- Thermals & graphite pad vs. paste
- Pressure sensor testing
- PCB quality & VRM analysis
- Gaming performance
- Acoustic performance
- Power consumption
- OpenCL performance
As always, the article carries on after the embedded videos.
You can find our previous tear-down here. Our VRM analysis is also below:
Testing methodology has completely changed from our last GPU reviews, which were probably for the GTX 1070 Ti series cards. Most notably, we have overhauled the host test bench and had updated with new games. Our games selection is a careful one: Time is finite, and having analyzed our previous testing methodologies, we identified shortcomings where we were ultimately wasting time by testing too many games that didn’t provide meaningfully different data from our other tested titles. In order to better optimize our time available and test “smarter” (rather than “more,” which was one of our previous goals), we have selected games based upon the following criteria:
- Game Engine: Most games run on the same group of popular engines. By choosing one game from each major engine (e.g. Unreal Engine), we can ensure that we are representing a wide sweep of games that just use the built-in engine-level optimizations
- API: We have chosen a select group of DirectX 11 and DirectX 12 API integrations, as these are the most prevalent at this time. We will include more Vulkan API testing as more games ship with Vulkan
- Popularity: Is it something people actually play?
- Longevity: Regardless of popularity, how long can we reasonably expect that a game will go without updates? Updating games can hurt comparative data from past tests, which impacts our ability to cross-compare new data and old, as old data may no longer be comparable post-patch
Game graphics settings are defined in their respective charts.
We are also testing most games at all three popular resolutions – at least, we are for the high-end. This includes 4K, 1440p, and 1080p, which allows us to determine GPU scalability across multiple monitor types. More importantly, this allows us to start pinpointing the reason for performance uplift, rather than just saying there is performance uplift. If we know that performance boosts harder at 4K than 1080p, we might be able to call this indicative of a ROPs advantage, for instance. Understanding why performance behaves the way it does is critical for future expansion of our own knowledge, and thus prepares our content for smarter analysis in the future.
For the test bench proper, we are now using the following components:
GPU Test Bench (Sponsored by Corsair)
This is what we’re testing!
Often the company that makes the card, but sometimes us (see article)
Radeon VII Graphite Pad vs. Thermal Paste Benchmark
Thermal testing is very interesting for this one: We discovered that AMD is using a Hitachi HM03 graphite thermal pad for Radeon VII. Under normal circumstances, despite the higher thermal conductivity – although the measurement of this is non-standard and difficult to compare – paste should be better as a result of being significantly thinner. A pad would be better for uneven contact, for instance, and so that’s the first thing we tested after we completed all standard benchmarks. Again, to be clear, this process didn’t begin til everything else was done.
After carefully removing the graphite thermal pad, we applied pressure paper to determine how well the cooler coldplate contacted the die. Despite the epoxy resin coating evening-out the die, the cooler coldplate itself did not make even contact, allowing a significant area of the die to go with minimal or no contact when no thermal interface is applied. This could caused by the coldplate or by the mounting hardware used by AMD. We tested this three times, carefully using the right torque and same mounting patterns each time. This supremely uneven contact is what drove the need for a thermal pad, as the pad allows for more error and is thicker, ensuring contact. We next applied Thermal Grizzly Kryonaut and ran a round of thermal tests.
First, for stock performance but with the fan locked to 2900RPM and ambient logged, we measured junction temperature as climbing rapidly to 107-108 degrees Celsius. Junction temperature is new to Radeon settings and is effectively the GPU-z hotspot temperature: AMD uses its network of 64 thermal sensors across the chip to find the hottest spots, then clocks according to that one hotspot. This is as opposed to guessing based on an edge temperature, which is what the traditional “GPU temperature” is – we’ll plot that line now. GPU temperature is measured at the edge of the die and is cooler than the center of the die, which is where Junction will pull data. Junction temperature has a TjMax of 110 degrees stock and 120 degrees when overclock settings are applied, assuming any of them worked. More on that later. Junction temperature is the most important now, as it will dictate boosting headroom. For that second line we drew, that’s the stock GPU temperature at about 80 degrees Celsius with our fixed fan RPM.
Kryonaut Junction temperature ramps harder and hits about the same range, but it’s clear that we’re throttling against the TjMax of 110 degrees. We’ll validate with frequency momentarily. Here’s where it gets really interesting, though: If we plot the GPU temperature for Kryonaut, we see that it’s actually lower – and a good bit, too, dropping from 80 degrees to 75 degrees. This establishes that, yes, thermal paste does perform better than a pad, but only in instances where contact is already good. The thinner interface always excels versus a thicker, higher conductivity interface. You want as little in the way of transfer as possible; unfortunately, despite leads in edge temperature of 5 degrees, junction temperature is maxing-out and causing thermal throttling. Our earlier pressure-mount testing demonstrated exactly why this happens: thermals at the center of the die are bad because contact is non-existent or very limited without that thermal pad. This is some very interesting data and establishes that AMD’s decision to use a thermal pad was correct, but it would still be better for AMD to just use a cooler that properly contacts the die and then use paste.
Just to validate that data, here’s a quick frequency plot. Some spikes are normal as part of this power virus load, so we can ignore those as artifacts of the program’s behavior; it’s the more frequent, smaller dips of the Kryonaut test that are relevant. In this example, the loss of contact at the center of the die is causing us to trip TjMax and throttle down from about 1670MHz to 1520MHz. This is noteworthy and would show up in data if run as such. We did this after all other testing, though, and we have already replaced the thermal pad with the same model pad as the original.
We have a whole host of additional thermal testing on Radeon VII, but we’re going to shunt that to the end of the content. Let’s get into overclocking discussion and gaming performance.
Sniper Elite 4
Sniper Elite 4 comes up first. This is a title where AMD has traditionally excelled in comparison to other games, largely because Sniper Elite 4 is a well-built DirectX 12 title with asynchronous compute support, leveraging additional hardware on AMD for asynchronous command queuing.
Tested first at 4K/High, before showing frametimes, we see that the AMD Radeon VII stock GPU averages at about 85FPS, with lows at 73 and 71FPS 1% and 0.1%. These lows are indicative of good frametime pacing and, given the stack-up in this game, are also indicative of just generally good development practices and the Asura engine. Comparatively, the NVIDIA RTX 2080 FE runs at about 84FPS AVG, which is within our error margins for this test – it is functionally the same as the AMD Radeon VII and inside of error. The GTX 1080 Ti operates at 87FPS AVG, within range of both the 2080 and Radeon VII, albeit slightly outperforming both. Compared to our modded Vega 56 with power mods and heavy overclocks, maintaining a 64FPS AVG, the Radeon VII is about 33% faster. Keep in mind that Vega 56 with power mods is easier to overclock than Vega 64, an advantage of the reduced CU count.
As a reminder, frametime plots are the best way to illustrate frame-to-frame pacing without obscuring data with averages. We’re looking first for consistent frametimes, illustrated on the left axis, and next for lower numbers. The time observed is the time from one frame present to the next, with a present happening to update the screen.
The AMD Radeon VII plots first, illustrating a rough 11.4ms to 13ms frame-to-frame interval. This performance is overall exceptionally fluid and largely thanks to Sniper Elite’s developers building the game so well. We want to avoid excursions from the mean greater than 8-12ms, as this becomes noticeable as a stutter to the user. The Radeon VII card does well in frametime pacing for this game.
As for the RTX 2080 FE, overall, performance is very similar to what we saw with Radeon VII. Frametimes are about 11.3ms to 12.3ms, on average, marking it as functionally identical in observable performance. The 1080 Ti will make this a messy line plot, but does show how well the 2017 card performs compared to its intended replacements of the modern era. Finally, for reference to a lower performing device, we’ll plot the RX 590 frametime performance at about 19ms to 27ms. Still acceptable with regard to frame-to-frame performance, but with much lower averages.
That’ll wrap Sniper Elite 4. 1080p bottlenecked on the CPU and failed to show differences between the high-end cards and a 2060, so it’s time to move to the next game.
We now turn to F1 2018 for an opposing look at a more abstracted API, standing as a modern representation of DirectX 11 as a lead-in to our more geometrically complex titles for later.
At 4K, the AMD Radeon VII GPU operates at 73FPS AVG, with lows at 53FPS 1% and 30FPS 0.1% low which, considering this game’s overall average low performance, isn’t all that bad. We’ll have to look at frametimes comparatively to get a better idea. As for averages, the Radeon VII card leads the overclocked 2070’s 68FPS by 7% FPS AVG, while being led by the 2080 FE and 1080 Ti SC2 at 81FPS AVG. The 2080 Ti, significantly higher in price, sets the ceiling at about 105FPS AVG.
Moving on to a frametime plot at 4K, we can take a closer look at whether those 1% differences were a fluke or repeatable improvements. Plotting Radeon VII first, we see the usual single-frame spike that we encounter in this benchmark. The spike jumps to 60ms before dropping to about 8ms, which is noticeable as a jarring stutter in gameplay. This only happens once throughout the entirety of the test, fortunately, with the rest of the frames being within the range of roughly 12.1ms to 14.2ms. This is objectively good from a frame pacing standpoint. Although that one spike is at risk of being averaged out, its extremity still drags down the low values noticeably.
Plotting the RTX 2080 FE, the line starts overall better than Radeon VII, albeit by differences that are mostly measurably, not necessarily perceptible. The 2080 hits its own spike to 83ms, a bit worse off than the Radeon VII, but this isn’t all that interesting. What’s interesting is the repeated frame spikes up and down at other parts of the benchmark. That’s the cause of those reduced low values. When we plotted the other 4 passes of each device, we saw almost identical behavior in each run. AMD’s Radeon VII peaked at 60ms, 65ms, 63ms, and 64ms, with this rendition of the 2080 being its own best. Finally, the RTX 2070 OC encounters similar spikes to the RTX 2080 FE. To be fair, this card is overclocked where the others aren’t, but its proximity to 2080 and Radeon VII performance made it an interesting line to plot. The running average isn’t all that different, but the frame pacing is markedly worse at irregular intervals.
1440p framerates will give us a fuller perspective of performance. With this resolution, the AMD Radeon VII card ends up at 120FPS AVG, allowing the 1080 Ti SC2 and its 132FPS AVG a lead of about 9.5%, although Radeon VII again plots higher 1% low and 0.1% low performance than its flanking NVIDIA counterparts. For other points of reference, our power modded Vega 56 card ends up at 104FPS AVG, maintaining similarly strong low performance metrics. Radeon VII leads this modded Vega 56 card by 15.5%.
1080p testing is more for plotting scalability than anything else. At 1080p in F1 2018, Radeon VII ends up at 153FPS AVG – again with higher lows – and affords the 1080 Ti an AVG FPS lead of 8.7%. Interestingly, at 1440p, the 1080 Ti had a lead of 9.6%, which then expanded to 10.8% at 4K. AMD’s card appears to lose some ground at the higher resolutions, purely from the perspective of relative average framerate, although it’s not that much of a change overall. An interesting item of note, but not particularly relevant to this title. Either way, again, we see similarly strong low performance in our modded Vega 56 card, approaching Radeon VII’s stock performance.
Far Cry 5
Far Cry 5 is our next game, running on the Dunia engine and representing a geometrically complex game with heavy use of screen-space reflections and long view distances. At 4K, the Radeon VII card ends up at 60FPS AVG, ranking it marginally ahead of the RTX 2080’s imperceptibly different 59FPS AVG and ahead of the 57FPS on the 1080 Ti SC2 stock card. This is the best relative performance we’ve seen thus far for Radeon VII, still planting it at about the same spot as the 1080 Ti and 2080. At this point, the argument would boil-down to cost, availability, power consumption, and noise. We’ll look at those soon.
At 1440p, the Radeon VII card ends up about tied in AVG FPS with the 1080 Ti SC2 and 2080 FE. Lows are within error of the 2080 FE. There is no meaningful difference in performance between these three devices when left to stock performance.
1080p is where we see our earlier observation break: Where we saw NVIDIA resolution scaling at 4K pull ahead of Radeon in F1, we see almost the opposite in Far Cry 5. The 2080 FE and 1080 Ti SC2 are functionally tied with one another, leading the Radeon VII card by about 6%.
Shadow of the Tomb Raider
Shadow of the Tomb Raider is the last game we’ll be showing in this video to leave time budget for other, more interesting tests. At 4K, the AMD Radeon VII card runs at 51FPS AVG, leading an overclocked 2070 by 12.2%, being led by the 1080 Ti SC2 by about 5%, and led by the 2080 FE by about the same. The Radeon VII keeps up, but does not outpace either of the primary two competitors.
At 1440p, we see the Radeon VII frametime advantage of previous games defray a bit, with the average framerate ending up at 87FPS, led by the 93FPS AVG of the 1080 Ti SC2. The 2080 FE holds a lead of 9.4% over the Radeon VII card.
Finally, at 1080p, we see the RTX 2080 and 1080 Ti cards pull further ahead of the Radeon VII, establishing a 10% lead. Radeon VII runs closer to an overclocked 1080 FTW when overclocked, again seeing its low performance advantage eroded by Shadow of the Tomb Raider.
AMD claims high compute performance for Radeon and used this in some of its marketing materials. To validate these claims, we ran three passes of Luxmark, an OpenCL benchmarking utility, against only the RTX 2080. These charts aren’t that populated because we don’t typically use this software.
Note that memory isn’t too heavily impacted in these tests, so AMD leads for architectural reasons, not necessarily capacity reasons. In this benchmark AMD does hold a firm advantage in the quicker “Ball” render test, commanding a lead of 72% over the RTX 2080 FE card. The microphone test positions Radeon with a lead of 50%, with the much more geometrically complex hotel scene allowing NVIDIA to claw away some of those gains, defraying the Radeon advantage to a still-respectable 15%.
At least in Luxmark, AMD does outperform NVIDIA for OpenCL testing. Of course, for software that is better accelerated with CUDA – or exclusively accelerated with CUDA – that leaves NVIDIA as the de facto option at this $700-$800 price-point, whereas AMD would be the better choice for this type of OpenCL compute workload.
Back to thermals for one chart before talking noise, the complete stock test of Radeon VII, including the pad and with no changes to software, puts the junction temperature at about 106-107 degrees as its target, with GPU temperature holding at about 80 degrees Celsius. For VRM thermals, we measured two VRM components and found that the right-half of the VRM is hotter, running at around 68 degrees Celsius, with the left-half at 60 degrees. These are completely respectable VRM thermal numbers for the MOSFETs and are well within acceptable ranges. AMD has done well on VRM component cooling, despite troubles in the core. Concerning the fan speed, this full auto, stock operation puts the fan at about 2900RPM in order to maintain its targets, but we need to know what that means for noise.
For noise levels, 2900RPM puts the card at about 50dBA, which is one of the louder operating noise levels we’ve recently encountered. The RTX 2080 FE’s average operating RPM is about 1850, which puts it at 39dBA. This is significantly quieter than the Radeon VII card; on this scale, the 10dBA is perceivable to a human as a rough doubling of noise. Remember, this is the operating RPM in our thermal test bench, so NVIDIA is heavily advantaged here. Overall, the range is below our 26dBA noise floor at idle, so we can’t accurately measure it, then about 32.5dBA at 1300RPM, 39.2dBA at 1850RPM, 53.9dBA at 3300RPM, and 59.4dBA at 3800RPM, but the default fan curve should never reach this high.
Comparatively, this is not dissimilar from the RPM-to-noise response curve we saw on the 2080 FE or Titan RTX. It’s just the operating fan speed that matters most, at that point, and NVIDIA does tend to operate closer to 1850RPM rather than 2900RPM; despite the same RPM-to-response readings, the actual operating noise levels favor NVIDIA.
Power consumption over time puts the AMD Radeon VII as peaking at about 420W total system draw during our Ashes of the Singularity power draw benchmark. Remember that this is total system draw, not individual component draw; the system is heavily controlled so that we can compare deltas, but it is not an individualized number. Average load power consumption is closer to 370-380W. For reference, the RTX 2080 FE card follows a similar pattern. It peaks at about 415W and averages in the 360W range for total system power consumption. We noted earlier that a power modded Vega 56 can get within 15% of the Radeon VII for much cheaper, but here’s the cost of that – our total system power consumption peaks over 600W with those mods, so it’s not exactly a fair comparison, and there’s no telling what those mods do to the life of the silicon.
Overclocking & Driver Bugs & Conclusion
Because AMD completely overhauled its API calls for this card, no current software utilities work for it. Afterburner is broken, GPU-z needs an update (and its creator is on vacation), and Wattool is also largely non-functioning. This leaves us with AMD’s WattMan, which is also presently in a largely unusable state.
Aside from innumerable other bugs encountered, some of which we’ll list below, the most noteworthy was that manual overclocking yields worse performance than running stock in at least 9/10 cases. That one time it doesn’t is typically within variance. All “overclocks” must be validated with performance testing, as misreporting of the clocks will lead users to believe the OC is actually working. Here are two tables illustrating what’s really happening:
|NOTE: ALL BELOW RESULTS ARE INVALID REPORTING.You have to validate overclocks with performance. These "overclocks" would apply and be "stable,"
but running a simple TimeSpy test proved that performance was significantly degrading, not improving.
WattMan misreports frequencies and the "overclock" ends up worse than stock.
|AMD Radeon VII Overclock Stepping | TimeSpy Extreme Stress Test | GamersNexus.net|
|Peak Frequency||AVG Frequency||Core Offset||Core Voltage||Memory Frequency||Memory Offset||Power||Fan Speed||Active Temp||Pass/Fail|
|Frequency Dips Regularly|
|NONE OF THIS IS WORKING. Performance is NOT scaling with inaccurately reported clock speed.|
|Still no performance scaling. Invalid.|
|TimeSpy Extreme OC Validation with Score|
|Reported Frequency||Reported Memory||Voltage Settings||TimeSpy Extreme GFX Score||TSE FPS1||TSE FPS2|
|1961||1100||1200mv||4431 (only succeeded once)||31.7||23.56|
|Auto Overclock||N/A||N/A||Black screen / shutdown||fail||fail|
|Auto Overclock||N/A||N/A||Black screen / shutdown||fail||fail|
|1790||auto||1100||OS freeze / shutdown||fail||fail|
|1790||auto||1090||OS freeze / shutdown||fail||fail|
We spoke with Roman (Der8auer) about this also, and he also experienced issues with manual overclocking even when using dry ice. One of his best OC results was with auto OC and -25C core temperatures, illustrating that the broken manual overclocking is at least partially a result of driver issues, but check his channel for more on that.
AMD’s drivers have largely improved over the past months, which is perhaps why it’s so disappointing that the Radeon VII drivers are so riddled with bugs. The company has worked hard to eradicate this perception of bad drivers, and has done well to fix its image and its driver packages, but botched the entire thing in one go with Radeon VII. Here’s a small list of what we encountered – we didn’t write all of them down:
- Occasional black screen & restart issues (full stock, no OC applied). Suspected related to ASUS motherboards
- Black screen / lock that requires hard shutdown (full stock, no OC applied)
- Stock/auto/out-of-box crash event during benchmark triggered hard reset, ultimately killing the ability to open Radeon Settings on the system. DDU and AMD’s clean uninstaller did not remedy the issue. “Driver gremlins” left behind, post-crash, completely broke AMD drivers. We re-imaged the system to bypass the problem.
- Some games hard crash, like Ghost Recon: Wildlands
- Some crashes cause fans to lock to 100% fan speed until power button is held/system is cold booted
- Manual overclocking seems to not do anything
- Power offset sometimes does not work (validated with power meters and clamps)
- Cannot adjust fan speed to 90%, but all other ranges work fine?
- Fan speed sometimes gets stuck at 100% and cannot be lowered, could not determine root cause
- Clock occasionally misreports, e.g. as “7800MHz”
- Crashes during OC stability testing can sometimes wipe-out drivers and require a clean reinstall as Radeon Settings will stop opening
- Performance monitor sometimes does not log for more than a few seconds on some installs (but works on others – root cause not found)
- Stats read-out in Wattman sometimes completely disappears, seemingly without reason (even under stock/unchanged settings)
- Fan options sometimes revert to old version (min/max RPMs rather than fan curve), seemingly without reason
We will revisit overclocking as soon as it works. We will also be posting some targeted benchmarks following the Radeon VII main review; unfortunately, as a result of losing at least one entire day to AMD driver issues, we pushed some of our feature tests out into other content pieces.
It is unfortunate that AMD has torpedoed its launch with drivers that aren’t ready, particularly coming off of relatively strong driver improvements in its recent past. The product in general needs more time. This launch was rushed – like most recent launches (see: initial RTX lineup) – and it really could have been a lot cleaner. Radeon VII seems to have more OC room than AMD is letting on, but bugs are holding it back.
Ignoring these issues, though, the card did “work” when we got it into the right system and as long as we didn’t want to overclock or launch specific games. In those instances, its gaming performance equates the 1080 Ti or RTX 2080 FE. OpenCL performance flies past NVIDIA on the Radeon VII card, offering a saving grace, though users of CUDA obviously see no benefit from this. If you specifically work with OpenCL and have a budget of about $700-$800, the Radeon VII card is a far better choice than NVIDIA. If working with CUDA, that sort of sets a requirement to use NVIDIA.
Noise levels end up about 10dBA higher than the 2080 FE at auto/stock settings, establishing a clear lead for NVIDIA’s reference card. This can be sort of swept aside once partner models come into play, but that’s a separate story. The PCB and VRM of the Radeon VII card are of excellent build quality, for what that’s worth, making the reference unit reasonable for waterblocks (once overclocking bugs are fixed, anyway).
We aren’t really weighing RTX features here, seeing as we presently are of the opinion that they provide minimal value to the consumer. There is value to animators or 3D artists – at least, there will be more once Blender adds official support – but gaming has few improvements. The conversation has seen an interesting shift in the past few weeks: Rather than “it has RTX, but no one cares” (regarding GeForce), we are now seeing “it doesn’t have RTX, but NVIDIA does” (regarding Radeon). For us, the lack or presence of RTX has no bearing on the value of the product. It simply is not a heavy enough influencer.
What does matter is noise levels (where NVIDIA leads with strength), gaming performance (where NVIDIA’s two-year-old 1080 Ti ties the Radeon VII), functional overclocking (NVIDIA shockingly holds a lead at present), and production performance (some testing still TBD, with a large OpenCL lead for Radeon).
We have some follow-up targeted feature testing for Radeon VII that will get separate content pieces; unfortunately, because of the time lost to driver defects, we had to push some testing back for a separate content item. For now, though, from a gaming and enthusiast standpoint, the Radeon VII card is difficult to recommend. At price equivalence, at best, you get rough equivalence in frame throughput, a good PCB and VRM, and maybe good overclocking features at some point, depending. That has been our primary reason to recommend Vega 56 lately – its overclocking is genuinely fun for enthusiasts, something that NVIDIA has shied away from and nearly altogether dropped. With Radeon VII losing all of that, it is harder to justify. Our primary hope would be that driver updates resolve much of this, but we’ll have to check back for that. We do not review based on promises, just like we didn’t for RTX.
Options are good and we want to see AMD succeed in more of its GPU pushes. Competition fuels inspiration – something we know first-hand from having to compete in the media space – and we’d like to see a stronger volley back-and-forth. At present, the product simply isn’t ready for launch. It needs another few weeks in the incubator, at which point we’ll revisit its viability as we expand testing to more production applications.
Editorial, Testing: Steve Burke
Video: Andrew Coleman