The GAMING CARD They Tried to HIDE From You! (RX 7900 GRE Review)

This, rather curious beast, is AMD’s RX 7900 GRE, a card that has technically been out for about eight months now, and yet it’s only just hitting the market for anyone to buy. It’s a pretty confusing product, so let me explain what’s going on here. The first clue is in the name – “GRE” – that stands for “Golden Rabbit Edition”. This card was a China-only GPU. It launched at the end of July in 2023, and was retailing for around $650. Technically, that places it nicely between the 7800 XT and 7900 XT, although that isn’t the price you’ll find it at today…

Spec wise, it’s actually a lot closer to the 7900 XT, with 5120 cores, 320 texture units, 160 ROPs, although actually matches the 7800 XT with its memory config with 16GB of GDDR6 and a 256 bit memory bus. The 7900 XT only has 5% more cores, although the game clock is more like 8% faster, so in theory anyway it should be pretty close there. It’s actually an exact spec-for-spec match of the 6900 XT, save for the process node shrink and chiplet layout instead – the same cores, TMUs, ROPs, memory and bus width. So it should be pretty close in performance to that, and seeing as that’s the only card I have to compare it to, it seems too perfect not to!

Testing at 1440p, in Starfield on low settings you can expect functionally the same performance between the two cards. There’s only one FPS splitting them on average, with a bit of back and forth on the 1% and 0.1% lows. Cyberpunk has the GRE ahead by around six FPS average, although not much faster in the 1% lows, and slightly behind in the 0.1% lows. Realistically this is functionally identical performance. CS2 sees the 6900 XT taking the slightest lead, although especially for CS2 this is well within the margin of error. Fortnite sees a slight lead across the board for the GRE – although really not by much. Just 4 FPS average, which means this is, for all intents and purposes anyway, identical. Microsoft Flight Simulator again shows a functional tie, with just 0.1 FPS average between the two cards. Seriously, it’s that close. Hitman’s built in benchmark lets me break out the CPU and GPU performance separately, and of course we are looking at the GPU data here. Strangely, the 6900 XT actually has a compelling lead here – to the tune of 20 FPS average, and in the 1% lows, with only the 0.1% lows remaining the same. That’s a bit of a surprise to me, considering the spec itself is the same, and the two Gigabyte OC cards I’m using here both have the same 2050MHz game clock rating. Even more strangely, Siege shows an even larger performance gap, with 380 FPS on the 6900 XT, but “only” 330 FPS on the GRE. That’s about 15% faster from the XT, versus about 8% from Hitman. Happily, things return to a close race with Shadow of the Tomb Raider, with the XT only holding a 3 FPS lead over the GRE.

Taking the average of all eight games, the 6900 XT comes out with 221.9 FPS average, versus the GRE with just 213.1 FPS average. That makes the XT around 4.2% faster overall, or 8.8% faster in the 1% lows, and 13.1% faster in the 0.1% lows. That’s kind of a lot considering the spec is so similar, but that’s what I got anyway. It is worth noting that the 6900 XT will draw more power – it’s rated for 300W, whereas the GRE is more like 260W instead, and only has two 8 pin PCIe power connectors, rather than the three the 6900 XT has. That’s the benefit of a die shrink I suppose! The one benefit the GRE has here is the price tag – it’s now at $550, or between £500 and £550, which is basically the same price as a 7800 XT, and a good £200 cheaper than a 7900 XT, or about the same price as a used 6900 XT.

The question I’m left with though is, why does this card even exist? It’s in a weird position, technically cannibalising AMD’s own product line, so why did AMD bother to make this? Well the general answer is that when manufacturing GPU dies, the process of making them is a bit of a dark art, with a fair amount of luck involved. Typically that means that the dies TSMC makes for AMD aren’t always perfect. They have what’s called a “defect rate”, meaning for every X dies they try to make, Y of them will generally have some form of defect. TSMC claims their defect density for their N5 process node – the one AMD is using for the main core die here – is under 0.1 per square centimetre – around 0.06, give or take. Given that this is a pretty massive die, something like 30 by 10 mm, you should be able to get around 180 possible dies from a 30 centimetre wafer. With a 0.06 defect density, that means you get 147 “good dies”, 8 “partial dies” and 29 “defect dies”. So, let’s pretend that the “good dies” get to be 7900 XTX’s. There is further nuance there, of course, but let’s pretend. So now those 37 otherwise non-perfect dies can generally be tested and reconfigured to be offered as lower end cards, like the 7900 XT, and this GRE. Now some of those good dies might actually turn out to be not stable enough to run at the full XTX speed so they get dropped down too, and some of the defect dies might just be completely dead, so that isn’t a set thing either, but you can see the spread forming. The assumption, then, for this GRE, is that AMD must have had a bunch of dies that weren’t good enough to be XT’s, so they cut the memory bus down and shipped it as the GRE, starting in China where I assume the market for this type of card is perhaps a bit more fruitful for AMD, then when they ended up with more stock than they thought, they launched it as a global product instead. At least, that’s my theory anyway!

The result is the same though, we get a slightly faster card than the 7800 XT for basically the same price. That’s not to say it’s perfect – I think a wider memory bus would have helped a lot in eking more performance out the of the core itself, but it’s hard to argue that, at least at this price point, it isn’t a great option.