significantly below the level of the actual memory. The measured temperature of around 69 °C on the backplate appears moderate at first glance, but at the same time it documents only very weak dissipation of the storage waste heat via this area. The temperature gradient between the storage tank and the backplate is clearly too large to assume functional heat dissipation through the rear.

This points directly to a fundamental problem in the layer structure of the cooling package. As I have already explained in my article comparing putty and thermal pads, the backplate is only able to make a significant contribution to memory or VRM cooling if it has defined contact points, a suitable interface material and a sufficiently rigid construction. A (different) thermal pad would have enabled a more stable and significantly more efficient heat transfer if it had been correctly designed and had a clearly defined pressure window. However, the present design lacks precisely this targeted thermal connection. The backplate acts more as a purely mechanical element and at best as a large-area heat spreader with no direct function for the critical components. The memory therefore has to dissipate the majority of its heat loss exclusively via the PCB and the front, which places unnecessary strain on the cooler and impairs the uniformity of the thermal profile.

For a card of this class, it would therefore not only be desirable but essential to integrate the backplate more actively into the heat flow. This would not only have reduced the storage temperatures, but also slowed down the thermal ageing of the components and improved the overall stability under continuous load. This is particularly true in a professional environment, where high continuous loads are common and thermal reserves are a decisive factor for long-term reliability. The measured 90 °C shows that potential has been wasted here. A revised design, for example with defined contact surfaces, improved interface materials or a modular backplate segmentation with high mechanical preload, would have achieved noticeable improvements. In its current form, the backplate of the Radeon AI R9700 is a purely decorative element with very limited functional relevance for memory thermal management. I would also like to refer you to my basic article here:

Thermal putty or thermal pad, which is better? Myths about thermally conductive materials, thermal conductivity and temperatures

Clock rates

The recorded clock curve of the Radeon AI R9700 under continuous full compute load shows a very consistent, but thermally explainable behavior that allows clear conclusions to be drawn about the internal regulation. Immediately after the load is applied, the GPU clock initially increases very quickly and reaches its maximum boost window within a few seconds. This initial peak is briefly above the later stable frequency, which suggests a temperature-based boost logic that utilizes its maximum reserves as long as the silicon temperature allows. However, as the hotspot temperature rises in parallel, the clock rate falls back to a sustained plateau relatively quickly. After about a minute, the hotspot temperature approaches the 80-degree mark, which apparently serves as a soft or fixed limit for the card. During the same period, the clock then stabilizes at around 2850 to 2900 MHz. This frequency remains almost constant until the end of the test, with only minimal fluctuations that can be clearly attributed to the thermal micro-fluctuations of the hotspot.

This means that the card is not limited in its continuous operation by the power consumption or the power supply, but primarily by the temperature. It is particularly noticeable that the hotspot remains just below the 90 °C mark during the entire test period, while the GPU edge sensor displays significantly lower values. This difference illustrates that very high temperature ranges form locally in the silicon, which is typical for GPUs with a high packing density. However, the behavior also shows that the cooling reaches its performance limits in these critical areas. Since the maximum possible boost frequency is obviously higher than the range permanently maintained in the test, it can be seen that the thermal control mechanism is permanently active and deliberately lowers the clock in order to control the hotspot temperature. This slight but continuous limitation explains why the clock cannot stay in the 3000 MHz range, but remains around 150 to 200 MHz below it.

This gives a clear picture: the Radeon AI R9700 achieves a stable but thermally throttled operating frequency under full load. The hotspot temperatures are so high that they have a direct influence on the applied clock rate and limit the boost mechanism. More efficient heat dissipation, particularly in the hotspot and memory areas, would extend the boost margin, increase the clock rate and also reduce the thermal load on the GPU in the long term.

Fan speeds and control

The analysis of the fan speeds of the Radeon AI R9700 under full load shows very clearly how much the cooling is required in this scenario and what consequences this has for the noise development. Even in the first few seconds of the load increase, the fan control reacts with a rapid speed increase, which is intended to absorb the initial temperature peak. The course of the curve shows that the control is designed aggressively and that there is hardly any downward leeway as soon as the hotspot temperature rises towards the known thermal threshold.

After around two minutes, the fan speed has reached an almost constant operating level and settles at around 3200 to 3400 revolutions per minute. This level is maintained over the entire duration of the test, whereby small fluctuations repeatedly reflect the dynamic reactions of the card to microthermal changes. The hotspot temperature is constantly just below 85 degrees during this phase, which shows that the fan speed was deliberately set to an upper limit in order to ensure thermal stability.

A speed above 3000 revolutions per minute means a very clearly perceptible operating noise in practice. The sound pressure of a typical axial fan does not increase linearly, but disproportionately with the speed. At just under 3400 rpm, you can expect an acoustic impression that is far above the usual everyday gaming noise and is more reminiscent of a workstation that is permanently running at high capacity. It must therefore be assumed that the operating noise will be loud and disturbing in quiet rooms, especially if the card is operated in compute or rendering workloads for a longer period of time.

The fan curve ultimately reveals two points: Firstly, it confirms that the card’s cooling solution has no significant thermal reserves under high continuous load and therefore has to regulate aggressively. Secondly, it indicates that this high speed is not only necessary in the short term, but permanently in order to keep the critical areas of the GPU in the safe range. For practical operation, this means that users must expect a very high noise level that cannot be reduced without interventions such as a manual fan curve or the use of an alternative cooling concept.

Frequency spectrum and measured noise level

The frequency spectrum recorded with Smaart shows an overall strongly increased level over the entire audible band, which already indicates a high broadband flow and motor noise of the fans. The measured SPL of 57.9 dB(A) under load confirms the visual perception of an acoustically dominant noise level, which in this form is clearly too high for a professional workstation card.
It is noticeable, however, that the usual uniform noise curve is interrupted by several narrow-band level increases. These peaks are typical of electrically induced oscillations of coils and inductances, i.e. classic coil beeping. The spectral distribution shows striking vertical structures in the range between approx. 700 and 1000 Hz as well as further energy increases in the upper mid-range between 1.5 and 2 kHz.

These frequency groups correspond to a very typical pattern of heavily loaded DC-DC converters that lock into fixed switching frequencies in compute mode. The repetition of these harmonic lines in the spectrogram confirms that these are not fan resonances, but clearly electrically induced coil whine, which clearly stands out both tonally and energetically.

Analysis of the audio recordings of the whine

During the warm-up phase, the electrical load on the voltage converters increases in proportion to the power consumption of the GPU. It is precisely during this phase that the load profiles change from short-term boost behavior to a stable constant load range. It is precisely during this transition that the strongest changes in the acoustic spectrum occur. The rising temperatures lead to a slight shift in the switching frequencies, which explains the observed frequency jumps between around 690 and 930 Hz in the first recording. As the GPU gradually reaches thermal saturation at over 80 °C, the frequencies condense into clearly defined tones as the VRM switching frequencies fall into stable grids.

Although the fans ramp up considerably shortly afterwards and the broadband airborne noise increases significantly, the coil whine remains prominently visible in the spectra. Even at levels above 3000 rpm, at which a masking effect would normally occur, the whistling remains tonal. This implies that the sound pressure of the electrical vibrations is disproportionately high in relation to the fan volume – a rather unfavorable noise profile. This is a significant disadvantage, especially in the professional sector, where the Radeon AI R9700 is actually intended to be used, as workstations are often operated in quiet development environments or near sound studios. The coil whine is therefore not only technically relevant, but also ergonomically problematic.

The Heating Up Sequence recording shows that the whine is already audible during thermal heating. The spectral peaks around 700, 810 and 920 Hz already occur before the fans turn up significantly. The beeping is clearly recognizable in this phase and has a stable frequency structure, which is typical for load currents that still increase moderately and evenly.

The subsequent numerical analysis of the audio recordings confirms this visual observation. The FFT analysis of the Heating Up Sequence file shows repeated coordinate clusters of pronounced peaks at:

• ~690-700 Hz

• ~810 Hz

• ~920-930 Hz

This corresponds exactly to the harmonic patterns that occur in GPUs when load changes in the VRM or in the memory voltage converters excite certain natural frequencies of their inductances.

In the Coil Whine 01 and 02 files, the whine increases in both amplitude and complexity of the harmonics. The fundamentally dominant frequency around 1005 Hz is very present and produces a penetrating, high-frequency whistle that remains clearly perceptible even when masked to a certain extent by airborne sound. The subharmonics around 502 Hz also produce a pulsating whistling pattern that can be subjectively perceived as “trembling” or “fluttering”. The higher harmonics around 1509 Hz lead to an unpleasantly sharp characteristic, which is particularly annoying in office environments, as they are in the range of maximum human hearing sensitivity.

In the Coil Whine 02 file, the peaks are even more pronounced. Particularly high amplitudes occur at:

• ~1004-1007 Hz (dominant fundamental)

• ~502-503 Hz (first subharmonic)

• ~1509 Hz (second harmonic)

The energetic dominance of these tone complexes leaves no doubt that the coil whine – not the fans – is the critical noise source of the Radeon AI R9700.

Power consumption

The power consumption of the Radeon AI R9700 reveals a clearly structured, yet demanding load behavior in the practical test, which clearly distinguishes it from classic gaming graphics cards. Measurements were taken directly on the graphics card so that the GPU chip as well as the memory, voltage converter and fan control were reliably recorded. Even in idle mode, the card behaves inconspicuously, as the consumption is usually in the range of around 30 to fifty watts, which is absolutely typical for a workstation GPU in this performance class when the fan is running. Under light load, the power consumption increases moderately and reaches around 100 watts, for example with simple 2D applications or minor GPU acceleration.

However, as soon as full 3D visualization kicks in, the picture changes significantly. In this discipline, the R9700 requires a constant 240 to 300 watts and is therefore in a range that is not unusual for high-end hardware, but clearly shows that the chip reaches areas of high utilization relatively early on. This applies in particular to ray tracing workloads, in which both shader and memory paths are fully utilized. Here, the consumption in the test setup regularly rose to values between 330 and 360 watts. Only in pure compute and AI load scenarios, in which the card works practically continuously at the limit, did the maximum power consumption of around 400 watts become apparent, which is far above the specifications and pushes the card to its thermal limit (see above).

These consumption values clearly place the Radeon AI R9700 in the upper class of professional workstation GPUs, whose efficiency behavior depends heavily on the respective application. It is noticeable that AMD allows very generous power distribution in the compute area, which on the one hand explains the high computing power, but on the other hand also significantly increases the thermal requirements. For systems that are to work under full load for long periods of time, a high-quality power supply unit is just as essential as sufficient case ventilation.

Perf-per-watt comparison between Radeon AI R9700 and Radeon RX 9070XT

If we compare the measured consumption values with the previously determined performance in rendering, CAD and AI workloads, a differentiated efficiency picture emerges. The Radeon RX 9070XT works somewhat more economically in many tasks, but does not quite reach the throughput values of the AI R9700 in compute-intensive areas. The AI-optimized variant often achieves a small but reproducible advantage per watt of computing power, as it converts its higher consumption reserves better into actual output. The RX 9070XT is more efficient in classic graphics workloads, while the R9700 clearly shows its strengths in AI, simulation and complex shader pipelines. Overall, the efficiency of the two models is close to each other, but is decisively shifted by the type of application.