Over the weekend, AMD announced its new R9 285 graphics card (look for reviews coming soon). This GPU is essentially a slimmed down R9 280X — it’s analogous to the old Radeon 7950, except that it has less RAM (2GB instead of 3GB) and the features that AMD introduced to the R9 290X family. That means TrueAudio support and the Asynchronous Command Engines that can handle eight commands instead of just two. It also supports AMD’s XDMA engine for better multi-GPU scaling.
Sapphire wasted no time announcing two versions of the card — a small Compact Edition at 17.1cm (6.7 inches) long and a standard model at 10.3 inches long. Both are double-wide cooler models, but the Compact Edition is still far more svelte. That raises an interesting question: why are graphics cards so big, anyway?
What big GPUs you have
The simple answer to this question is that the graphics card is far more than just the GPU — but it turns out that die size, TDP, and total card size tend to all follow each other. Below, we’ve graphed the TDP and die sizes for five different top-end Nvidia graphics cards, starting with the 55nm refresh of the GT200 family (the GTX 285) and continuing through the 28nm GTX 780 Ti.
Note that TDP and die size tend to follow each other at nearly the same slope. The GTX 680 is
clearly the winner here from an efficiency standpoint — it was markedly smaller and drew less power than the GTX 580, but performed significantly better than that card. (I used Nvidia for this comparison because the GTX 680 stands out as an extremely well-positioned GPU that delivered a particularly strong set of efficiency and power consumption improvements at the high end).
Original image by IXBT
Still, look at the back of a GTX 680 — our efficiency winner of the past six years — and you’ll see that the GPU itself takes up only a small part of the PCB. The internal square bounded by the four screw holes is the actual GPU — so what’s taking up all the rest of the board space?
Original image by IXBT
The front provides the answers. The memory chips sit in an array around the outside of the GPU die, surrounded by a further mesh of power circuitry. Voltage regulators, and third-party controller chips (if any were present). Most of the hardware on the card isn’t GPU. You can also see the silkscreen for the fan’s location — it’s the circular shape.
This is where physics gets in the way of our plans for tiny GTX 680s. The tighter your PCB layout, the smaller the gap between all of the various circuits. Designers can compensate to a degree by using higher-end components that draw less power, building better cooling solutions for the entire GPU, and binning parts to ensure only the best chips get used for the smaller form factors, but at the end of the day there’s a lower limit on just how small things can be. Past that point, you’re packing more and more heat into a smaller and smaller area.
The other downside to packing more components into smaller areas is that the heatsink becomes smaller in turn (there’s less area to cover). This typically means the system requires a smaller, higher RPM fan — and higher RPM fans tend, inevitably, to be louderfans. Thus, we inevitably end up with situations where making cards smaller also means making them louder and consumers won’t trade size for sound past a certain point.
Vastly increased efficiency without dramatic power consumption
One of the improvements that shines through a comparison like this, however, is the degree to which both AMD and Nvidia have improved performance without dramatically increasing power consumption. The official TDP on Nvidia’s top-end single GPU card in 2008 was modestly lower than today (204W vs. 250W), but the GTX 780 Ti would blow the doors off any GTX 285 configuration in existence.
Granted, Intel and AMD have long stuck with 140W TDPs as a sort of unofficial maximum as well, but even Intel hasn’t delivered the same kind of increase in real-world applications as Nvidia or AMD has over the same period. Therelative rate of increase in HPC applications might be similar given that Intel has added AVX and AVX2 between 2008 and 2014, but outside of scientific computing, those instruction sets don’t give the company an overwhelming advantage.
Incidentally, the fact that we require large GPUs and PCBs is part of why
APU graphics will always be in a permanent state of catchup. The GTX 780 may be fantastically efficient compared to the top GPUs of 2008, but it’s still drawing 250W with dedicated power circuitry and onboard RAM. There’s just no way to integrate that kind of configuration into a conventional socket — if there was, we’d have never needed discrete cards at all.
