THREADRIPPER 2970WX

AMD’s new workstation CPU rips up the benchmarks

AFTER SMASHING a personal target in a Cinebench overclocking push, it’s hard to ignore the sensation of victory hovering overhead. Over 5,000 points in Cinebench, competitive gaming scores, low idle power draw, x264 over the 81fps mark—it’s all there. A kick-ass processor, and we haven’t even stepped out of the opening paragraph.

That said, the 2970WX isn’t for the faint of heart, and it’s not aimed at gamers either. It’s a finely honed tool, tempered to maximize your efficiency in any and all workloads. There are 24 cores thumping away at the heart of this chip, six per CPU die, each die communicating with one another to break up those application processes across the 48 threads. Those zeros and ones are catered for by a huge 78.25MB of cache (64MB of L3, 12MB of L2, 2.25MB of I-Cache). On top of all of that, there’s still support for quad-channel memory, 64 PCIe lanes, and all the features that the mainstream Threa-dripper series has had for the last year or more (we’re talking precision boost, and all the software tweaks you can muster through SenseMi). Oh, and at stock, it’s cool. Seriously cool. We’re talking maximum load temps of 50 C, under a 240mm AIO.

But before all that, let’s talk about memory. Basically, only two of the dies have direct access to the memory channels, each one having two. The other two dies have to send and receive their memory requests through those first two dies directly, in a sort of piggyback mode. In short, you can expect memory latency all the way up above 100ms per request—not a problem for professional workloads, but not good for gaming. With two lanes going into two dies, and there being four lanes total, the obvious question is: Why not redesign the chip so that one memory channel comes from each die? The reason for that is bandwidth. Split the channels up, and you cut the bandwidth in half for those two main dies, reducing performance significantly. Then why no octo-channel memory instead? Well, that would mean a new socket, and new motherboards.

How great an effect does this have, then? Well, overall, it really does depend on which game you’re testing. Across our suite, only Total War: War hammer I I suffered because of the memory allocation mode, with frames dropping by half. It’s very much a case by case scenario. Switch to game mode, however, and AMD’s Ryzen-Master drops to local mode memory only, with Total War’s performance sitting where we would expect. That’s a bit annoying, because it does require a full system restart, but in today’s age of superfast SSDs, it’s just that—an annoyance at most.

Computational performance, though, is nothing short of spectacular. Over 4,400 points in Cinebench, 172 in single-core mode, x264 above 80fps, and load temps of 50 C make this a killer production part. We’ve gone for a thin spread of paste across the entire IHS on this one, similar to our “Build It,” and that seems to work best. As for overclocking, cranking the voltage up to 1.43V was enough to grab 4.15GHz across all 24 cores. With temps sitting happily around the 75 C mark, we managed over 5,000 points in Cinebench R15. The irony is it took five consecutive attempts to get that figure, each one quicker than the last, as the 2970WX completes the run so fast, the cores don’t have enough time to hit max load before they’ve completed the test. Keeping CPU utilization high through successive runs was the only solution—a rolling start.

Is this CPU for you, then? Well, it depends on what you want. For the price, it’s an incredible feat of engineering brilliance. If CAD/CAM, 3D design, or rendering is your game, those 24 cores are simply awe-inspiring. If you’re solely after a gaming chip, though, the 2950X is both better value for money, and less annoying overall in games that need that local memory mode. –ZAK STOREY