Date: February 7, 2019
Author(s): Rob Williams
AMD’s Radeon VII might be the first 7nm gaming GPU, but the reality is, this Vega 20 chip is keen on tackling compute-intensive benchmarks, such as those from our workstation test suite. Having taken a look at gaming performance in another article, this one takes care of encoding, viewport, and rendering applications.
When AMD announced its Radeon VII GPU at CES, the mention of a 16GB framebuffer sent an immediate chill down our workstation PC’s back. That 16GB framebuffer operating at 1TB/s made it obvious that Radeon VII is a compute beast, despite its gaming-focused marketing.
1TB/s of bandwidth on a gaming card in an age when 500GB/s is beyond rare is quite a feat. It’s the kind of spec you’d only expect to see from the enterprise, and wouldn’t you know it, the Radeon Instinct MI50 also has 1TB/s of memory bandwidth to go with its 3840 cores.
To prove that this card was made for workstations, AMD’s briefing for the card tackled creative workloads ahead of the gaming ones. Blackmagic Design’s DaVinci Resolve was name dropped first, followed by Techgage frequent fliers Premiere Pro, Blender, and LuxMark. In Premiere and Blender, AMD says the VII is slightly faster than an RTX 2080, but in LuxMark? It dominates.
It may not affect many creative users, but AMD decided ahead of launch to improve the VII’s FP64 performance. After CES, we established with AMD that the card wouldn’t have accelerated double-precision performance, which gave us a rough estimate of 862 GFLOPS. Soon after, AMD changed its mind, doubling the effective FP64 performance over RX Vega.
As we were compiling some of our results, we noticed better-than-expected FP64 performance. We expected 2x at best, but saw 3x and beyond. It’s only by coincidence we have this information, as it’s tackled in our benchmarking script as an “in case” sort of thing. 14 hours ahead of embargo lift, AMD shot us an email which clued us in.
Long story short: due to “customer interest”, AMD decided to increase Radeon VII’s FP64 performance to 3.52 TFLOPS, which is significantly better than any other gaming GPU out there. Two of these cards in parallel would deliver FP64 performance close to NVIDIA’s Tesla V100.
In the gaming performance look that complements this article, I mentioned that AMD didn’t seem to know where the Radeon VII should be targeted. It’s called a gaming card first and foremost, yet AMD doesn’t hesitate to jump into discussion of the card’s creative abilities ahead of the gaming ones. And now we have FP64 that’s 1:4 FP32 instead of 1:16, a design change that suddenly adds more people to the list of those who will want a VII – namely those involved with science or finance. Gamers don’t need FP64, and the vast majority of creative professionals don’t, either.
Nonetheless, I draw a bit of a comparison to NVIDIA’s TITAN series with AMD’s Radeon VII. The TITAN series is NVIDIA’s best for gaming, but focuses largely on compute, and especially deep-learning. Radeon VII, meanwhile, gives focus to both gaming and workstation equally, and at seemingly the last minute, high-precision workloads, as well.
Overall, the Radeon VII seems to be best targeted at those who want a powerful GPU for their creative workloads, but also one for their gaming. These professionals wouldn’t need optimizations that might exist only on Radeon Pro, and wouldn’t care about higher-tier customer service, or 10-bit color in DX applications. With 16GB of HBM2, Radeon VII’s memory certainly isn’t going to hold it back.
To better understand AMD’s current offerings both for gaming and workstation, here are some specs:
|AMD’s Radeon Gaming GPU Lineup|
|Cores||Base MHz||Peak FP32||Memory||Bandwidth||TDP||Price|
|Radeon VII||3840||1400||13.8 TFLOPS||16 GB 4||1 TB/s||300W||$699|
|Vega 64||4096||1546||12.6 TFLOPS||8 GB 4||484 GB/s||295W||$499|
|Vega 56||3584||1471||10.5 TFLOPS||8 GB 4||410 GB/s||210W||$449|
|RX 590||2304||1576||7.1 TFLOPS||8 GB 3||256 GB/s||225 W||$279|
|RX 580||2304||1340||6.1 TFLOPS||8 GB 3||256 GB/s||185W||$229|
|RX 570||2048||1244||5.1 TFLOPS||8 GB 3||224 GB/s||150W||$179|
|RX 560||896||1175||2.6 TFLOPS||4 GB 3||112 GB/s||80W||$119|
|RX 550||640||1183||1.2 TFLOPS||2 GB 3||112 GB/s||50W||$99|
|Notes||1 GDDR6; 2 GDDR5X; 3 GDDR5; 4 HBM2|
Architecture: Radeon RX 550~590 = Polaris; Radeon VII, RX Vega 56 & 64 = Vega
|AMD’s Radeon Pro Workstation GPU Lineup|
|Cores||Base MHz||Peak FP32||Memory||Bandwidth||TDP||Price|
|SSG 9||4096||1440||12.3 TFLOPS||16 GB 8||484 GB/s||260W||$6999|
|WX 9100||4096||1200||12.3 TFLOPS||16 GB 8||484 GB/s||230W||$1399|
|WX 8200||3584||1200||10.8 TFLOPS||8 GB 8||512 GB/s||230W||$999|
|Frontier||4096||1382||13.1 TFLOPS||16 GB 4||484 GB/s||300W||$499|
|Pro Duo||2304 x2||1243||5.7 TFLOPS||32 GB 3||448 GB/s||250W||$449|
|WX 7100||2304||1188||5.73 TFLOPS||8 GB 3||224 GB/s||130W||$549|
|WX 5100||1792||713||3.89 TFLOPS||8 GB 3||160 GB/s||75W||$359|
|WX 4100||1024||1125||2.46 TFLOPS||4 GB 3||96 GB/s||50W||$259|
|WX 3100||512||925||1.25 TFLOPS||4 GB 3||96 GB/s||50W||$169|
|WX 2100||512||925||1.25 TFLOPS||2 GB 3||56 GB/s||50W||$129|
|Notes||1 GDDR6; 2 GDDR5X; 3 GDDR5; 4 HBM2|
5 GDDR6 (ECC); 6 GDDR5X (ECC); 7 GDDR5 (ECC); 8 HBM2 (ECC)
9 Includes 2TB of solid-state storage on-card.
Architecture: WX 2100~7100 = Polaris; WX 8200, 9100 & SSG = Vega
AMD’s Frontier Edition was a monster when it first released, and it continued on as AMD’s fastest GPU until this Radeon VII came along. The VII is not only 700 GFLOPS faster, it also doubles the memory bandwidth, with the same amount of HBM2. The VII also takes after the FE with its 300W TDP, a value we don’t think tells the full story (as our numbers are better).
With most everything successfully brain-dumped to this first page, we’ll quickly tackle test stuffs below, and then jump right into Radeon VII workstation testing.
On the following pages, the results of our WS GPU test gauntlet will be seen. The tests chosen cover a wide-range of scenarios, from rendering to compute, and includes the use of both synthetic benchmarks and tests with real-world applications from the likes of Adobe and Autodesk.
Six graphics cards have been tested for this article. More would have been tested had time allowed it, but since it didn’t, we chose what we felt were the six most relevant GPUs to help paint a useful picture of workstation performance.
That GPUs include NVIDIA’s top GeForce, 2080 Ti, the new Quadro RTX 4000, priced $200 above the VII, last-gen’s TITAN Xp, which NVIDIA blesses with workstation enhancements, as well as last-gen’s top normal Quadro, the P6000. AMD’s Radeon Pro WX 8200, priced $100 higher than the VII, is also included.
Here are the specs of the test machine:
|Techgage Workstation Test System|
|Processor||Intel Core i9-7980XE (18-core; 2.6GHz)|
|Motherboard||ASUS ROG STRIX X299-E GAMING|
|Memory||HyperX FURY (4x16GB; DDR4-2666 16-18-18)|
|Graphics||AMD Radeon VII (16GB; Jan 22 Press Driver)|
AMD Radeon Pro WX 8200 (8GB; 18.Q4.1)
NVIDIA GeForce RTX 2080 Ti (11GB; 417.71)
NVIDIA TITAN Xp (12GB; 417.71)
NVIDIA Quadro RTX 4000 (8GB; 412.16)
NVIDIA Quadro P6000 (24GB; 412.16)
|Storage||Kingston KC1000 960GB M.2 SSD|
|Power Supply||Corsair 80 Plus Gold AX1200|
|Chassis||Corsair Carbide 600C Inverted Full-Tower|
|Cooling||NZXT Kraken X62 AIO Liquid Cooler|
|Et cetera||Windows 10 Pro build 17763 (1809)|
|For an in-depth pictorial look at this build, head here.|
Benchmark results are categorized and spread across the next four pages. On page 2, Adobe’s Premiere Pro and MAGIX’s Vegas Pro lead our encoding tests, with both AVC and HEVC codecs taken care of. On the same page, Sandra’s financial and scientific performance can be seen, as well as the cryptography.
On page 3, a few renderers are taken care of. These include the popular open-source design suite Blender, as well as LuxMark, and Radeon ProRender. ProRender is being tested with the help of 3ds Max, a suite that normally handles Redshift and V-Ray as well – but alas, neither work (or work well) on AMD graphics cards.
Page 4 is home to viewport performance, covered with the help of SPEC and its SPECviewperf suite. In total, 8 test results are featured here, covering important design suites like CATIA, SolidWorks, Siemens NX, Creo, as well as Autodesk’s 3ds Max and Maya.
For those interested in gaming performance, we have an article dedicated to that. So, don’t fret about missing 3DMark scores – those are tackled in that article.
Without further ado, let’s get this train moving.
With encoding that’s entirely CPU-bound, only the CPU matters, and as such, it’s easy to generate results that show great scaling. When the GPU is added into the mix, the CPU begins to matter a bit less, as the GPU can handle so much of the workload on its own. For 1080p encodes, the GPU is not going to be touched too much (in most cases), but if that 1080p is backed by a full project, such as one of our YouTube videos, the GPU can still help a lot.
Across these three different tests, there’s no winner to really speak of. The only result worth calling out may be the WX 8200, which fell a bit behind the others with our own project encode. Had these charts included lower-end GPUs, greater differences would be seen.
With MAGIX’s Vegas, we have a few interesting results to pore over. It might be important to get the odd WX 8200 results out-of-the-way: yes, those wound up with identical times between AVC and HEVC, even between two completely different projects.
In this application, the Radeon VII doesn’t shine like we hoped it would, and in fact, we never expected both AMD cards to sit at the bottom here. It feels like the GPU driver itself could help with this. Especially since AMD used to perform quite well in Vegas.
AMD has been strong in cryptography on both the CPU and GPU side in recent years, and with Radeon VII, that strong performance continues to shine through. Interestingly, strengths on either side trade places, but NVIDIA’s 2080 Ti is hard to beat – especially on the SHA2-512 side. With SHA-256, the VII sits ahead of every GPU aside from the 2080 Ti, and it continues to deliver strong performance when the security is increased.
With the financial test, we once again see the 2080 Ti strutting its stuff in a rather noticeable fashion. The Radeon VII performs extremely well on its own, though, as the 2080 Ti is the only GPU it wasn’t able to overtake here. It’s important to note that double-precision wasn’t used in this particular test, else we’d see even greater differences.
We get to wrap up this page with some great performance out of the VII. In both the GEMM and FFT Sandra tests, the VII places ahead of every other GPU. Again, this was using FP32.
Further tests will happen in the near-future. Time was the major issue here! More in-depth Sandra testing, including double-precision, can be found here.)
Both AMD’s CPUs and GPUs have been great for Blender performance the past couple of years. In particular, Radeon RX Vega 64 impressed us quite a bit at its launch, much like the Radeon VII is impressing us with the performance seen here. NVIDIA’s top dog pulls ahead, but it also costs at least $300 more, so all things considered, VII is a strong contender.
The release of Blender 2.80 is right around the corner, and we plan on digging into its performance once it drops. Unfortunately, our initial performance tests with the 2.80 beta have been a miss, although more testing needs to be done. In particular, heterogeneous rendering (CPU+GPU) is enabled by default, and in our initial tests, using that mode is slower than using the GPU by itself.
We’ll dig in a lot more soon, but if you’re reading this and have had great experiences with heterogeneous rendering in the 2.80 beta, please comment below to let us know.
During our product briefing, AMD mentioned that LuxMark performance on the Radeon VII is good, and it gave us an example of just how good. We didn’t really expect it to best the 2080 Ti, which already impressed the hell out of us with its performance a couple of months ago. Here, AMD didn’t just beat the 2080 Ti with the LuxBall render, it leaped ahead.
It’s humorous that when Ryzen launched, AMD’s best bud quickly became Cinebench. Now that Radeon is so hopped up on compute steroids, LuxMark suddenly has the company’s interest – and really, who could blame it? This performance is outright amazing.
What relevance does LuxMark have, you ask? It’s relevant to those who use the open-source renderer LuxRender, now known as LuxCoreRender due to project abandonment. If you use Blender, you can plug LuxCoreRender in, or use standalone software to go the fully manual route.
It’s strange, but AMD’s own Radeon ProRender performs better on NVIDIA GPUs than Radeons, something we found out in our look at the WX 8200. There’s something about NVIDIA’s RTX that accelerates this workload, based on both the 2080 Ti and Quadro RTX 4000 result. But, even the older Pascal GPUs perform better than AMD’s latest and greatest.
Based on the fact that the Quadro P6000 pushed ahead of the faster-clocked-but-equally-spec’d TITAN Xp, it seems like Quadro might have some optimizations in place that helps boost performance. At the same time, architectural enhancements brought to the 2080 Ti help it keep comfortably on top of this chart.
Unfortunately, the Radeon VII didn’t perform well here against the competition, even having been beat out by the Quadro RTX 4000. That card costs about $300 more than the VII, but is technically slower (it’s a pro card, thus it costs more).
SPEC’s 3ds Max viewport test is the only one of this bunch limited to 1080p resolution, and fortunately, we begin to see the VII exercising its strengths a bit better here with Maya. At 4K, the card comes ahead of the RTX 4000. At 1080p, the performance oddly suffers, a reality that probably won’t be evident based on the performance being extremely good at that point to begin with.
With the CATIA test, Radeon VII has its chance to shine, placing well ahead of the 2080 Ti. Somehow, RTX’s enhancements haven’t benefited this test as much as it has some others, much to the VII’s benefit. CATIA is one such example of software that would justify 16GB of GPU memory.
These results also highlight that CATIA is an area where NVIDIA chooses to deliver optimizations to the TITAN Xp (and possibly other TITANs). The Quadro P6000, despite being technically slower than the TITAN Xp and 2080 Ti, tops the chart. The RTX 2070-esque Quadro RTX 4000 likewise places well ahead of the RTX 2080 Ti.
This review is going live a week before SOLIDWORKS World 2019, held in Dallas, Texas, so it’s too bad we didn’t have some great news about VII performance to share. Fortunately, the VII isn’t alone in suffering crippled performance here. The 2080 Ti and TITAN Xp are also held back. Meanwhile the Radeon Pro WX 8200 and both Quadros occupy the top-half of the chart.
Siemens NX helps prove that not all workloads are built alike. The TITAN Xp, which didn’t fare as well as the Quadro P6000 in SolidWorks, had no problem delivering explosive performance here. It didn’t quite match the P6000, but the real-world benefits at that point are going to be harder to distinguish.
That’s not so much the case when we’re looking at the gaming cards. With those, you’re definitely going to notice a serious performance hit. Radeon Pro and Quadro are your go-to here.
Creo is another application that tends to run a lot better on certified workstation GPUs than the gaming equivalents. The WX 8200 performs admirably here, with the 2080 Ti proving that its gaming-focused self can handle Creo no problem, and in fact can deliver market-leading performance. It’s a bit odd that NVIDIA’s gaming GPU performed so well here when AMD’s Radeon VII didn’t, but it’s likely a matter of differing priorities (but to be clear, AMD loves Creo, so this performance is sure to only get better.)
We sometimes bundle Energy and Medical together, but Radeon VII makes things too interesting to do that this time around. In the Medical test, the VII outperforms all of the others, and if you’re suspecting that the 16GB of VRAM helped, you’re probably correct.
With this Energy test, 1080p and 4K performance deltas can vary to a much greater degree than most of the other tests here. At 1080p, the Quadro P6000 sits far ahead of the pack – a result that seems unrealistic, but matches our testing from the last time we put the card through its paces. At 4K though, the differences in performance between the GPUs is harder to distinguish – a good thing for users.
To test for power consumption, a Kill-A-Watt that the PC itself is plugged into is used for monitoring a Far Cry 5 4K benchmark run. Admittedly, we don’t have the best methods for power testing, and would love to improve them in the future, but for now, we use what we have, and that’s our eyeballs. Over the course of the one-minute benchmark, a rough average is pulled.
7nm can of course mean less power draw, and in Radeon VII’s case, it manages to draw ~50W less than the Vega 64, while easily outperforming it at the same time. Of course, RX Vega’s high power draw can be remedied with Wattman, but sadly, most people won’t know to use it, so it’s nice to see VII’s power draw to be a lot more modest out-of-the-gate. It could possibly be improved further if you want to spend time in Wattman, but up to this point, time hasn’t allowed us to dig into that.
In our look at AMD’s Radeon VII performance in gaming, the card was a little challenging to sum-up for the simple fact that it feels like more of a workstation card – even though it can handle both scenarios just fine (ignoring some driver issues we experienced, which will be fixed.)
On the pro side, the VII is easier to draw up some conclusions for. It helps that this is the fastest GPU AMD has ever made, so if there’s a scenario that will take advantage of OpenCL, this card should perform great, based on our experiences with such tests as Blender and LuxMark. Oddly, the OpenCL Radeon ProRender continues to give NVIDIA the performance lead, defying logic – but there could be many things attributed to it.
When time allows, it’ll be worth exploring ProRender in other design suites to see if the performance outlook is any different, but at the same time, ProRender also gave us slower-than-expected performance on the CPU side as well, especially with the 32-core Ryzen Threadripper. Either way, once it makes sense to test further, we’ll jump on it.
In high-end software solutions like CATIA, the Radeon VII performs well. It pays to be cognizant of each vendor’s gaming vs. workstation strengths, though, since in SolidWorks, you’d definitely want a pro card, while in CATIA, a gaming card will probably suffice (but in that case, the VII is the easy choice over the 2080 Ti.)
Not all of the performance seen throughout the article was that impressive, with Radeon falling behind in MAGIX’s Vegas, but that could possibly point to software updates being needed. With Adobe’s Premiere Pro performing just fine with the VII, there’s definitely room for some improvement with Vegas.
SiSoftware’s Sandra helped us establish that the Radeon VII is one hell of a compute machine, outperforming all of the other chips in the scientific test – even without the use of FP64. But on the topic of FP64, if you have workloads that require it, then the VII is going to outperform any other gaming GPU you’ll bring home. We may dig into this deeper at some point, but the scenarios typically used for double-precision are a bit outside of our wheelhouse.
While gaming wasn’t tackled here, it was in this complementary article. In it, we see that the VII keeps up to the RTX 2080 fairly easily overall, with two games in particular catering much more to NVIDIA than AMD. We also experienced some odd driver issues during our tests, but AMD has assured us that the first post-launch driver will patch things up nice.
And with that, you hopefully have a better idea of whether VII is suitable for you in compute workloads, but if not, leave a comment with what software you’re using, and we’ll try to help.
It might not be ideal in all situations, but AMD’s Radeon VII is a very well-rounded graphics card. It delivers great performance in gaming, potentially excellent performance in math-heavy workloads, and shines with strong OpenCL performance in applications like Blender and LuxMark. With 16GB of memory on tap, the card’s $699 price tag seems more than fair.
For those looking for a pure gaming card, the RTX 2080 may be the better choice. You’ll lose 8GB off of the memory by going with RTX 2080, but you’d gain things like Microsoft DXR and NVIDIA DLSS support, as well as smaller features like the screenshot tool Ansel. If you believe 8GB will not be enough for your gaming workloads over the next few years, then the allure of VII will only grow stronger.
As always, it pays to know what you’re looking for in your graphics card. If luck is on your side, the VII will complement your workloads just fine. The results scattered throughout this article can help you decide which way you should be swaying.
Games might not cater to those who want this card for compute workloads, but if you happen to be a gamer as well as a creative genius, you may be interested to know that a VII purchase will qualify you to receive three AAA titles that recently released, or will release soon. The three games are a $180 value, so even if you can’t use them, you may have family or friends who can. Ultimately, this bundle helps improve the value proposition here, and does it well.
Copyright © 2005-2019 Techgage Networks Inc. - All Rights Reserved.