Hey everyone! Longtime Topaz user here and up until now, a dedicated Mac one at that. But I’m hitting a point now where it’s becoming clear that in order to fully get everything I want out of the latest versions of Topaz Video, a PC is truly what I need. I’ve reviewed benchmarks here many times over and while that’s great to give me an idea of what people’s individual rigs are capable of, I’m looking to build my PC from the ground up. I already have a large 15" cube ATX case with 6 fans, I just need to pick out the components. I know users here have talked extensively in the past at a few points about how Topaz taps into different areas of the hardware and it’s clear that a good processor and graphics card are needed, along with plenty of RAM. As I look to embark on this in the next few months, I’m currently looking at going with an Intel processor vs an AMD one because 1) Intel’s Core Ultra 7 and 9 series look like they give serious performance while undercutting the price of AMD at the same level by quite a bit, and 2) As a Mac user, I have a lot of external drives and hardware that are Thunderbolt capable, so going with Intel seems a no-brainer. What have people’s experiences been with Intel vs AMD processors for the latest Topaz Video releases, especially now that the Starlight models have taken center stage? I’m looking at going with an NVIDIA GPU because clearly Topaz favors them vs AMD overall. But aside from all of that, any other PC builders here who can recommend other great components (motherboards, RAM, power supplies, cooling systems) that they love and have found work great for their setups? Thanks so much!
| Argument | My take |
|---|---|
| “Topaz requires a good CPU” | True, but secondary |
| “Intel is better for Topaz” | Not substantiated |
| “Core Ultra 7/9 is significantly cheaper than equivalent AMD CPUs” | Highly model/price dependent |
| “Intel is preferable because of Thunderbolt” | Yes, this is a valid argument |
| “NVIDIA is preferable for Topaz/Starlight” | Yes, especially for Starlight |
| “A lot of RAM is useful” | Yes — 64 GB makes more sense than 32 GB for a high-end system |
| “The GPU is crucial” | Absolutely — this is the key component |
Considering the only aspect intel > AMD is memory bandwidth, thus :
| CPU | Memory | Channels | Max. official speed | Theoretical bandwidth |
|---|---|---|---|---|
| Intel Core Ultra 9 285K | DDR5 | 2 | 6400 MT/s | 102.4 GB/s |
| AMD Ryzen 9 9950X | DDR5 | 2 | 6000 MT/s | 96.0 GB/s |
| AMD Threadripper 3900X series | DDR4 | 4 | 3200 MT/s | 102.4 GB/s |
| AMD Threadripper PRO 3900WX series | DDR4 | 8 | 3200 MT/s | 204.8 GB/s |
| AMD Threadripper 5900X series | DDR4 | 4 | 3200 MT/s | 102.4 GB/s |
| AMD Threadripper PRO 5900WX series | DDR4 | 8 | 3200 MT/s | 204.8 GB/s |
I use a 3995WX. It absolutely destroys any high-end desktop setup, by far.
And forget multi-GPU. Load balancing is atrocious, except for a few models that are already obsolete.
Threadripper 5000 allows some memory overclocking, whereas the PRO models are locked to their validated memory settings.
Intel’s E-cores are not “useless”, but on a serious workstation I’d much rather have full-performance cores than a mix of P- and E-cores. AMD’s homogeneous CPU designs avoid the scheduling and performance asymmetry of Intel’s hybrid architecture.
About graphics cards, there are three brands I would not buy:
Gainward, PNY and INNO3D, because they often save more than others on components they use on the PCB.
Regarding the PSU, if 1200W is sufficient, take a look at the Enermax PlatiGemini EGN1200P, there are fair market prices
…and if you don’t need bling-bling fans, the Arctic P12/P14 are good. You can get a value pack with 5pcs for about 25$
That’s a fantastic breakdown and very useful information! I was looking at Tom’s Hardware 2026 reviews of CPUs and from some of the benchmarks there, it seemed the Intel Core Ultra 7/9 series was pumping our serious performance while being much cheaper than their AMD counterparts. But the information and experience you’ve provided is exactly what I was looking for, and if Thunderbolt compatibility ends up not being something I really care that much about, it might just tip me back towards considering AMD for the processor rather than Intel.
Many thanks for that advice! I think I’m leaning towards an Asus ROG NVIDIA card at this point. The case I have already has 6 built-in fans so no need to buy any of those right now. I have no idea what PSU I’m going to go with because I don’t even know what wattage I really will need yet. 850W and 1000W seem to be good enough in general, but I hadn’t really considered going higher.
At around 50% load, power supplies generally operate at their highest efficiency point. A lot of people oversize their PSU, which isn’t optimal either and also makes the PSU more expensive. On the other hand, running a PSU close to its maximum load isn’t ideal either.
My system, with an RTX 5090 undervolted to 850 mV and an AMD Ryzen 9 9900X with turbo-boost disabled (Windows power settings cpu limited to maximum 99% usage), draws between 550W to 650 W during all SLP renderings, measured at the wall. That puts my 1200 W PSU pretty much right in the sweet spot, air coming out of the PSU isn’t even lukewarm.
So, depending on which GPU and CPU you choose, I’d recommend using realistic power figures and adding them up roughly and I recommend calc with 50-60% PSU load, and then you still have air up for future upgrades.
Absolutely right and very good advice
With one minor caveat regarding transient spikes, which generate brief power surges of 1.5x to 2x the nominal consumption, somewhat justifying the oversizing.
I use a RTX 5000 Pro 48 GB.
I use both Photo and Gigapixel at the same time with Neuroserver Models - hitting 35 GB Vram at max 335 watts.
The GPU does not care if if do game at the same time (Gray Zone Warfare at 120fps) or just keep editing images.
If that helps with the decision.
Yeah, exactly. But this is where good PSUs come into play: these spikes only last for milliseconds, and cheap power supplies can sometimes have trouble dealing with them.
But if you ignore the spikes and calculate the power draw based on the normal values without those extreme transient spikes, then size the PSU so it runs at around 50% load, those spikes are guaranteed not to be a problem - unless the power supply uses really cheap components
My Idea is use a Stresstest (burn-in) Tool which drives cpu+gpu to 100%, and then measure power draw. The idea is you take this value, double it and then you have wattage of your PSU to buy and you are safe in any cases and air for hardware upgrades. SSDs, HDs etc can be ignored, because consumption is relativly low - in reality Topaz Video is a bit lower than the max power draw from the stresstest.
I recommend this setup-calc for heavy load PCs - longterm renderings as we do with topaz slm slp etc models. Normal PC users can also choose a bit smaller power supply, as they rarely run both the CPU and GPU at full load for extended periods.
But people who say they don’t have gpu+cpu longterm full loads, for ecample because Gigapixel is the tool, and it is not planed using others that did that, then yes you can go with a smaller PSU.
This is actually the first time I’ve gone Intel, and only because Threadripper CPUs and motherboards have gotten extremely expensive, so I wanted to share my 2025 build:
I noticed you’re considering a Threadripper workstation, so I wanted to share something relevant. The issue shown in the video I posted affects both AMD and Intel workstations, not just Threadripper. The thread on the LT forums is mostly AMD users (so the community support for AMD workstations feels large compared to Intel), but I had the exact same symptoms on my Intel workstation. In my case, the root cause ended up being faulty ECC RAM, and the vendor replaced it with almost no questions — which makes me think there may be a known issue with certain DDR5 ECC modules.
My thoughts on my current Intel build:
-
E‑cores: Surprisingly useful. Switching to “performance mode” pushes heavy AVX workloads onto the E‑cores, which keeps heat down when P‑core performance isn’t needed.
-
ASUS Pro WS W680‑ACE: Much cheaper than the AMD equivalent, and prices have dropped even more since I built the system.
-
ECC + non‑ECC support: Being able to swap to non‑ECC made troubleshooting way easier.
-
No overclocking: This board can’t overclock — first system I’ve run at stock settings.
-
Intel iGPU (QSV/Quicksync): I use this way more than expected. Great for fast H.264/HEVC exports. Topaz can use it too, though I mostly export to ProRes.
I had a lot of issues early on (as shown in the video), and it took almost a year to fully resolve. But once the ECC RAM was replaced, the system stabilized immediately. and based on Amazon reviews from this RAM vendor I am not the only one that had quality issues:
However I didn’t have a problem with their customer service as they replaced my DDR5 ECC quickly with very little questions asked.. so who knows… I haven’t used this RAM company that much and this is only my 3rd or 4th set of ECC RAM from them.
As for GPU:
I’m currently shopping for a replacement specifically for Starlight Precise performance. Right now it looks like nearly any NVIDIA 4000/5000‑series card performs better than AMD for this model. The price/performance balance is strange — even something like a 5060 Ti might perform poorly on most models but still do reasonably well on Starlight Precise. Hard to confirm because lower‑end user reports are rare.
Nothing to do with Intel or AMD: DDR4 and even DDR5 run hotter than ever.
You really need active cooling on them as soon as you load them intensively, which is typically the case with Topaz Video.
- The PMIC adds even more to the problem
- And XMP/EXPO voltages at 1.35-1.45V instead of JEDEC’s 1.1V add even more heat
Below: running like a charm 24/7 with brutal Delta fans on — and only under that condition.
| Topaz workload (2x32 GB) | AI workload (8x16 GB) |
|---|---|
| SK hynix A-die (best) 1.425V | Micron R-die (value) 1.35V |
| 38°C max - Delta SH 120mm on it | unknown °C — 2x Delta SH 80mm on it |
Workstations don’t stay in the server room anymore. They invite themselves onto our desks.
Problem is, they were designed for where they used to live: a datacenter-grade wind tunnel, where nobody cared how loud the cooling got.
DDR4
| Die | Node (nm) | Daily VDD | Bench ceiling | Voltage scaling |
|---|---|---|---|---|
| Samsung B-die | 20 | 1.50 V | 1.55 V | Excellent — holds 1.5 V 24/7 |
| Samsung C/D/E-die | 10 | 1.40 V | 1.45 V | Poor, gains flatten out fast |
| SK Hynix MFR | 25 | 1.50 V | 1.60 V | Only starts to scale past 1.48 V |
| SK Hynix AFR / CJR | 21/18 | 1.45 V | 1.48 V | Decent |
| SK Hynix DJR | 17 | 1.45 V | 1.50 V | Decent |
| Micron Rev.E | 16/18 | 1.45 V | 1.50 V+ | Good, but heavily disputed |
| Micron Rev.B (16 Gb) | 18 | 1.45 V | 1.55 V | Good |
| Micron R-die | 14 | 1.40 V | 1.45 V | Weak, can scale negatively |
DDR5
| Die | Daily VDD/VDDQ | Bench ceiling | Notes |
|---|---|---|---|
| SK Hynix A-die | 1.40 V | 1.45 V | Best voltage scaling in DDR5 |
| SK Hynix M-die | 1.40 V | 1.45 V | Close to A-die, one notch below |
| Micron B-die | 1.35 V | 1.40 V | Weak scaling, don’t bother pushing |
| Samsung | 1.35 V | 1.40 V | Vendor absolute max is 1.40 V |
Reliability and overclocking headroom are two different things, and people keep mixing them up.
Lowest RMA rates: Crucial (Micron), Kingston, Lexar — all under 0.3%.
Best binning if you actually want to tune: G.Skill, Kingston Fury. Note the irony though — G.Skill sits at 1.48% RMA, above Corsair’s 1.12%. Tuner-friendly doesn’t mean trouble-free.
The real industry-wide drama nobody talks about: same part number, different dies depending on the batch. Every brand does it. Corsair just makes it worse by gluing non-removable heatspreaders onto proprietary PCBs, so you can’t even check what you bought without dumping the SPD.
Die lottery, buy at your own risk: Netac, KingBank, Gloway — great value when they land Hynix, no guarantee they will.
Bottom line: buy the die, not the brand. Thaiphoon Burner before you commit, or accept the lottery.
And the reason the tuner-friendly brands return more often isn’t bad build quality: it’s shipping 1.40V XMP profiles under an RGB heatspreader with no airflow planned for.
Got an Intel’s 270k+, MSI PRO Z890-S WIFI, 64 GB G.Skill Flare X5 F5-6000J2836G32GX2-FX5 CL28 on the way. Already have an Asus 5070ti tuf gaming oc and 2 2 tb 9100 pro’s and 1 4tb 990.
Thunderbolt 4 and hopefully 5 when MSI releases the add on card.
Just the ram was 1000 euro but the rest was a steel for it’s price i think. 200 for MB and 289 for CPU. And ram is now back to 2k.
I don’t think there is one single best recommendation, especially since hardware prices have gotten quite expensive in some areas. If you find a good deal, go for it.
I went with an AMD 9900X myself and haven’t regretted the purchase. 12-cores don’t seem like a bad choice. I don’t want to use liquid cooling, the CPU can still be cooled quite well with a good air cooler, more problematic are 170W TDP cpus
My philosophy is that if a computer is running 24h a day, it should run as reliably and efficiently as possible. That’s why I set the maximum CPU usage to 99% in the Windows power settings, so the CPU can’t enter turbo mode. I also limited my RTX 5090 to 850mV.
Depends on what you do with it. That Z890 has 8 lanes of pci 4 chipset to the CPU and AM5 only has 4. And also takes away 4 lanes for USB4. And the Z890 can do a 2 8 lanes pci 5 split on more expensive board in a future buy. The socket may be over but i can go for a second gpu on a more pro bord. I also have a PLX pci 3 switch to 4 m2 slots for models and KVCache on raid 0 mlc ssd’s.
Intel had to catch up hard on AMD flagships but if you want more pci lanes you have to go Threadripper and that is crazy expensive. But if you happy that is most importance. I took me almost 1,5 year to understand new PC’s and i still don’t get most hahaha.
Waiting was not an option anymore so i compared bundles sets and knew Intel were undercutting the price so it was now or never but those bundles were the same price as buying as parts so i bought from different suppliers to be cheaper and i hope still good.
Anyway thanks for that Artic fan tip. I got 2 5 pack’s 120 and 140 mm but i got the pro’s and also bought 4 360 degrees 120 to 140 mm for 9 euro each on Aliexpress because i also need ram fans and these look like will fit with my Noctua D15 cooler. I also want no water cooling.
Anyway those 360 degrees fan frames are also handy for my thunderbolt 4 NVME/Wavelink 10 Gb on my m4. So i bought 4 of those 360 mounts.
) now I can adjust rpm of the slot coolers
Thanks but doesn’t that also blocks the airflow going out from the front fans? I’m going to try to put a 120 mm fan on the outside that pulls the hot air away like yours but more like a Mac Pro 1,1 where the fans in front pushing the cool wind in and the fan at the end sucks the hot air out and looks more like a tunnel.
I got a Fractal Design North - TG Clear Chalk White and maybe it fits but i think I should have got the mesh version instead of the glass version and use 2 140mm from the side sucking hot air through the mesh.
It depends on how your airflow is set up. In my case, I have four front fans pushing air into the case, plus two fans on the top and one at the rear.
Ideally, all the air coming in from the front should flow through the entire case and then be exhausted at the back. The problem is that I only have one rear exhaust fan. so i need more at the back and that’s why i use slot fans.
If roof fans pushes out, this create an airflow from bottom to top, so the airflow from front to back and bottom to top are crossing each other, which isn’t ideal and if both top fans exhaust air out of the case, they pull away fresh air coming in from the front fans.
If one top fan is intake and the other is exhaust, there’s a risk that the air you’re just bringing into the case gets exhausted again from the other roof fan.
Ideally, you want one consistent airflow direction. But because of limited space, I’m using a mid-tower case, this is a problem having one rear exhaust fan.
So for now, I’ve set it up like this: the top fans run very slowly, while the front fans push a good amount of air into the case and the rear run fast pulling out. The two PCI slot fans adds extra airflow, which helps move the air from the front to the back trough the entire case.
I will probably change the top fans to a push-pull setup, but I’ll add a plastic divider between them to prevent a kind of “short circuit” in the airflow. That way, my CPU cooler can get fresh air from the top instead of pulling away air coming from the front, that is supposed to flow past the GPU.
P.S. When you have the Airflow from Back to the Front, then Slot Coolers are not a good Idea!
It sounds like you’ve built a very complex workaround for problems that don’t really exist anymore.
Running a PCIe 3.0 PLX switch in RAID 0 with old MLC SSDs for LLMs and KV Cache in 2026 is like putting dual carburetors on an electric car—a single modern PCIe 4.0/5.0 NVMe completely destroys that bandwidth without the latency or complexity.
Same for the Z890 split: x8/x8 consumer multi-GPU for AI workloads is heavily bottlenecked compared to proper enterprise interconnects.
Bragging about Intel’s DMI 4.0 x8 chipset link for AI is missing the point entirely.
AM5 delivers 28 direct PCIe lanes from the CPU via Infinity Fabric for your GPU and main NVMe drives—they never touch the chipset link. DMI bandwidth only matters for secondary chipset I/O.
Unless your LLM runs on a drive multiplexed through the chipset while you saturate three SATA ports, that DMI x8 gives you strictly 0% extra performance for inference.
If you enjoy the hardware Frankenstein/DIY aspect with Aliexpress 360-degree fan brackets under a Noctua D15, go for it! But purely from a performance, latency, and throughput standpoint for LLMs, you’re overcomplicating things to match specs that an off-the-shelf single NVMe and a clean airflow setup beat out of the box.
Thermalright Phantom Spirit 120 + 2x Delta AFB1212SHE is all you need to blow AIOs.
Did you actually buy G.Skill Flare X5 for an Intel Z890 board? Flare X5 is an AMD EXPO-tuned kit. Running EXPO profiles on a LGA1851 platform is asking for stability nightmares—you should have bought XMP (Ripjaws/Trident Z).
Also, claiming you paid €1,000 for a 64GB DDR5 kit that is now worth €2,000 is pure delusion. No standard 64GB 6000MHz CL28 kit costs as much as a high-end RTX GPU.
You’re pairing top-tier Samsung 9100 Pro SSDs with the MSI PRO Z890-S, which is literally MSI’s bottom-tier budget board, while dreaming about Thunderbolt 5 add-on cards that won’t even have the proper PCIe lane allocation on a €200 motherboard.
You really need to stop buying random parts before checking basic specs and RAM compatibility list (QVL).
