AMD Ryzen Threadripper 1950X 16C/32T CPU gets Geekbenched

[qasdfdsaq]

AMD's SMT implementation IS significantly more efficient than Intel's. It was one of the big discussion points when Ryzen benchmarks first started rolling out, and since there's an 8C/16T Intel chip to compare it with it's actually one that's very easy to see. The 1800X is ~15% slower than a 6900k in PiFast, but ~ 15% faster in handbrake. The 6900K has a slightly more agressive boost clock, but only to the tune of ~ 4% more above base, so you can't put a 30% performance swing from ST - MT down to more or less aggressive boosts - the simple fact is that AMD's SMT implementation is very efficient...



You do realise that this is exactly the same point that Friesiansam made all of 11 posts ago, don't you? And that the same arguments against it apply? The score is far too low, compared to other processors (including other Ryzen processors), for this to simply be an optimisation issue, so people are looking for reasons that might explain the score being very low. It's not just a case of "it requires optimisation" - it should be scoring close to twice that much. It's a serious anomaly - serious enough that it's bound to cause speculation. If the score is right, 16 core Threadripper is no faster than 8 Core Ryzen. And that's a BIG issue from AMD.

But +80% efficient? That's more than Power8 (designed for 8-way SMT), Intel's SMT, *And* the old Pentium HT combined

[scaryjim]

But +80% efficient? ...

tbf I didn't have much of an idea what kind of throughput SMT should be getting. Hunting around for reviews that test Ryzen with SMT on and off, it looks like it can get north of 50% additional throughput from SMT in ideal situations, but certainly nowhere near that 80% figure.

But all that tells us is that something is even wonkier with these results than we initially thought. That's not the result of a 16C/32T processor running slower than an 1800X, equally it's unlikely to be the result of a 16C/16T run at similar clocks to an 1800X.

So what is it? Well the leaked Threadripper line-up lacks a 1950X, but does have a 1956X, which is a 12C/24T @ 3.2GHz base. Running that with SMT off might produce comparable results - it's got 50% more cores than an 1800X, but is lower clocked, so the 1800X wouldn't need as much as +50% throughput to get in the same ball park for the MT result.

*shrug* it's all conjecture of course, but I find this quite an interesting puzzle - the performance is so far off what you'd expect from the comparison processors...