Comment by Roark66 | Hacker News Reader

Comment by Roark66 | Hacker News Reader

I'm currently on Ryzen 3700X and asus tuf gaming x570 MB. I'd love more single core performance and more nvme/disk speed. Also I have a dual GPU setup, 2 M2 nvme, and 6 sata occupied.

I'm considering upgrading end of the year. I briefly considered thread ripper (for it's pcie lanes) until I found there is still no zen4 thread ripper and the price is likely yo be eye-watering.

So the choice is, upgrade to Ryzen 5900x and keep the rest of the system, or spend a lot more and upgrade to a AM5 ryzen plus a new mobo(I considered Intel too, until I found it tops at 20 pcie lanes).

I've always been buying mostly gigabyte and asus mobos, but I got burned more than once by them so I might go with another manufacturer this time.

What do you think? Is it worth upgrading to AM5 for just single core and disk performance? What about Intel? Considering I'd love more pcie lanes for a 10gb adapter perhaps(so I can move some of these spinning disks out) Intel is probably not a good choice.

I'm happy enough with the multi-core performance of ryzen 3700x. If I got a new MB I'd definitely want at least two nvmes in a mirror(for speed) - perhaps more and sata ports for my 6 spinning disks.

For your mirrored drives idea: I've had two nvme drives in mirror for a while but there are a lot of caveats to getting maximum performance. Not sure how it is on the AMD side but most Intel boards for example would have one m2 slot connected to the CPU and the other three via the chipset. Which makes them share a common bottleneck (also with your 10gb ethernet card)

For me it was in the end much faster to get a new board with pcie 5.0 support and a large enough single SSD. Total IOPS and throughput both turned out higher than my previous RAID 0 array.

I've got a 5950x and two NVME drives. The setup is the same as you describe. One has dedicated bandwidth and the other shares via the chipset. I just use them separately rather than trying to RAID them since the speed is fast enough for just about everything I need.

Indeed, that's what I ended up doing with the other drivers. Use them separately for different data types.

With my personal desktop (2x 1TB NVMe drives), they're mirrored with ZFS for data integrity. Depending on what Roark66 has them mirrored for (speed vs data integrity, etc), the alternative single-drive approach might not be workable.

Well he wrote that it's for speed, so that would be raid0

> If I got a new MB I'd definitely want at least two nvmes in a mirror(for speed)

raid0 isn't really a mirror though.

raid1 is a mirror, and can have speed benefits over just a single drive (able to read from multiple drives in parallel).

But yeah, you might be right. When talking casually, things like "mirror", "raid" etc can be a bit fuzzily applied. :)

You're right. The primary use of mirroring on raid 1 is for redundancy. Raid 0 is striping and more associated with performance.

What kinds of workloads do you run that you are capped by nvme speed?

Loading ML models into GPU memory.

I have 5900X and I wish I’ve waited for 5800X3D instead.

What’s the rationale for that? Gaming?

Very likely. I jumped from 3700x and the difference is staggering. My 6800xt came alive and stutters are a thing of the past.

I kind of want an OLED but they don't come cheap, vertical resolution is the same, and they are not amazing for coding. Maybe in time for a Zen 5 based 3d cache chip.

>My 6800xt came alive and stutters are a thing of the past.

Potentially because of resizable BAR. 3** CPUs didn't support it, 5** did.

if you need more go 5950x for double the cores with only 4/3 of the price or 3d cache cpus for that factorio UPS. 5900x is not that of an upgrade

Reply on news.ycombinator.com