AMD Threadripper Is Looking Good, At Least Under Windows

[Gusar]

Test Drive wasn't too shabby in all its CGA glory

CGA? You spoiled brats. Hercules Monochrome graphics, that's where it was at!

If I remember correctly, a lot of the early 3D games did not need a GPU. I think one of the last big titles that could be played entirely on a CPU is Half Life (1998).

Unreal Tournament (1999) does not need a GPU. Need for Speed Porsche (2000) could work in software mode too. It seems it's somewhere around the turn of the millennium that games started to require GPUs.

[ssokolow]

Unreal Tournament (1999) does not need a GPU either. Need for Speed Porsche (2000) could work in software mode too. It seems it's somewhere around the turn of the millennium that games started to require GPUs.

I remember reading that Unreal Tournament's CPU-rendered mode is used as a compatibility hack for modern systems in some contexts... I can't remember the details though.

[Gusar]

I remember reading that Unreal Tournament's CPU-rendered mode is used as a compatibility hack for modern systems in some contexts... I can't remember the details though.

Graphics aren't the problem for UT99, modern CPUs are. For one, frequency scaling - so the game doesn't constantly go through speed up/slow down cycles, you need to use a fixed frequency CPU governor. The other problem is the game running super fast (like really super fast), this is hacked around by having the launch script stress the CPU 100% while the game launches. This is for the Linux version I have installed, no clue what's going on in the Windows world, though a quick googling shows me there's a D3D11 renderer for the game nowadays!

[linuxgeex]

Michael - if you've got time to kill during the benchmarks, I'm really curious how the 1950X will perform using softpipe or llvm-pipe. With so many threads, I figure at this point some games (older ones, like OpenArena) would end up being playable, at least on low detail settings.

ETWolf used to be playable on my Turion ML-34 with software rendering at 400x300 with Transgaming's SwiftShader. It was an order of magnitude faster than whatever the GLX-indirect renderer was 12 years ago. The current soft renderers are not designed for speed. They are designed for compliance. If we could disable much of their compliance we could make them way way faster.

Oh shizzle... I remember now this wasn't an OpenGL soft renderer, it was a DX9 soft-renderer and I had to run it via cedega.

[edit:]

OpenSWR has been merged into upstream Mesa as of late 2016 and supposedly provides up to 51x the performance of llvmpipe on an 18-core xeon, so...

$ LIBGL_ALWAYS_SOFTWARE=1 GALLIUM_DRIVER=swr openarena

Should give you the fastest software rendering on the latest Mesa, however they say in their docs that there's been performance regressions during the merge with Mesa so you might also want to try downloading the older version from github, building the libGL.so.1 replacement and using LD_PRELOAD instead of the above directions.

[salsadoom]

Test Drive wasn't too shabby in all its CGA glory

Wow this brings me back. Remember turning on the radio and rocking out to the PC speaker music? lol!

[hoohoo]

It never ceases to amaze me that as important as Linux is in the commercial high profit server world, AMD does not automatically send review samples to the most important mass market Linux tech site, Phoronix. I'll be watching for your test results, Michael!

[curaga]

Nice! It'll have been roughly 25 years since games were capable of being played comfortably using strictly a CPU. I'll be impressed if this Intel system and/or Epyc build can do the same.

I'm comfortably emulating ps2 in all software mode, on a Phenom II x6.

[hauberg]

Michael It would be very interesting to see benchmarks comparing different implementations of BLAS on ThreadRipper. A lot of benchmarks seem to rely on BLAS implementations that are optimized for Intel so it would be great to see which implementations work best for AMD.

[edwaleni]

How is it that AMD always seems to not want to send you samples, then decides to send you samples, but late? I thought they would have learned after Ryzen.

In the "old days", samples were sent to many of the technology publishing groups, like Byte, Penwell, IDG and ZD and they were shared in many cases between the editorial teams. Easier to control the editorial content and most time they also were large advertisers.

Today, there is an explosion of tech oriented websites, all requesting/demanding samples of everything, not just CPU's. Some sites are so desperate for ad clicks and SEO, they merely copy instantly what someone else has published verbatim.

If AMD and Intel served a sample to everyone who asked, it would probably kill sales.

Once upon a time, Intel used to call us and ask if they could send us a new build sample to test out. Today? To get a sample I get deferred over to an OEM. And even then if you don't do $1M, don't expect a big response.

As for Phoronix, I kind of like the fact that he relies on premium folk to acquire his hardware at times. It keeps his testing honest and unbiased.

If a publication that happen to be a large receiver of Intel advertising found a serious errata in a sample, would they notify Intel first or go to press?

If a website that had zero ad revenue from Intel found an errata, would they notify Intel or go to press first?

[linuxgeex]

It never ceases to amaze me that as important as Linux is in the commercial high profit server world, AMD does not automatically send review samples to the most important mass market Linux tech site, Phoronix. I'll be watching for your test results, Michael!

I appreciate the work Michael does in benchmarking overall performance, but he doesn't scratch the surface when it comes to understanding why the performance has changed. Both THG and Anandtech provide much deeper analysis, and it's no surprise that technologists and enthusiasts are going to read their articles first as a result.

Anandtech in particular does much deeper analysis of the hardware, providing data like core idle and uncore idle power consumption, L2 vs L3 cache timings per thread count, memory timings across NUMA boundaries with DDR4-2133-3200, power usage per core per thread count, 99th percentile FPS with common top GPUs, and insightful commentary on performance anomalies.

Michael's frequent "so there you have it" or "chalk another one up for AMD" type comments don't really mean anything to me, but he does have insights from time to time. ie "Dota 2 with OpenGL is one of the cases where Mesa Git is believed to be in a regressed performance state that has yet to be resolved on Mesa Git master." That provides some context for the result, instead of just blah blah agreeing with it. Even better would have been to cite an earlier benchmark which demonstrates the regression, instead of introducing it as ambiguous heresay... again its these little things that make a review site a lot more credible.

So, even as a hard-core Linux user who admins dozens of linux boxen, and even if Phoronix got samples first - I would still most likely read both those sites before coming here, because I'm a lot more interested in knowing why rather than just trusting an automated test harness and some "there you have it" comments to enlighten me about the values/qualities of the latest hardware.