Hacker News Reader

Comment by ehutch79 | Hacker News Reader

So, does this mean threadripper is unusable and we shouldn't buy them?

If your workload is exclusively max parallel AVX2 workloads then you probably do want to hold off a bit, yes.

For everything else, though, there doesn't seem to be any cause for concern. It's currently a single stress test in a single very specific configuration that's failing.

For reference the tests that pass:

    MemTest86 v8.3, 4 passes (~8 hours), RAM at 3466 MT/s [PASS]
    MemTestPro 7.0 (paid version of HCI MemTest), 700% (~5 hours), RAM at 3466 MT/s [PASS]
    AIDA64 6.20.5300 System Stability Test (full system test except local disks) [PASS]
    IntelBurnTest v2.54 (based on Intel Linpack), maximum stress level, all available RAM, 20 runs [PASS]
    OCCT 5.5.3 test, large data set, 64 threads, AVX2 [PASS]
    Google stressapptest 1.0.9, all available RAM [PASS]
    Prime95 v29.8 build 6 torture test, 64 threads, min/max FFT = 4K, in-place FFTs, AVX2 (~1 hour) [PASS]
    Prime95, min/max FFT = 8K, in-place FFTs, AVX2 [PASS]
    Prime95, min/max FFT = 16K, in-place FFTs, without AVX2 [PASS]

The tests that fail:

    Prime95, min/max FFT = 16K (in-place FFTs or not), with AVX2 [INSTANT FAIL]

Note that an AVX2 workload is both in the passing & failing camp. So it's not as clear cut as "AVX2 doesn't work"

The link is in response to @DerAlbi, who suspects the VRM of hisGigabyte mobo, not the CPU. DerAlbi is techo enough to use an oscilloscope so he’s more likely to have a clue than many.

Read DerAlbi’s (recent) reply on thread:

https://forum.level1techs.com/t/3970x-prime95-stability/1532...

I.e. if an AMD CPU suits your needs, you should get one (unless you are doing Prime95 as a load!), and perhaps avoid a MoBo with that VRM. Or wait a couple of days for the truth to percolate up through the murk.

When it becomes reachable through WASM or browser Javascript optimization (if it isn't already), that seems like a big problem (easy DoS, probably a tricky RCX). Ideally browser vendors will hold off on that; besides this, there are the Intel throttling issues for some AVX2 instructions that could be an annoyance. But AVX2 is good for benchmarks...

I'm not sure why I'm being downvoted. This is a legitimate question. Does this bug make the cpu unusable? Does the work around for it slow down the per core performance to the point of unusability?

Whether the issue is with the CPU hardware, the mainboard design (VRM, etc), mainboard BIOS, kernel, or the Prime95 app itself still appears to be an open question.

Based on oscilloscope analysis of the VRM output in a linked thread elsewhere in the comments it looks like the board’s VRM design, or its configuration by the board’s BIOS, may be the most likely suspect.

But there are less-researched reports of similar issues on other boards as well, which makes things a bit more murky.

Given the uncertainties there it may put some people off from buying into the TR/sTRX40 platform in general. But to offer a blanket recommendation to avoid is a bit premature.

It's a little early to start condemning things. It might be motherboard power delivery, a prime95 bug or something else. If it is the CPU there could be a firmware correctable flaw. This is the bleeding edge; these devices have only been available for 13 weeks and someone is indeed bleeding, it seems.

It could even be an uncorrectable flaw in the device design, in which case AMD will likely do as they did in 2017: replace the parts.

> This is the bleeding edge; these devices have only been available for 13 weeks [...]

Shouldn't these devices have been tested by AMD and motherboard manufacturers for months before they were released? Including on workloads like Prime95 which are well known to uncover system instabilities.

Keep in mind also that this is not the first time in the Zen/Zen2 history that AMD has shipped buggy products: there was the segfault bug, the random number generator instruction bug, and the unbootable Threadripper on recent (at the time of release) Linux kernel. The worst part, to me personally, is that they were all "surface bugs" that could have been detected by AMD with a little bit of testing.

I've been wondering why Supermicro isn't releasing any Threadripper motherboards (while they do for Xeon-W). Maybe this explains it: the consumer CPU business at AMD is a circus and Supermicro wants none of it?

Unless it shows up in other workloads, or your workload is specifically Prime95, it doesn't make it unusable, it just means you can't do Prime95.

I overclocked a threadripper 3950x on a Gigabyte card and never encountered any problems in the various stress tests I ran. That machine convinced me that AMD's presently marketed consumer CPUs are superior to Intel's offerings, dollar for dollar. (I also benchmarked a comparable Intel chip. Both machines were cooled with a Corsair H100, ie liquid pumped through a radiator.)

> So, does this mean threadripper is unusable and we shouldn't buy them?

There's only a couple of datapoints. No official confirmation of an issue. This could be the fault of the CPU, Chipset, specific motherboard, or a bad batch.

Reply on news.ycombinator.com