AMD Threadripper 9970X: What Real Users Actually Think

Threadripper 9970X perf/watt is 0.53 vs 0.45 for the 7970X — a meaningful efficiency jump. And the raw performance gap is even bigger: 156 vs 128 geometric mean across all tests.

I want more PCIe lanes more than I want more cores. More more.

Incredible performance, as expected. It's nice to see AMD keeping HEDT alive. Normal consumer CPUs have gotten so small compared to consumer GPUs they're almost funny to look at.

Those compile times are delicious.

The 9950X was a 60% uplift in these same types of workloads compared to the 7950X — that fact just got overshadowed by all the people only concerned with gaming. Intel has not been remotely competitive since Zen 3.

I guess this is why AMD has 50% of server space vs Intel and still going up. With the Threadripper 9980X and 9970X, it's pretty evident that Zen 5 brings a range of refinements that directly impact real-world workloads.

AMD just increases their prices on HEDT with no competition until it's just a little cheaper than Threadripper Pro. Especially since it would otherwise eat into workstation sales.

The real-world speed of DDR5-6400 with four CCDs is going to be very close to the maximum speed of DDR5-5600, so it probably won't make a large difference at all.

This all depends on what exactly the VMs are doing. You should consider what the requirements of all the individual applications would be (cores, memory, storage, etc) and sum them if they are running simultaneously. If you go with Threadripper, R-DIMMs are not a choice — they are mandatory and very expensive.

I like Gigabyte TRX50 AI TOP because it supports Pro and non-Pro CPUs and also 4/8 channels of memory. This way you have potential of upgrading to Pro CPUs in the future and double the size of RAM and memory channels.

As long as you are aware that you won't be able to use all 8 DIMM slots — you'll need a 9975WX or any TR Pro to use 8 DIMMs. I'd also highly recommend running 6 fans on the radiator in push/pull config — it makes a noticeable difference.

That is the wrong type of memory — you need RDIMMs which unfortunately are more expensive. 6000 MT/s is the sweet spot for Zen 5 due to the multiplier. I got 6400 just because it was not that much more.

6400 MT/s has about 14 percent more bandwidth than 5600 MT/s. That's just math. What is variable is how much of that shows up when you run real workloads — latency, timings, thermals, and other bottlenecks all play a role. But the 14 percent figure itself is just the ceiling, not something to agree or disagree with.

64GB ECC DIMMs run very hot, so you'll want an active cooling solution to push air across the RAM. Source: I sell these builds to businesses and universities for a living.

Gigabyte AI TOP is somewhat special — you can install a Pro CPU later and get full PCIe 5 lanes and 8 channel memory.

Gigabyte AI TOP runs 3 lanes at x16 PCIe 5.0 (48 from the CPU) and 1 at x16 4.0 for non-pro TR. M.2 slots use PCIe 4.0 lanes. So OP will be fine with that board for 3-4 GPUs.

R-DIMMs or Registered RAM is usually meant for servers and high end workstations. It has extra hardware to self-correct any bit flips or errors. This allows the computer to process data for days/weeks/months without stopping and not crashing due to memory issues. Better, more stable, more expensive.

You probably won't notice the difference between 5600 and 6400. Save your money on the stock 5600.

This build screams I should be on Ryzen instead. All that said, this is clearly a Ryzen build for most of these use cases.

HEDT and workstation platforms have three main selling points over consumer platforms: higher core counts, higher memory bandwidth and capacity, and copious I/O bandwidth (PCIe lanes). I've run two Threadripper systems myself — 5975WX and a 7970X. For plugin-heavy DAW work with an absurd amount of VSTs loaded from an array of 8 NVMes, I benefit tremendously.

I am kind of dialing back a bit and am close to just getting a new Ryzen setup that would support 2x GPUs at 8x PCIe 5.0. The Threadripper build costs just aren't justifiable for my use case.

Order the V-Color RAM straight from their website. I ordered from Spain and it took around 1 week straight from Taiwan — also no import taxes as they declared $100. It is almost impossible to find them in Europe otherwise.

The TRX50 boards only have 48 PCIe lanes. That's fine for two GPUs at full x16, but once you add a third card there aren't enough lanes left — the third GPU would only get x8 bandwidth, which might limit performance depending on the workload.

If there's one thing about Threadrippers that anyone can miss — the non-Pro Threadrippers only work on TRX50 motherboards. WRX90 is for Pro chips only.

These parts depreciate very slowly, even after end of life for a generation. If you find you need more cores or RAM, it's super easy to upgrade and sell your old stuff.

Latest TR has pretty similar single core performance to 13th gen Intel, and beyond compiling or exporting, most content creation software is not particularly multithreaded or heavily dependent on memory bandwidth. Sounds like a cooling issue if performance drops are happening with an i9-13900K.

Have you consulted Puget Systems Benchmarks for your workloads? Lots of 3D work enjoys a beefy 3D vCache setup with fewer cores, others see small improvements from doubling core count and memory bandwidth, and a few tests have significant overall improvements with Threadripper.

9000 non-PRO CPUs have 80 PCIe 5.0 lanes, vs 7000 non-PRO with only 48 PCIe 5.0 and 32 PCIe 4.0 lanes. That's a significant improvement for multi-GPU or multi-NVMe workstation builds.

PCIe lanes and the core counts are my primary motivation for getting a Threadripper.

Honestly you might as well have gotten a 285K if you didn't care about the memory bandwidth or the PCIe bandwidth. Save yourself the 200 watts of energy. Unless you really do need 512GB of system memory and two RTX Pro 6000.

I just built a 7980X on AI TOP board with V-Color RAM, 4090 and RTX 4000 — it has been hella stable. Make sure on initial boot you let it post for 30-45 mins if you don't get display, as initial memory training on TRX50 is much longer than normal.

I have OEM Hynix 4x48GB RDIMM with active cooling (a fan) blowing on it, and they still get up to 80C sometimes. I'm running 2 RTX 5090s. If you are using 5090s, you'll need FE or AIO versions, or the 2nd one won't fit without a riser.

I recently built almost the same thing with 7970X — same mobo, same memory. The mobo will not use PCIe 5 for the SSD unless you have a Pro series processor. I ended up needing an ASUS M.2 hyper PCIe 5 card for my SSDs.

If this is a professional situation, get your system made professionally. When you're relying on a machine like this professionally, you cannot afford downtime from compatibility issues. Companies like Falcon Northwest make sure everything is compatible, runs perfectly, and is setup out of the box.

That motherboard supports only quad channel RAM. It does not have enough physical slots to reach the maximum memory bandwidth of the platform.

If you have 8 channels DDR5 I would go the faster RAM. An 8 channel platform will have a longer life before being obsolete — might as well pull the trigger now and have more shelf life. If you're on a 4 channel platform just take the best deal.

The PCIe lane limitations have nothing to do with the chip generation — 7000 and 9000 non-Pros all have the same available lanes to use. The Pros, across both generations, also have the same 128 available Gen5 lanes. It has to do with how the board and chipset is laid out.

I'm getting 6000 MT/s rock solid stable with 4x64GB sticks rated at 6400 on the 9970X platform.

TRX50 can do up to 3 GPUs, but the third will run at x8 and you won't have much PCIe lanes for anything else. Very do-able though.

The 13900K gets laggy and choppy when multiple heavy 3D assets are open and when compiling and rendering shaders and heavy textured files, it crashes 50% of the time. I live in India and average room temperature here is 34-35 degrees. Great cooling has been provided but it just can't keep up.