(Tech Stuff, maybe a Tech Babble?) GA106, Sash's affair with the Mid-Range Ampere Graphics Processor
Where do I start? Oh yes, I know where to start; about the subject of this post. That would be a good place to start. No massive digression when I start actually typing, uh. Yes. So, this post is because I am a bit Autistic and somewhat excited about the fact (even though I have had these for over a month now) that I have a pair of GA106 GPUs in my PC. They are not even matching graphics cards, but the GPUs that power them are of course both GA106. The two cards I am referring to are RTX 3060 and RTX 3050. You might ask why I have this weird combination of cards, and if that is so, you're in luck as I was going to explain that right now. Lucky you, huh?
Well, the original plan for Sash's graphics processing capabilities in 2022 was to purchase an RTX 3060 12GB; because, despite the somewhat mediocre perf/£ over previous generation or competing AMD cards, the 12GB VRAM size and good Ray Tracing performance was beneficial for my new blender projects with OptiX acceleration, among other things (like UHD Texture packs). AMD doesn't offer the 12GB until you hit the 6700XT series which is a LOT more expensive and still not below £500 here, which is, well, quite out of my budget.
So, that was the plan. I was hoping to get one for MSRP (lol) but that wasn't going to happen. Instead, I happened to stumble upon SCAN.co.uk's website just when they got a batch of RTX 3050s in. They had a deal on one, for MSRP (I know right?),. that is £239.99, and I grabbed it. I couldn't pass up the opportunity to snag RTX 3060's crippled little brother for what it was actually intended to sell at, so there we go.
Well, fast forward to today, and I had asked my long-term friend, Borb, to lend me the money to buy an RTX 3060 12GB too, since I found one for "only" £399.99, and I was already pushing the 8GB on the 3050 with my projects in Blender. So, the Bank of Borb obliged and I ended up with a 3060, too. The plan was to sell the 3050 to pay some of Borb back, but my mum's new partner (he's a great guy) is awesome, and lent me the money to pay Borb back (since he used Interest free credit with a time-limited interest free period), and I would pay Garry (that's his name) back without the strict time limit. This allowed me to consider keeping the 3050, as I can pay Garry back in instalments easily without selling it. That, and the price of 3050s on second hand market was going down, anyway.
So why keep the 3050, you might ask, huh? Well, the answer is quite simple and actually has two distinct reasons. Neither are really related to gaming, but a lot of my PC these days is tooled for other purposes. The primary purpose is that Cycles Renderer in Blender, using OptiX, can use both the 3060 and 3050 together for Ray Tracing the same scene. That's around a 50% speed-up there (eyeballed testing), it's not quite as high scaling as the hardware but it's a very good result. Trust me, on my latest project, where I have nearly 5000 frames that would take around 15-20 seconds each on the 3060 alone, the extra RT performance makes a huge difference when I'm waiting >12 hours for the scene to finish.
The second reason is transcoding. I have two pre-sets for Handbrake using the gpu=0/gpu=1 selector arguments, and can encode two videos at once at full speed with NVENC, which greatly speeds up the compression of archival gaming footage for my increasingly bloated collection of gaming videos. Oh, and I think Adobe Premiere Pro can use both GPUs, too. I think. Don't quote me on that, but the 3050 clocks up and has some stuff in memory when I'm using it so I assume it's doing something.
Didn't I actually want to talk about the hardware? I mean, yes, but apparently the entire post was filled up with why I have them, and now I am a bit burned out with this post, so I'll make this quick.
GA106 is a mid-sized graphics processor based on NVIDIA's 'Ampere' architecture. It's about 280mm2 (CBA to google) and has ~13 billion transistors (ish) on Samsung's 8nm process. The use of Samsung Foundries for NVIDIA GPUs is not new; Pascal-based GP108 and GP107 processors used Samsung's 14nm lithography, but the use Samsung over TSMC for much larger, higher performance GPUs (in fact, the entire consumer Ampere series) is pretty much a new occurrence.
Anyway, the GA106 represents what I consider to be a spiritual successor to the TU116 chip in the Turing 16-series (the weird, cheaper versions of Turing without the RT/Tensor cores). This is because the apparent series predecessor, TU106 (RTX 2060, RTX 2070 series) is much more similar to the larger TU104-class GPUs but with toned down GPC count, than the more value oriented mainstream 106 class. TU106 in that regard is pretty unique as it's the first NVIDIA 106 class GPU with a 256bit memory interface (they typically have 192bits, this goes back to the first 106 class, the Fermi-based GF106 in the GTS 450). Aaaanyway...
If you compare GA106 to TU106 in terms of performance you will be a bit disappointed with Ampere's gains, as the two GPUs are about comparable in raster performance, well, I think GA106 is a bit faster - we don't have a fully enabled GA106 on desktop, since the 3060 still has a single TPC (2 SM, 256 FP32) disabled (28/30 SM, 3584/3840 FP32), and the RTX 2070 is the fully activated TU106. Though, compared to the relatively crippled RTX 2060 (which itself had a cut down bus to 192bit), the 3060 looks a bit better. I am digressing again...
Ampere's SMs are stronger than Turing's (largely thanks to the addition of FP32 math pipes on the INT32 ports), so NVIDIA probably decided it was time to shift the 106 class back to its historical configuration of 192bit/3GPC. In that regard, from the top level, GA106 is much more comparable to an upgraded TU116 (GTX 1660 series) than the TU106. In a sense, GA106 is a TU116 with 30 SMs instead of 24, but everything else is pretty much the same, well except for the Ampere improvements of course. (I am talking in terms of layout, each GPC has 2 more SMs).
So, consider the RTX 3060 and 3050 to be the "Ray Tracing enabled, upgraded versions" of the GTX 1660 SUPER and GTX 1650 SUPER, respectively. Because that's what I consider them to be.
And on that bombshell, I am going to (maybe) play Elite Dangerous outside on my mini-lap. But then again, I might not. Actually leaving my house these days is becoming increasingly more difficult.