Linux Format

AMD Threadripp­er X1950..

Sometimes they let Jonni Bidwell play with fancy hardware. But the tantrums that occur when he has to return it are oh so tiresome…

-

Jonni Bidwell takes the 32 threads of the latest AMD processor for a very quick spin around the block. It’s a monster.

We were impressed with AMD’s erstwhile flagship processor, the Ryzen 1800X (see LXF223). Now AMD’s back with a new beast, the 16-core, 32-thread ThreadRipp­er 1950X. Its predecesso­r gave Intel’s i7 profession­al class 6950X (which costs twice as much) a run for its money in many tests, particular­ly multi-threaded workloads. For gamers though, a more suitable comparison was found in Intel’s quad core 7700K, which cost less and performed significan­tly better at single-threaded workloads than the Ryzen. This time round it’s a similar story: there are expensive Intel chips for it to compete against, but there are also cheaper, fewer-cored offerings from Intel if all you care about is gaming.

AMD’s innovative Infinity Fabric enables its silicon to scale naturally. ThreadRipp­er can be thought of as two Ryzen 7s stuck together in a single package, wrapped in an attractive industrial orange bracket. Those Ryzen 7s can in turn be thought of as two four- core complexes (CCXes) Infinity Fabric’ed ( pretty sure that’ s not ave rb− Ed) together. Going the other way and sticking up to eight CCXes together, you get AMD’s enterprise level Epyc CPUs, aimed at datacentre­s. Such magic glue (you’ ll never make a marketing person− Ed) enables 64 cores to fit on a dual socket motherboar­d. That’s textbook journalist­ic oversimpli­fication, of course, and there are subtleties between the mainstream (Ryzen 5 for example), ThreadRipp­er and Epyc components that we won’t go into. What’s worth making clear is that ThreadRipp­er can handle up to 64 PCIe lanes, as opposed to the paltry 24 (of which only 20 are usable) of the Ryzen 7. This makes it much more tempting for those looking to build multi-GPU setups or house lots of high speed SSDs.

It’s a big ‘un

Size isn’t everything, but ThreadRipp­er’s dimensions are something to behold. It measures 72x55mm, which make it not that much smaller than a Raspberry Pi B+. It also has 4,094 pins and so doesn’t fit in an AM4 socket: this one calls for the new TR4 socket. Installati­on is reasonably hardcore, there are three torx screws that clamp down a retention bracket. Once this is released a tray can be lifted up and then two covers removed. Then the great orange thinking unit can be slotted in place, and everything battened back down. The ThreadRipp­er chips have a maximum power draw of 180W, so things are going to get hot (AMD recommends liquid cooling).

Looking at the base clocks, the 1950X’s 3.4GHz versus the 1800X’s 3.6GHz, it’s reasonable to expect that the 1950X won’t improve on the singlethre­aded results. Both have cores that boost to 4.0GHz (although this is handled differentl­y on ThreadRipp­er) so we’d expect performanc­e to be similar. With 32 threads, though, we would expect to see some pretty awesome results for parallelis­able workloads.

Our test bench featured the eminently covetable Asus ROG Extreme Zenith motherboar­d armed with 32GB of 2,400MHz DDR4 RAM, a nippy 512GB Samsung 960 Pro SSD and a GTX 1080. We found Fedora 26 booted okay on this, but working off a live USB was incredibly slow. Once installed and updated (to Kernel 4.12.5) we encountere­d no problems, not even with Nouveau. We put the chip through its paces with a selection of Phoronix

TestSuite benchmarks. As predicted, the single-threaded tests ( FFTW, GraphicsMa­gick, Lame,

FLAC and Blender) are very similar to Ryzen, which in turn was bested by Intel’s 7700K. Comparing these to results that are available on www.

openbenchm­arking.org, we see that the i9-7900X does even better here, particular­ly at multimedia workloads. It’s then reasonable to conjecture that the 7900X will be a better chip for games, which tend to not scale beyond around four cores. That said, we didn’t test AMD’s Gaming Mode (that would have required booting into Windows), so things may not be so clear cut.

The multithrea­ded results are a mixed bag, but before we go into any analysis, it’s worth rememberin­g that this is a new CPU and any anomalies we noted may be addressed in upcoming firmware or kernel updates. There could be issues within the benchmarks too, the FFmpeg test for example uses

-threads argument but actually passes in the number of cores. This may not matter for a quad core chip, but with a 16-core monster the difference will be marked. They could also be the result of errors, oversights or misconfigu­rations on our part. Maybe we should’ve tried Game Mode. With that aside, let our wild speculatio­ns begin.

JohnTheRip­per ought to be doing about twice as much, er, ripping on the ThreadRipp­er as we saw. It uses OpenMP for distributi­ng workloads across cores, so perhaps something is awry in our toolchain. Rendering with

TTSIOD was promising: we wouldn’t expect a straight doubling in performanc­e here, since this is a complex job that won’t always parallelis­e nicely. Likewise C-ray, and since this test runs for such a short time it’s reasonable to expect more complex ray tracing jobs to exhibit an even greater differenti­al.

Compiling kernels

We were really impressed with the kernel compilatio­n results, although the speedy SSD probably helped a lot here. This is a good real-world benchmark that tests raw processing power, I/O and concurrenc­y. The kernel (4.9 in this case) is so sprawling and GCC so clever that it rarely has to wait for one compile job to finish before others can start, so it achieves awesome concurrenc­y. We remember when compiling kernel’s took several cups of tea. The FFmpeg test decodes an H.264-encoded HD stream to NTSC DV, and was rather disappoint­ing with Ryzen. This time around we see an improvemen­t in line with what we’d expect from twice as many cores, but this test seems to do better on Intel silicon. ThreadRipp­er excels at the OpenSSL test (4,096-bit RSA signatures), doing a little better than results reported for the 7900X.

Intel’s soon-to-appear 12-, 14-, 16and 18-core monsters will see the blue team better compete with Threadripp­er’s multithrea­ded amazingnes­s (and continue to trounce it on single-threaded workloads). But with the 10-core 7900X costing the same as the 1950X (£900, $999), its bigger siblings won’t be competing on price. Intel’s top-of-the-range i9-7980XE will be on sale by the time you read this for $1,999 (probably Brexit priced at £1,800). That’s twice the price for two more cores.

ThreadRipp­er is a mighty powerful chip, but we can’t help but wonder how big this mythical high-end desktop market segment is. It’s great to see these numbers this side of £1,000, and there’s a certain class of people that will be sold on numbers alone. There are also people that spend days compiling/ rendering/transcodin­g − but these aren’t mainstream users. So maybe this potent chip is solving a problem that, for the many, doesn’t (yet) exist.

 ??  ??
 ??  ?? There’s something funky about our John the Ripper results. Running it in tandem with GPUTest failed to improve them.
There’s something funky about our John the Ripper results. Running it in tandem with GPUTest failed to improve them.
 ??  ?? This is what compiling a kernel in under 40s looks like.
This is what compiling a kernel in under 40s looks like.

Newspapers in English

Newspapers from Australia