Oh I didn’t see that differentiation before. But if was interested in TR, I wouldn’t mind. Who needs PCIe 5 anyway and who needs more than half the slots explicitly being >4.0? (“Oh no I can’t use 4x 400GbE NICs in my new TR”)
92 is a lot. Depending on how manufacturers design the board, it’s a slot or two less. Or they go for massive I/O and storage and put 3-4 full slots as PCIe. We never had HEDT with so many lanes…TR 3000 had 48 in total if I remember correctly. And Intel with W-2400 has way less than 92.
And if the boards ditch the SlimSAS and OCulink in favor of MCIO x8 ports, that’s a lot of fast storage on the I/O side of things. There just isn’t enough space to fit 10x M.2 slots for 40 lanes of storage. HEDT might be more tilted towards PCIe slots and less storage…we’ll see
Gigabytes board has 3 x16-slots, the rest is Storage-related.
Riser-card could reasonably fit 4, and occupy a single x16 slot.
Including an including a few Oculink to 4x SATA cables would also not take up any meaningful lanes.
hmmm… Gigabyte lists 8000MHz DDR5 for it’s X670E AORUS PRO X board though, I’m not sure if I trust their QVL, not that the memory won’t work, but that it will work at the advertised speeds.
Thats the weird thing, as far as RDIMMs are concerned, it goes 4800MHz>5600MHz>6400MHz for the RCD part of the DIMM; and pretty much all the memory vendors abide by this except the ones that load XMP profiles into their RDIMMs, then they go much higher; I don’t think I’ve ever seen a 5200MHz RDIMM, they are all faster than that save for the 4800MHz variety.
Perhaps AMD is just being conservative.
Gigabyte’s TR motherboard page was up awhile ago, it’s got 3 x16 PCIe slots and 4 m.2 slots, no slimsas/oculink/mcio connectors… all adds up to 64 exposed PCIe lanes, 44 of them pcie 5.0 and 20 pcie 4.0
Very much looks like a consumer approach. I remember TR 3000 boards being priced reasonably low. Cheap M.2 slots, no PCIe 5 slots at the edge of the board, sub-spec amount of lanes. Looks like a good deal if you don’t want to spend 800$ on a board. I’m pretty sure we will see some premium TR options that cost as much as TR Pro boards and offer more slots and different I/O connectivity.
Interesting times.
Getting 8 sticks stable is always more trouble than doing it with two of them. As far as I’m concerned, RDIMMs run at 4800 and everything else is a bonus. If 5600 works, the better.
I don’t see TR Pro having overclocking in their main audience. But TRX50 certainly shares and overlaps with consumer values and expectations. And AMD probably doesn’t want Intel to be the only one offering overclocked memory on enthusiast/HEDT platform. I remain confident about the memory…but 8000 MT/s on QVL sounds like a stretch.
Increasing success by lowering expectations and let others deal with the teething issues.
Im sad the lowest end one is a 24 core for $1500. I dont really care about the cores, I just want a platform with more PCIE on it and the extra memory channels are a nice extra too. I wish they had a 16 core for around $1000-1100. Funny how the Pro models go down to 12 core SKUs but not the regular threadrippers.
Is there any rumors of a X3D refresh as well down the line? I havent been following these at all since I thought they would be Pro only again.
It’s like those qualified 4k++ MT/s sticks on AM4 boards for chiplet CPUs. They probably can clock that high but you’re running the CPU desynced. Technically true but borderline misleading…
I’m still slightly salty I can’t get much above 6000MHz with 8 sticks on the competing HEDT platform.
looking at the gigabyte memory QVL addendums, they did list:
When running EXPO/XMP at DDR5-5200 or higher, the system’s stability may vary by AMD processor and memory module’s margin of capabilities.
When running EXPO/XMP at DDR5-6800 or higher, the memory performance gain may not be proportional due to AMD processor current architecture limitation.
sounds like someone on the qualification team was able to hit 6800MHz to realize there are problems which is a good sign.
Which is still a lot theoretical peak bandwidth compared to past generations. Intel had 6-channel of 2667 afair. DRAM still sucks, but >2x in a generation is far from what we got in the past.
If you want or need more…dual socket Genoa with 12 channels each. Sure, you have to deal with NUMA and the corresponding small latency drop. But you are very close to 1TB/s bandwidth with 4800 RDIMMs. HBM products don’t grow on trees and you never get high capacity with HBM.
Overclocking RAM at the scale of 8 channels is very new to the industry. We had some overclocked 3600 DDR4 ECC UDIMMs in the last generation. Maybe we’ll see 6800 RDIMMs at some point. Workstation market opening up to non-standard SPDs may lead to new products over time.
I’m personally salty about TRX50 not having 6-channel memory. They have this on Siena, yet TR is stuck with 4-channel. Market segmentation is a bitch. 16 cores is really stretching dual-channel and TR with 64 cores just gets 4 channels? and has to deal with 2DPC if >256GB cutting down clocks in the process, possibly even with RDIMMs as seen in EPYC 2DPC specs.
EDIT: On second thought, the if this photo (from Computex) is true, the ASMedia implementation is also mindbogglingly stupid… as despite being fed with 4 lanes of PCIe 4.0 it only allocates 2 lanes of PCIe 3.0 per port. LOL
The official product page, sadly, does not clarify. It’s commonly known that many USB4/Thunderbolt 4 devices claim 40 gbps connectivity but is actually capped at a fraction of that, and I wonder if it’s also true of these claims.
comsol multiphysics, it is so memory starved that a 16 core intel hedt with the decent memory bandwidth will perform at 160% the speed of a 64 core threadripper pro… I have thousands of compute-hours of simulation ahead of me.
I am very tempted by this… if only it wasn’t so expensive.
If I ever do go this route I think I would create my own RAM water cooled heatspreaders, buying 24 heatspreaders would be like 1k USD if purchased commercially.
I kind of wish the dual socket workstation platforms would come back, dual socket overclocking was interesting, à la evga sr2. If AMD ever released a dual socket threadripper with memory OC support I would immediately buy it.
vcolor just started selling 6800MHz RDIMMs, they are only available in single ranked configurations though.
With a good enough shroud (3d printing maybe?) I think this is doable, but even the server guys with their 10k screaming fans are supposedly going to go water cooled for DIMMs starting this generation and next generation more broadly because they can’t keep the memory temperatures down under a stressful workload.
Each DDR5 RDIMM can put out ~20 watts when running fully pegged.
32GB 5600MHz DDR5 RDIMMs can be had for ~80USD a piece now… but yeah the entire system gets very expensive very quickly, mostly the fault of the CPUs themselves, the workstation-y F-sku genoas start at 3k USD each.
I was all hyped up for the new Threadripper Pro, but I’m kind of disappointed by it having just 8-channel memory and being only $500 cheaper than the Epyc 9654.
Since I’m into those memory-hungry apps, there’s no way I’m going for the 7995WX. I’m going to get the Epyc 9654 instead, because it gives about 50% more memory bandwidth.
Asrock and Asus just showed off their TRX50 and WRX90 motherboards, it doesn’t look like 2DPC is going to be a thing for this platform considering Gigabyte, Asus and Asrock boards are all 1DPC.
SPR is only officially rated for 4800MT/s below the capacity limit (2TB for Wx-2400 and 4TB for wx-3400) where they are further limited to 4400MT/s. Since both platforms support DRAM overclocking it literally doesn’t matter what the base support rating is.
Assuming their overclocking potential is the same this is true, but this example highlights that intel has been conservative with advertised memory clocks as of late while AMD hasn’t; between AMD juicing vSOC to support higher ram speeds on 7000 series in early BIOSes and the 2DPC Genoa issues I’m slightly worried.
Perhaps this is a moot point, Asrock is claiming their TRX50 board will do a 7600MHz memory OC now… hopefully that isn’t some kind of golden CPU IMC example.
