Wanting better disk and memory throughput. Is a general AM5 build the answer?

  • Thread starter Thread starter JNZ
  • Start date Start date

JNZ

JNZ

Associate
Joined
16 Feb 2024
Posts
12
Location
UK
Hi,

I'm debating the pros/cons of upgrading. My current system is several years old and frankly going very strong even in 2024. The only sticking point is disk and/or memory IO is not the fastest, but I did build this system originally for performance (as it was 8 years ago).

My current system is an i7 9700K running at 5 GHz on air, 32 Gb 3466 MHz DDR4 16-18-18-36 memory, and 4 Tb 990 Pro SSD among a variety of other SATA SSDs and HDDs.

I've been questioning the wisdom of throwing the best part of £3000 at a new motherboard, CPU, and memory.

I have long been debating the following spec:

AMD Ryzen 9 7950X3D
2x32 Gb Dominator Titanium 6000 MHz 30-36-36-76
ASUS AMD Ryzen ROG STRIX X670E-E GAMING WIFI AM5 ATX
Antec HCG1000 Pro Platinum 1000 Watt ATX 3.1 PSU
24 Tb Ultrastar DC HC580 SE

£2278

...I must say, that is quite a bit less than I was expecting (I already upgraded my GPU earlier in the year...). The 24 Tb drive is nearly £600, so the actual system is ~£1680.

Hmm. I see no reasons to not get the 7950X3D. Even in benchmarks where they say the 7800X3D is very slightly better, the 7950X3D generally out-performs it (especially on 1% lows). https://www.techspot.com/review/2821-amd-ryzen-7800x3d-7900x3d-7950x3d/

All other hardware will be moved from the current system.

My GPU is a 4070 Ti Super, and I have no plans to change it any time soon. I use my system for game dev, so want another system that will last another 10 years as this one has. I'm actually not looking forward to replacing my current system because it has been the best computer I ever had. I'm not even sure I can justify changing it.
 
Last edited:
Disk throughput... while technically PCIe4/5 has faster bandwidth, thats very difficult to notice in real life
Memory throughput.. DDR5 definitely improves, how much really depends on if CPU can use it. AM5 is good up to DDR5-6200

Your 7950X3D looks good. Combination price/productivity/gaming can't be currently beat

Since it reads like there isn't much urgency with this, mandatory section "wait for next stuff":
- New AM5 AMD X870 motherboards rumored to have better memory traces, may make out of the box DDR5-8000 configuration viable (end of Sep)
- Intel Arrow Lake should be ok (end of Oct)
- 9800X3D/9950X3D somewhere after Jan

PS. spinning rust storage prices are insane
 
  • Like
Reactions: JNZ
The only sticking point is disk and/or memory IO is not the fastest, but I did build this system originally for performance (as it was 8 years ago).
Can you describe this in more detail? Why do you need them? For what purpose? How do you know that is the bottleneck?

If I speak without answers to those questions, then AM5 is unquestionably the best consumer platform, but if you REALLY need those things then you'd likely be looking at a threadripper, epyc or xeon PC, since they have way more PCI-E lanes and much higher memory bandwidth than what AM5 can offer you.

An example of why AM5 is the best:
- With PCI-E 5.0, AM5 can have up to 2x CPU connected PCI-E 5.0 M.2 slots with no impact on the PCI-E lanes, though you'll usually need a X670E board for that.

24 Tb Ultrastar DC HC580 SE
I'm confused about this choice. If you want higher disk I/O, why are you using a huge hard drive?

There's usually two things you would do in this situation:
1. Shift your storage onto a NAS and use a network cable.
2. Use a fast cache (i.e. SSDs) for your frequently used stuff (e.g. your current projects) and move the archives onto slower HDDs, or the NAS.
 
  • Like
Reactions: JNZ
Hi,

Great questions!

@Tetras :

Can you describe this in more detail? Why do you need them? For what purpose? How do you know that is the bottleneck?
If you want higher disk I/O, why are you using a huge hard drive?
I appreciate this appears to be an oxymoron.

I already do (2) in your list with my current set up. I have previously considered NAS, but is it faster than just having a drive hooked up to SATA? I know SATA isn't the fastest interface, but for my needs it is fast enough as an interface.

I'm not doing video production so don't need SUSTAINED disk IO to be super-fast (that would require an array of some kind). The disks are not being maxed out pretty much any time during a read operation. It's read...process...read...process... every 1/20th of a second or so. It's the processor/memory bandwidth that is the limiting factor.

I need the drive for storage, plain and simple. The actual amount of data being accessed "in use" is probably only a few tens of gigabytes at most at the peak. Once it is loaded, it is fairly moot, but the load times are the killer because I need to keep loading things to test.

I already reduced load time from 6 minutes down to 2 minutes with the new SSD (990 Pro). That is considering that it is in the PCI-E slot attached to the chipset, too. I have another SSD in the other slot and would need to dismantle half the computer to get it out. I'm not sure the gains are worth the effort given the rest of the system specs which is why I didn't do it the first time (I doubt it would be much faster than it already is).

The fact the disk isn't being maxed out during load suggests there are much better areas to deal with first. After the first load, most data is cached anyway, so subsequent loads are from memory, and that is where the system could be improved for sure. Compared to SSD access times, the memory access is only around 10% faster (cold load vs. cached load). In fact, I'm going to go and time it to get an accurate benchmark.

@alec:

Disk throughput... while technically PCIe4/5 has faster bandwidth, thats very difficult to notice in real life
Memory throughput.. DDR5 definitely improves, how much really depends on if CPU can use it. AM5 is good up to DDR5-6200
The disks are not maxed out as it is, so my primary focus is on memory IO. Seems this would be a good upgrade.

I can wait a month to see what happens. Will it push the prices down much on the current hardware? This build is already some £500 cheaper than when I first looked at it about 6 months ago.

PS. spinning rust storage prices are insane
Cheap or expensive? I think I paid £250 for 8 Tb 18 months ago. It's around 25% cheaper per Tb today.

Thank you!
 
Last edited:
I did some testing...

Seems the SSD is bottlenecked by the memory speed (???). SSD is capable of 7000 MBps, connected to PCIe 4.0 bus. System memory is 3466 MHz, so I'm guessing the SSD is going to max out at 3 GBps if I'm lucky. I can't see how it can exceed memory speed as where does the data go? So memory of this system must certainly be the limiting factor.

With that said...

Steam takes 28 seconds to load.

FS2020 takes 2 mins 30 to get to the Welcome screen, and another 1 minute to get flying for a total of 4 minutes.

I guess just from an interface perspective, the limiting factor will be main system memory, so assuming I built a 7950X3D system with 6000 MHz RAM, then I should be getting close to the theoretical max performance of the SSD drive I have? It would remain PCIe 4.0 spec as it is a PCIe 4.0 disk.

Anyone have a Zen 4/5 system with a Samsung 990 Pro to test? Otherwise, I can't see any feasible way of improving much on what I already have (I did consider RAID but given the disks aren't maxed out, this won't help).

Game-performance wise, I play at 4K, so most times will be GPU limited. While I expect some improvement, I'm not expecting miracles. I'm primarily trying to reduce load times (the reason that matters is I'm a software dev and I need to load things constantly to test).
 
Last edited:
System memory is 3466 MHz, so I'm guessing the SSD is going to max out at 3 GBps if I'm lucky.
The bandwidth of RAM is way higher than that. SSDs are usually super slow in comparison to RAM.

With that said...

Steam takes 28 seconds to load.

FS2020 takes 2 mins 30 to get to the Welcome screen, and another 1 minute to get flying for a total of 4 minutes.
I'm primarily trying to reduce load times (the reason that matters is I'm a software dev and I need to load things constantly to test).
This is complicated, since it depends on the app, but you can check task manager to observe the utilisation.

There's been a lot of benchmarks done with PCI-E 5.0 drives and in comparison to the fast PCI-E 4.0 drives the difference is like.. a second or two :o

From what I've seen when monitoring my own usage, it is often just a single core of the CPU that is holding everything up.

I'll try and find some relevant benchmarks which might give you some pointers about when RAM is a bottleneck and when it is not (and the same for SSD).




If I tried to cover all bases, I'd suggest something like:

My basket at OcUK:

Total: £826.96 (includes delivery: £7.99)​

Then get a fast PCIE5 drive, if you still find it is bottlenecking.

The reason I went 9000 instead of 7000 is because of two things:
1. The X3D CPUs have lower single core/thread and I think you want the highest burst performance you can get?
2. The 9000 CPUs do have some advantages like improved AVX512 support, which can help app performance in some circumstances.

The 2x CCD CPUs are more complicated to handle if you don't need the extra cores.

(PassMark: single, multi)
i7-9700K: 2875, 14460
7800X3D: 3750, 34260
7950X3D: 4145, 62468
9700X: 4534, 37541
9900X: 4704, 54230
9950X: 4758, 67020
 
Back
Top Bottom