I was very excited to make this post. But my testing wasn't really as long or as detailed as I wanted, because, well, I am using my main PC for this test. I actually overcame the anxiety to place the R9 280X in my main PC to test with, against the Radeon VII. So there's that. Anyway...
Back in 2012 AMD released the first ever GPU based on the, at the time, brand new Graphics Core Next architecture. An architecture that we have all grown to know and love these long 7 years. Well here is what I thought would be interesting to test: how much performance has AMD gained from this architecture, in those 7 years? Specifically: I wanted to get a small idea of the performance gain from early 2012's flagship single-GPU GCN part, the Radeon HD 7970, to the flagship gaming-oriented single-GPU GCN part you buy today, in 2019, the mighty Radeon VII.
What also makes this interesting is that the Vega 20 GPU powering Radeon VII, is actually around the same size as the Tahiti chip in HD 7970. So it's sort of a measure of "how much performance has AMD squeezed from a ~330-350mm² die area in 7 years", roughly.
Okay let's be exact here:
This is Tahiti. This processor is approximately 352mm² in size, and contains ~4.3 billion transistors on a 28nm lithrography made by Taiwan Semiconductor Manufacturing Company (TSMC). That is about 12.2 million transistors per every square milimetre. Back in 2012, this was cutting edge technology. Things have changed a bit since then so let's take a look at what 2019's GCN king looks like~
I love this chip. The HBM2 makes it look so cool. Okay, so this is the Vega 20 GPU. Well, that chip in the middle is. It is, of course, surrounded by four stacks of High Bandwidth Memory (2nd Generation). Each stack on this particular package contains four 1GB DRAM chips stacked on top of a logic die, then connected to an interposer and wired directly into the GPU via a 1024-bit interface. That's wide. 4096-bits wide, actually, if you add up the entire interface. Anyway! That GPU in the middle is approximately 331mm². It's actually slightly smaller than Tahiti - but the impressive bit, is it packs 13.2 billion transistors into that space. That is around 40 million transistors per square milimetre, or approximately 3.2X the density of Tahiti. That is made possible by using TSMC's latest 7 nanometre manufacturing technology. Cutting edge to cutting edge.
It's also worth pointing out, that Vega 20 used in the Radeon VII isn't actually a full enabled chip. The die has four Compute Units fused off, to improve yields. That reduces the Stream processor count from 4096 to 3840, and the texture unit count from 256 to 224. A relatively minor reduction. HD 7970 is a full chip, so keep that in mind. it's still valid since Radeon VII is the flagship single GPU GCN gaming graphics card for 2019. At least I think it will be, since AMD has announced that Navi-based lineup will co-exist with Radeon VII this year.
Okay, now for my boring notes. :D
Well, I could bore you with more specifications and details, I just love typing about GPUs, but I want to get to the results. But before that, I must make a few notes...
Firstly, I do not actually own an HD 7970. But I assume, like, everyone who is interested in this information knows that the R9 280X that released as the third-best Radeon single-GPU card in 2013 alongside the R9 290X and 290, was actually simply an HD 7970 GHz edition with a new name. And that card is a factory overclocked HD 7970. That is rebrand-ception. But hey, it works.
So I am taking my R9 280X, the exact same silicon as the HD 7970 and I am setting the GPU core and memory clock speeds to match the reference launch specification for the Radeon HD 7970, that released in January 2012. That is 925 MHz GPU core, and 5.5 Gbps on the GDDR5 memory. I know many custom cards existed with higher clock rates, but since Radeon VII is a reference design card and I test at fully stock speeds, it would be apples-oranges to compare a factory overclocked HD 7970 to a reference Radeon VII. Higher clock speeds on Vega are part and parcel with the architecture. And, no, sorry I will not test overclocked results as overclocking graphics cards scares the crap out of me. Okay, maybe I will in the future, but please don't hold your breath over it, okay?
But your Radeon VII is liquid cooled! It's not a reference design! Oh snap, that's true. But it is running at reference clocks and honestly, in my testing it doesn't actually throttle during light benchmark results, on the stock cooler. When I used it for a day, then ripped it off because I had ringing in my ears. >_>. The Radeon VII is running at approximately 1750 MHz core, it fluctuates a bit above and below that, as per Vega's dynamic boosting algorithm, but it's about that, I checked it don't worry. HBM2 is obviously at 2Gbps. 1TB/s, yay.
My second note is about my test system. For this test as I mentioned earlier, I used my main Gaming PC. I do not usually do this as I'm sure I've said a billion times I have some dumb anxiety problem and putting hardware in and out of my main PC causes me stress, unless I am fully prepared to do it. I have a testing PC to mess with daily (multiple times a day) and that is built for that purpose, and blah blah. But I am not putting my liquid-cooled Radeon VII in the Ryzen 3 1200 PC. So here is the specification for the PC used to host these two GPUs:
Ryzen 7 2700X eight-core CPU using increased Precision Boost Overdrive values, ~4.2 GHz on all cores is observed in bench-marking workloads
16GB (2x8GB) 3200 MHz CL14-14-14-30 with tightened sub timings as per this extremely handy tool. These are Samsung B-die chips.
MSI B450M Mortar
Seasonic Focus Plus 850W Gold Powersupply
Radeon Adrenalin version 19.5.2
I know I'm typing a lot here, but I like to be thorough. Please bear with me for a moment :D. I would also like to note, that unlike previous tests, I am using "AMD Optimised" Tessellation settings for both cards here, as this is not a synthetic test of geometry performance. FreeSync is disabled when testing the Radeon VII, as the Tahiti chip doesn't support it. It probably wouldn't make a difference but it bugged me that it might somehow. Textures are often set to Low where possible, as 3GB really is completely inadequate the resolutions I was testing at (to keep the Radeon VII GPU-bound, even Intel CPU would struggle with 150-200 FPS honestly).
I made sure the GPU wasn't VRAM starving, which I had observed a few times, that produces inconsistent results (One of my benches was producing ~20% variance in average FPS each time due to VRAM issues. Lowering the resolution fixed this and the results are re-tested for verification purposes). I also monitor GPU load, clock speeds and VRAM allocation using HWINFO64, and analyse this information between runs to assess any potential issues.
Also, I didn't test as many games and benchmarks because I was struggling a bit with the anxiety and, you know. That just sums me up, but hey I think these results are interesting enough. Oh jeeze, I am typing a lot. Okay I will stop now and here are the results.