The Xeon w-3400 and w-2400 series CPUs were announced months ago, and the w-2400 series has been available for over a month now to order in pre-built systems and as components (at least in the UK). And some of the w-3400 series CPUs are available for immediate delivery. Yet all I can find in terms of actual reviews are:
Linus cooing over an HP workstation built around the platform (without benchmarks);
der8auer overclocking one in an Intel lab on favourable benchmarks;
an early set of Phoronix benchmarks which were a bit meh;
KitGuru doing an actual review and not being that impressed.
I understand that there is some kind of unusual multi-stage NDA in place, but I would have thought that once the components are on general sale there is nothing to hold the reviewers back (apart from funds).
Well, the products were announced, but they’re not available except to a few select reviewers for marketing purposes. I think we’re expected to see availability in the second half of 2023.
I’m supposed to have a w5-3435x in my hands tomorrow… only thing that will be missing is memory of which the high quality stuff has not hit the channels yet.
atm it appears benchmarking on windows is not advised due to scheduler/clock ramp issues, linux should be used if you want to see representative results, and it handily beats zen3.
There’s a thread on the forum where members have received theres:
One of the renowned Japanese reviewers (Jisakuhibi) has been reviewing 2455X and 2495X with gaming workload. Long story short, in gaming workload, and when overclocking to 5.0GHz, 2455X can archive similar performance to 12900K, and 2495X can archive performance somewhere between 12900K and 13900K, both at the expense of higher power usage than their consumer counterpart.
The reviewer did omit gaming tests from 3400X benchmarks, though.
I suspect part of the delay in getting reviews is how difficult benchmarking it fairly will be. Linus’ staff lacks the competence for that, and der8auer probably only got anything because Intel did the benchmarks for him. Phoronix probably could do a decent review, but would be hindered by gcc not yet having tune tables for sapphire rapids. In raw performance, EPYC should beat it. But if you compile your workload around the accelerators, which the businesses interested in these things would probably do, it should be a very strong product both in raw performance and efficiency.
Yes I have, I’ve been a little slow on the setup because there still aren’t any reasonably sized and fast memory kits available for it yet.
But running on 1 stick of 4800MHz memory it performs exactly like you’d expect a ~4.7GHz golden cove (alderlake) to perform:
I’ll probably be doing some Solidworks benchmarks in the other thread soon, I’m expecting it to be considerably faster than a Threadripper PRO 5995WX for model rebuilds, between 30-100% faster.
Once I get memory I want to run Clear linux on this, chances are Intel’s Linux distro is going to have all the scheduling nuances sorted out.
Certainly it would be possible. Easy, even, IIRC Sapphire Rapids is SoC and doesn’t need an external chipset. You can just solder the CPU up to some power and a serial output and use it as a ridiculously overspecced microcontroller. Give it some PCIe lanes and it could be an actual computer.
You’re wasting a tremendous amount of resources though. You’ll struggle to fit more than two or three channels of memory, even if you vertically stack SODIMMs, and as consumer platform ITX boards have shown, it’s tricky to get more than 32 PCIe lanes out of it (one x16 slot and three x4 M.2s is typically the most you can expect). You could probably do PCIe over miniSAS or USB-C, but then you’ll need to also make adapter cards, and you’ll end up not actually saving any room over just making a conventional SSI-EEB board.
Ultimately, Sapphire Rapids with only two or three channels of memory available is going to be less effective horsepower than a 13900K or a 7950X, because those two will happily overclock and get more bandwidth out of their memory than SR could. AM5 and LGA1700 both support ECC memory on the right motherboards (for AM5, any Asrock or the Asus ProArt and for LGA1700 any board with a W chipset). The extra cores and lanes offered by SR becomes useless when you’ll have a hard time physically fitting them on an ITX board, and without loads of memory bandwidth your cores will spend most of the time idling.
Six channels on ITX isn’t unheard of, it really just depends on how creative the manufacturer is willing to get.
Plus, plenty of programs will run fine out of caches; memory bandwidth is seldom the bottleneck outside multi-hundred gigabyte datasets (something like Ansys).
There is definitely a use case for compact systems with a lot of CPU horsepower, although an mATX sized board would provide a much better experience.
That’s the most grotesque motherboard I’ve ever seen, thank you very much for linking it. My preferred approach would have been to just lengthen ITX beyond spec to fit SODIMMs on the sides of the processor. I saw boards called “deep ITX” that did this a while back, though in this case the purpose was to fit four channels of full-size DIMMs on ITX. Wouldn’t be great for front to back airflow, but if you wanted that you might as well just make it a server blade instead.
Re cache, I guess you’re right. My primary use case for a workstation CPU is CFDs, where datasets in the hundreds of gigabytes aren’t uncommon, so bandwidth becomes a higher priority than cache. The use case cited was a TV media station, so a single memory channel would be more than sufficient, it just rankles me to stick such a powerful CPU on such a limited board.
If you wanted to make a point of it you could use 3D-cache EPYC instead, and just eschew memory altogether. Those things will happily fit a light Linux distro plus media player in cache.
Intel just released a presser on how the their CPUs outperform AMD in performance and power consumption, ahead of AMD’s big event tomorrow.
I want to see what AMD has to counter with because these claims seem too good to be true.
Interesting comparisons. Will give Intel benefit of the doubt for now.
Look at Max 9480…the HBM is effectively the performance booster I guess.
The 32 core Xeon are boosted by ASIC accelerations in quite a few benchmarks. Otherwise, likely worse than EPYC 9354 (same 32 cores). What’s the catch of Intel accelerator?
BTW, how are Max 9480 sold with HBM? The spec says it’s up to maximum 64GB HBM. There are four stacks on the package. Are they going to sell one stack, two stacks…up to four stacks. Or in some other formation?
HBM is a god send for CPUs, they have been so memory bound the last 15-20 years.
All the HBM variants of xeon are full stack variants, all with 64GB of HBM. Supposedly AMD is going to announce the MI300c tomorrow to compete with the Xeon MAXs.
I think Intel is referring to applications compiled with some of the new instructions sets like AMX or AVX512.
There are a couple of hardware accelerators probably being lumped into this too, DSA and QAT being the two important ones.
Someone on this forum mentioned that ZFS could take advantage of QAT and I still need to look into that because that is very interesting.
Yeah I think that’s a pretty good synopsis for the non-HBM parts.
I’m disappointed the MI300C was absent from today’s launch. Also kind of disappointing that the MI300A has shorter HBM stacks than MI300X, but the fact that the MI300’s can be multi-socket make up for all of this.
Should probably make a separate thread for this, these MI300’s are going to be game changers.
