Posted: By Alex Herrera 12.06.19
As is our practice when evaluating workstation-class graphics AIBs, we ran SPEC's Viewperf, in this case the newest version 13. Viewperf focuses workload on the graphics AIB, such that the rest of the system isn't (or at least shouldn't often be) the bottleneck. As a result, Viewperf will give a good idea of which AIB has the highest peak performance. However, it's worth noting that the magnitude of any superior numbers does not indicate the level of superiority it will have in a real-world environment where the rest of the system, OS, and application may impose other bottlenecks.
With this benchmarking exercise, we ran tests on all three AIBs housed in the same workstation, an overclocked Boxx workstation running Windows 10 Professional with very high single-thread performance (important as Viewperf computation is single-threaded and therefore will apply heavier stress to GPU with a CPU that’s especially fast at single-thread execution). All three were run on the same day, downloading the most current, publicly available driver.
|Quadro RTX 4000||Radeon Pro W5700||Radeon Pro WX 8200|
|Estimated street price||$900||$800||$990|
|Peak FLOPS (SP)||7.1||8.9||9.8|
|Memory bandwidth (peak)||416 Gb/s||448 GB/s||512 GB/s|
|TDP||125 W||205 W||230 W|
|Key specifications for the Radeon Pro W5700, the preceding Radeon Pro WX 8200, and Nvidia’s Quadro RTX 4000. (Source: Nvidia)|
|Key hardware resource metrics: The Radeon Pro W5700 relative to its predecessor (the W5700) and its most natural competitor, Nvidia’s Quadro RTX 4000 (all normalized to the Nvidia Quadro RTX 4000)|
The results: Performance, performance/dollar, and performance/TDP
With respect to raw performance, the W5700 essentially matched the Quadro RTX 4000, on average about 2% faster across all viewsets, while besting its predecessor by around 14%. One particularly notable observation to make on the Navi-based W5700 is how performance has gone up relative to the WX 8200, despite lower maximum theoretical FLOPS. This behavior clearly illustrates how a more efficient architecture can better utilize a processor’s available execution slots. AMD claims a 25% higher IPC (instructions per cycle) for Navi compared to the previous GCN architecture.
With its lower price point, the W5700’s price-performance looked notably better, averaging around 15% better than the RTX 4000 in scores per dollar and 41% better than the WX 8200 (a caveat of course here that street prices do fluctuate). That extra edge over the WX 8200 comes courtesy of both better architectural and 7-nm implementation efficiencies.
|Viewperf 13 results: The Radeon Pro W5700 relative to its predecessor (the W5700) and its most natural competitor, Nvidia’s Quadro RTX 4000 (all normalized to the Nvidia Quadro RTX 4000)|
|Viewperf 13 performance/dollar: The Radeon Pro W5700 relative to its predecessor (the W5700) and its most natural competitor, Nvidia’s Quadro RTX 4000 (all normalized to the Nvidia Quadro RTX 4000)|
Comparing performance per watt takes a bit more discussion. The WX 8200 had looked relatively poor compared to the RTX 4000 when calculating performance relative to TDP, thermal design power. Bear in mind, TDP does not indicate how many watts are being consumed at a specific level of performance, rather it indicates a worst-case extended period of power consumption that demands adequate dissipation of the resulting heat produced. Furthermore, in the case of both the W5700 and Quadro RTX 4000, the AIB’s power consumption will be increased if supplying significant power over the USB-C port.
In this benchmarking, I did not measure watts consumed by the AIB as A, that’s not easy to do, and B, it’s arguably less important for a deskside system (emphasis on arguably). Yes, the TDP won’t indicate exactly how many amps are consumed per specific workloads and how that will affect the electricity bill. AMD emphasizes that in its measurements at idle and a couple of company-selected workloads, the W5700 consumes less (about 18%) than the RTX 4000.
Now, in real-world usage, the RTX 4000 and W5700 probably often have comparable power consumption, for example, in solid modeling (one of the two workloads) where the GPU is not likely to be under consistent maximum load for an extended period. So it’s probably fair to assume both AIBs will have similar impact on electricity usage, and perhaps the W5700 would draw fewer amps on average across a typical CAD workflow, as AMD pitches.
|Viewperf 13 performance/TDP: The Radeon Pro W5700 relative to its predecessor (the W5700) and its most natural competitor, Nvidia’s Quadro RTX 4000 (all normalized to the Nvidia Quadro RTX 4000)|
With the Radeon Pro W5700, AMD has successfully held serve with compelling Nvidia-matching 3D graphics performance at the price point … but it’s unlikely to change the status quo
If you’re a default buyer of AMD Radeon Pro products, you should be quite satisfied with the W5700, as you have your most competitive option in a long time in the $750–$1,000 range. With the W5700, AMD has a workstation-caliber graphics AIB that holds its own on 3D graphics performance versus Nvidia’s most natural competitor and given today’s pricing (a disclaimer since both AIBs’ street prices are fluid), roughly 15% better price-performance. So if you tend to default to AMD products, shopping in that price range, and have a system that can accommodate a dual-slot AIB and both auxiliary power connectors (most, but not all), then you’ve got an AIB available you should be quite happy with.
If you’re a default Nvidia buyer, however, you’re not likely to see enough in the W5700 to entice you to jump ship. Yes, the W5700 is a nudge faster on professionally-oriented 3D graphics viewsets, albeit a very small margin, especially when considering the viewsets are designed to be graphics-bottlenecked. So for tasks that are compute or data throttled, you’ll likely see less to no difference. Furthermore, the RTX 4000 has been out for close to three quarters, so you’ll likely see a future Quadro leapfrog both the RTX 4000 and W5700 before you see another Radeon Pro do so. And finally, a buyer of the W5700 needs to consider a few potential issues they may or may not care about, depending on the workflow, the size of the AIB, the workstation that houses the GPU, and how long you expect to hold onto both.
First off, consider the size (as a function of TDP primarily) and power requirements of the AIB. While Nvidia was able to house the 125 W (TDP) RTX 4000 in a single-width AIB requiring an 8-pin auxiliary power connector, the 230 W WX 8200 comes in a dual-width package requiring both 6-pin and 8-pin auxiliary power connectors.
|The single-width, 8-pin powered Quadro RTX 4000 and the double-width, 6 + 8-pin powered Radeon Pro W5700 (center) and WX 8200 (R)|
Finally, while ray-traced rendering is still in niche usage across mainstream applications, it’s increasing in adoption quickly, a trend spurred both by the technology’s obvious appeal of photorealism and the fact that Nvidia’s Quadro RTX line is for the first time integrating ray-trace acceleration (via both RT and Tensor Cores). So if your workflow today would benefit from accelerated rendering—or if you think it will in the near future—then the RTX 4000 offers appeal the Radeon Pro W5700 can’t match. While it will take some time to ramp the software infrastructure to support ray tracing pervasively, support is ramping, for example, with Dassault and Solidworks, which has integrated RTX technology in its Visualize in-app renderer.
Add that all up, and I’d say AMD has held serve with the Radeon Pro W5700. It’s got an AIB that provides substantially better performance at a lower price point than its previous WX 8200, and for now, it can match a modestly higher-priced rival AIB in conventional 3D graphics performance. It isn’t, however, likely to change the status quo in unit share at that price tier, as default Nvidia buyers will likely stick with the RTX 4000—at least for now—and wait to see what is coming down the road in the next few quarters.