The AMD Radeon VII isn’t a GPU we ever expected to see as a consumer product. AMD repeatedly indicated it had no particular plan to bring a 7nm to this space, preferring to keep its first run of 7nm hardware reserved for servers and the AI/ML space. Then, at CES 2019, AMD CEO Lisa Su announced AMD would bring a new GPU to market to compete head to head with the RTX 2080 at the $ 700 price point. Fast forward to the present day and here we are.
The Radeon VII is based on the same 7nm silicon as the Radeon MI50 and MI60. AMD shrank the die significantly between the two GPUs and used the space savings to squeeze in another set of HBM2 chips, doubling available RAM bandwidth. If you care about sheer memory capacity in your graphics card — and to be clear, there are professional and scientific applications that benefit from large GPU buffers — the Radeon VII is the only card on the market that gives you this much RAM for under a thousand dollars.
Boosted FP64 Performance
The Radeon VII has another advantage over other cards on the market, though AMD kept this one tight to their vest. Until today, the highest FP64 performance you could buy in a consumer GCN GPU was the Radeon 7990, a short-lived dual-GPU product that’s nearly five years old. Initially, AMD communicated that its Radeon VII GPU was capable of just 0.88 TFLOPS (1/16th of FP32 performance). Instead, the Radeon VII is capable of a whopping 3.46 TFLOPs, or just under half of the MI60’s maximum performance.
This significant repositioning puts a new spin on the Radeon VII, positioning it more clearly for double-precision compute workloads. Unfortunately, this news wasn’t communicated until right before the review NDA, which means we won’t have time to take a particular examination of the GPU in that context. We intend to write a separate article investigating compute and scientific workloads on the Radeon VII, which will give us an opportunity to examine this side of the GPU more thoroughly. Today’s review will focus on the consumer side of the equation.
As with the first 7nm GPU, all eyes are going to be on the Radeon VII to see how well it performs on the process node. We already have some information on this point. As the table above indicates, the Radeon VII die is ~67 percent the size of the Vega 64 die. Clock rates have improved only modestly, with the base clock up by 1.09x and the boost clock increasing by 1.13x.
Most of the improvements between Radeon Vega and Radeon VII boil down to the huge increase in RAM bandwidth and total VRAM buffer. The clock gains from the 7nm shift are very modest. AMD did note to us that it had reduced the GPUs internal latency and made a few performance-enhancing tweaks to the architecture, but data here was limited. Those hoping that AMD would have a new major AI or machine learning initiative to announce — something akin to DirectX ray tracing support or new antialiasing methods — will be disappointed.
New Thermal Monitoring
The Radeon VII uses a new type of thermal monitoring system rather than the old, edge-mounted GPU thermistors it had previously deployed. Radeon VII has a total of 64 temperature sensors mounted across the die, 2x the number of Vega 64. Going forward, Radeon VII GPUs will use the maximum temperature measured across the die, known as the Junction temp, to control GPU behavior. According to AMD: “Controlling based on Junction Temperature from the extensive sensing network allows each GPU to reliably maximize its performance potential while reporting an additional temperature that is more representative of the hottest parts of the GPU.”
AMD notes that Junction temps will be higher than what gamers are used to seeing in the past. This shouldn’t be considered a problem, and the GPUs are designed to hold these temperatures safely. Junction temps of 110C are not unusual or considered problematic.
The Story So Far…
Before we dive into performance figures, let’s revisit how the GPU market has evolved in the past six months. Nvidia’s RTX refresh cycle this past fall didn’t do much to improve performance-per-dollar. Of its new high-end GPUs, only the RTX 2080 Ti genuinely moved the ball forward on performance. The RTX 2080 and RTX 2070 are slightly faster than the GTX 1080 Ti and GTX 1080 respectively (think 8-12 percent), but carry higher prices than their predecessor GPUs did. Nvidia’s justification for these price increases has been to point at its new ray tracing feature as justification. ExtremeTech historically takes a very dim view of buying hardware for features you can’t use, for reasons we explored in-depth as part of our RTX 2080 and 2080 Ti review.
But therein lies the rub. Even if you agree that Nvidia’s RTX technology is a risky bet, AMD hasn’t baked anything equivalent into the Radeon VII. In the past, one GPU vendor sometimes zigs when the other zags — Nvidia went all-in for 3-D glasses and monitors a few years back, while AMD threw its weight behind Eyefinity and multi-monitor gaming. In this case, RTX and DLSS were major launch features for Nvidia, while AMD’s major launch feature is the straightforward promise of more performance relative to its previous generation of graphics cards.
Gamers hoping that AMD would bring a GPU to market that shook up the status quo or at least forced Nvidia to lower its prices are going to be disappointed (again). AMD is targeting the RTX 2080’s performance at the RTX 2080’s $ 700 price point using Vega’s 300W power envelope.
All testing was done on an Asus Prime Z370-A using an Intel Core i7-8086K with 32GB of DDR4-3200 and the latest version of Windows 10. A Thermaltake Toughpower 80 Plus Titanium 1250W PSU was used for testing.
All of our results can be parsed in the slideshow below. Click on each graph to open it in a new window.
AMD promised that the Radeon VII would be capable of tackling the RTX 2080, and that’s generally what we see here. In our complete suite of tests, the Radeon VII offered 96 percent the performance of the RTX 2080. That’s within a 5 percent margin of error, and close enough to declare that yes, the two solutions are generally competitive.
The Radeon VII tends to exhibit superior scaling to the RTX 2080, which is to say, it tends to lose less performance than its competitor as resolution increases.
Power Consumption, Heat, and Noise
Note: My GPU test rig uses a V3 Voltair CPU cooler, which includes a Thermo-Electric Cooler, or TEC. This consumes additional electricity. As a result, our idle and power consumption figures may be higher than elsewhere on the web. Running 32GB of RAM in XMP at DDR4-3200 also substantially increases power consumption compared with stock voltage and DDR4-2400.
Our power consumption figures are taken from the third loop of a Metro Last Light Redux benchmark run at 1920×1080. All detail settings conform to those used for our GPU reviews.
AMD’s absolute power consumption has scarcely budged compared with Vega 64, but its performance per watt has improved significantly. Using the automatic undervolting option on our Radeon VII decreased power consumption by ~7.5 percent without harming frame rates at all.
Once you factor in the Radeon VII’s increased performance, the GPU is indeed significantly more efficient. The Radeon VII consumes roughly 75 percent as much power as the Vega 64 per frame of animation drawn. Activate its underclocking feature, and this drops to 70 percent. But the RTX 2080 consumes just 63 percent the power of the Radeon Vega 64. AMD’s 7nm GPU draws roughly the same amount of power as its Nvidia rival, but it isn’t quite as efficient on the whole.
Finally, there’s noise. I don’t own a dB meter, but folks — Radeon VII ain’t quiet. Overall, it’s comparable to the Vega 64, but there are moments when the fans on the Radeon VII kick harder. They also tend to ramp up faster. Everyone has their own personal tolerance for this sort of thing, but I consider the noise profile of these cards to be a significant negative.
At this point, the noise situation has become ridiculous. Ever since at least Hawaii, reviewers have hit AMD for the noise profile of its reference designs. To its credit, the company has at least attempted to address this, but its most high-profile attempt to fix the problem with a water cooler created an even bigger mess. Vega 64 was a loud GPU, louder than I’m personally comfortable installing in my own system. Radeon VII doesn’t improve on this at all. At this point, it’s clear AMD doesn’t actually have any interest in building or outfitting its reference cards with coolers that match the performance of what Nvidia ships (and what Nvidia ships isn’t always great, either, mind you). Hawaii launched over five years ago. Why are we still waiting for AMD to actually fix this in a high-end GPU that isn’t the Radeon Nano?
I expect a $ 700 GPU to have a better noise profile than the $ 12 box fan I bought at Aldi.
The Radeon VII is a rough match for the RTX 2080’s performance. It also lacks one of the RTX 2080’s specific liabilities. When Nvidia launched the RTX family, it asked customers to swallow a significant price hike relative to the company’s then-current GPU lineup. AMD isn’t making that mistake. The Radeon VII offers more performance than any AMD GPU before it. It’s 1.33x faster than the Vega 64 with no other pesky last-gen cards to muck up the stack.
Of course, given that the Vega 64 currently sells for $ 400, you’re buying 1.33x more performance for 1.75x the money. That’s exactly the opposite of the kind of ratio we prefer to see. It’s hard not to wonder what 7nm Vega’s performance might have looked like if the GPU had fielded more ROPs and texture mapping units to accompany its enormous bandwidth increase — again, we suspect that consumer games simply weren’t the primary market for this card and that its overall design reflects that fact. We intend to write a separate article focusing on GPU compute and will report back with how the Radeon VII’s additional memory bandwidth aids it in those contexts.
For me, the bottom line is this: I have consistently argued that Nvidia erred in raising prices when it launched the RTX family. I have argued that there were too many reasons to believe RTX features won’t be particularly useful during Turing’s lifespan to justify buying into the family, particularly when older GTX cards based on Pascal offered such competitive performance.
Now we have the Radeon VII. It lacks RTX/DXR features but doesn’t offer an alternative. Its larger HBM2 buffer and interposer costs may explain its pricing — I’ve wondered before if Nvidia’s decision to raise prices would actually make it make sense for AMD to bring 7nm Vega to the consumer market — but they don’t make it a particularly good value. How much value do I put on ray tracing right now? Not much. But a feature that’s only useful in a tiny number of games is still at least arguably more useful than no feature at all. Nor is there much reason to believe that the 16GB HBM2 buffer will get a workout any time soon. Game developers will always target their titles to the GPUs that people practically own; having one consumer GPU with 16GB of onboard memory isn’t going to move the needle in terms of future game VRAM requirements.
If AMD had managed to bring Radeon VII in at $ 600, it would have a genuine value argument to make relative to Nvidia’s product stack. If it had outpaced the RTX 2080 at the same price, it could argue for superior rasterization performance with higher noise levels as an acceptable trade-off. Instead, what we have here — at least in the consumer market — is a loud RTX 2080-equivalent without the admittedly dubious features Nvidia tried to use to justify its price increases.
You know what’s worse than an RTX 2080 with dubious features and a bad price point? A loud RTX 2080-equivalent with no new features at all and the same bad price point.
The story isn’t all bad here. AMD was able to take advantage of its shift to 7nm to improve its overall competitive standing against the RTX family, and the Radeon VII competes against the GTX 1080 Ti / RTX 2080 more effectively than Vega 64 fared against the older GeForce 1080. But I can’t hammer Nvidia for months over the price increases and positioning it introduced with Turing only to turn around and laud AMD for delivering a GPU that roughly matches on perf but offers fewer features and higher noise, uncertain as the value of those features may be.
This is not the 7nm GPU you’re looking for. We’ll have more to say on compute specifically in the near future.