AMD's RX 470 has been on our time table since May, when the pre-Computex press event informed us of a “mid-July” release. Well, it's mid-July – wait.
August 4th. It's August 4th. The RX 470 is available effective today, coinciding with embargo lift on reviews, and we've had time to thoroughly analyze the card's performance. The RX 470 is a partner card and will not be available as a reference model, though some partner cards may as well be reference models; they're using the reference RX 480 cooler, just with new colors, back-plates, or LEDs.
AMD has positioned its RX 470 in the sub-$200 market, listing its MSRP as $180. AIB partners will price their cards according to any custom coolers or pre-overclocks applied, though the floor has been set, more or less. That plants the 470 in a presently unchallenged market position: AMD's biggest current-gen competition in this price-range is its own RX 480 4GB card, the GTX 1060 being nVidia's lowest tier offering.
Before our deep-dive review on the Sapphire RX 470 Platinum, card architecture, thermal & endurance throttles, power, and FPS, let's run through the specs.
Current Gen GPU Reviews
AMD RX 460, RX 470, & RX 480 Specs
AMD RX 480 | AMD RX 470 | AMD RX 460 | |
Architecture | Polaris 10 | Polaris 10 | Polaris 11 |
Compute Units (CUs) | 36 | 32 | 14 |
Stream Processors | 2304 | 2048 | 864 |
Base / Boost Clock | 1120MHz / 1266MHz | 926MHz / 1206MHz | 1090MHz / 1200MHz |
COMPUTE Performance | >5 TFLOPS | Up to 4.9TFLOPs | Up to 2.2TFLOPs |
Graphics Command Processor (GCP) | 1 | 1 | 1 |
ACEs | 4 | 4 | 2 |
HWS | 2 | 2 | 2 |
TMUs | 144 | 128 | 56 |
ROPs | 32 | 32 | 16 |
Geometry Processors | 4 | 4 | 2 |
Peak Texture Filter Rate | 182.3GT/s | 154.4GT/s | 57.6GT/s |
Peak Pixel Filter Rate | 40.5GP/s | 38.6GP/s | 19.2GP/s |
L2 Cache | 2MB | ? | 1MB |
LSU | 576x32b | ? | 256x32b |
VRAM Capacity | 4GB GDDR5 @ 7Gbps 8GB GDDR5 @ 8Gbps |
4GB GDDR5 | 2GB GDDR5 |
Memory Interface | 256-bit | 256-bit | 128-bit |
Memory Speed | 7Gbps (4GB model) 8Gbps (8GB model) |
6.6Gbps | 7Gbps |
Memory Bandwidth | 224GB/s (4GB model) 256GB/s (8GB model) |
211GB/s | 112GB/s |
TDP | 150W | 120W* | 75W |
Power Connectors | 6-pin | 6-pin | None (Slot only) |
Display Port | 1.3 HBR / 1.4 HDR | 1.3/1.4 HDR | 1.3/1.4 HDR |
Release Date | June 29 | August 4 | August 8 |
Polaris 10 vs. Polaris 11 Specs & Architecture
Polaris 10 | Polaris 11 | |
Compute Units (CUs) | 36 | 16 |
COMPUTE Performance | “>5 TFLOPS” | “>2 TFLOPS” |
Memory Interface | 256-bit | 128-bit |
Architecture | Gen 4 GCN | Gen 4 GCN |
Playback Support | 4K encode/decode | 4K encode/decode |
Output Standard | DP1.3/1.4 HDR | DP1.3/1.4 HDR |
The AMD Radeon RX 470 and RX 460 graphics cards are considered “partner cards,” meaning reference variants will not exist for the low-end. We received the Sapphire RX 470 Platinum Edition for review, which we'll discuss with greater specificity below.
The RX 470 is on the Polaris 10 GPU architecture, as is the RX 480, and so low-level functions will remain mostly identical to what we described in our RX 480 review. The RX 470 is a slightly trimmed-down version of the RX 480 GPU, running 32 CUs (Compute Units) rather than the 36 CUs found on the RX 480. The Polaris 10 architecture runs 64 Stream Processors per CU, resulting in a chip with 2048 SPs for the RX 470. For reference, the RX 480 hosts 2304 Stream Processors.
AMD's core clock is listed as 926MHz base / 1206MHz boost for the RX 470, but partner models will have room to increase that clock-rate with pre-overclocks. We'll test overclocking later in the review. COMPUTE performance scales up to about 4.9TFLOPs on the RX 470.
The memory subsystem and data compression algorithms function as described in our original Polaris / RX 480 write-up, but at a slower speed. The RX 470 has a variable memory speed range of ~6.6~7.0Gbps (our card runs at 1750MHz actual), but is still on a 256-bit bus and capable of producing bandwidth upwards of 211GB/s. The RX 470 is available only as a 4GB card, with no 8GB models available; memory is GDDR5 for this platform, as it has been for the rest of Polaris.
The 4GB VRAM capacity means that AMD's closest current-gen competition is its own RX 480 4GB card, something we'll show in benchmarks later.
TDP for the RX 470 sits at 120W; interestingly, we were originally told to expect a 110W TDP at the May press event – but specs were unfinalized at the time. As a reminder, TDP is not a measure of actual power draw by the graphics card, but a measurement of cooling requirements for the chip. To this end, AMD and nVidia products are not comparable in their TDP values. You will need to look at actual power draw and thermal benchmarks to learn about comparative performance between manufacturers. A single 6-pin header and the PCI-e bus are responsible for powering the RX 470, though partners may modify this if desired.
Launch Day RX 470 Prices & Models
(Update - Here's a list of available card, as of this post going live:)
Brand | Model | Price |
PowerColor | PCS Red DEVIL RX 470 | $180 |
XFX | RX 470 RS Triple X | $210 |
XFX | RX 470 RS Black Edition True OC | $220 |
Gigabyte | RX 470 G1 Gaming | $200 |
Sapphire | RX 470 Nitro+ 8GB | $240 |
Sapphire | RX 470 Nitro+ 4GB | $210 |
Sapphire | RX 470 Platinum | $200 |
Polaris 10 Architecture Refresh
Architecturally, the RX 470 & 480 block diagrams will look largely the same – just with fewer CUs on the RX 470:
Above is the Polaris 10 block diagram that was provided for the RX 480. Mentally remove four CUs to create the RX 470 block diagram.
Polaris focused heavily on power efficiency improvements, working to leverage the new Global Foundries FinFET process in accompaniment with low-level optimizations. FinFET transistors, as we've explained a few times, use a three-dimensional design that extrudes a fin to form the drain and source of the transistor. The gate surrounds the fins, reducing power leakage at an infinitesimal level for cumulative efficiency improvement across the board.
AMD also optimized its datapath and data compression, both reducing bits transmitted and, as a result, power consumption. Memory uses a non-trivial amount of power and has seen reductions upwards of 40% in energy consumption per transacted bit. This is why AMD and nVidia spend each launch hyping up their new implementations of delta color compression, or the process of analyzing a scene for nearby comparable colors and compressing them into fewer colors. Calculations can then be applied to the compressed colors to extract the original value, hence “delta” color compression: Only a delta value is needed (the difference between, for instance, a compressed, neutral blue and the original light blue – like #0079f0). This reduces time crunching colors significantly, as color can be reduced on a scale of 8:1 maximally. A skybox is the easiest example of potential 4:1 or 8:1 delta color compression (DCC).
With memory energy consumption reduced, more power is potentially made available to the cores (but a net reduction in energy consumption is still the target scenario). Like the RX 480 – and every other known Polaris chip – the RX 470 uses 7 DPM states that scale from “sleep” (DPM1) to fully unlocked (DPM7). In between, the card adjusts along a volt-frequency curve to manage its power consumption for each frequency level, pursuant to the active DPM state.
Additional reductions to energy consumption were made by clock gating and power gating any under-utilized circuits, which cooperate with a heuristic pre-fetch routine to occupy all available cycles with instructions.
A single graphics command processor oversees instructions inbound from the CPU and dispatches those instructions, almost as a “forward,” to the new-ish Hardware Schedulers (HWS). We explain this process in our original RX 480 review, pasted here:
“The back-end of the render pipeline (GCP → Setup Engine → Scheduler / CEs) begins tasking incoming resources appropriately to low-level GPU components (some virtualized, some physical – ACEs, for instance, are a physical compute resources on the silicon). The Graphics Command Processor takes instruction from the CPU and sends it to the scheduler, which is a GPU component. The scheduler then begins the process of managing a familiar graphics pipeline (discussed here), e.g. drawing primitives and geometry, performing light/shading passes, eventually fetching textures, applying transforms, and preparing to rasterize the output. Post-processing, as always, happens at the end of the pipeline.”
And here's a look at the Polaris Compute Unit block diagram:
Basically the same in CU layout as the previous generation of GCN. See Gen 3 below:
Sapphire RX 470 Platinum Edition Specs
Sapphire's RX 470 Platinum Edition uses the reference RX 480 cooler, which we exposed and detailed in our tear-down of the 480. It's painted silver, hence “platinum,” but is otherwise the same as the reference cooler. Aside from a difference of paint, the RX 470 Platinum Edition also includes a back-plate and ~10MHz pre-overclock, though we rarely hit the 1216MHz “up to” rate advertised. Most of our logged testing still sat around 1206MHz, depending on the game.
So, then, the RX 470 Platinum Edition is about as close to a true “reference” RX 470 as we'll get, since AMD won't be making any. Even the board is effectively the same, which we'll look at in a video shortly.
Now, that stated, the RX 470 Platinum Edition does have a few differences in its fan configuration and programming. Our model maxes out at around 4000 RPM, a good deal lower than the 5200RPM screaming max of the RX 480 Reference card.
Let's move on. Enough theory and design discussion. Continue to page 2 for testing methodology.