AMD EPYC 7251 Provides Great Value At Less Than $500 USD

[pegasus]

I beg to differ with the "Great Value" claim; both Xeons mentioned, the Silver 4108 and the Silver 4112 have AVX-512 support, Epyc cpu's do not, in many floating point heavy applications, such as those in HPC workloads like various simulation type workloads, the Xeons will simply walk away from Epyc's.

This is true as long as your dataset fits into cpu cache. And that is seldom the case. And even when it is, such workloads are better handled on GPUs.

In real life, datasets barely fit into available memory so it is the speed at which you can feed your compute units that defines the speed of the whole system. And here 8 memory channels of epyc per socket give you advantage over skylake. That NPB LU.C test is a nice demonstration of that.

[pegasus]

I beg to differ with the "Great Value" claim; both Xeons mentioned, the Silver 4108 and the Silver 4112 have AVX-512 support, Epyc cpu's do not, in many floating point heavy applications, such as those in HPC workloads like various simulation type workloads, the Xeons will simply walk away from Epyc's.

True, as long as your dataset fits into cpu cache. Then avx512 can show some muscle, if your cooling is designed properly and cpu doesn't downclock too much

But if this is not the case, avx512 units on skylake get starved waiting for data and then epyc can jump over skylake, since it is able to feed its avx2 units with more data.

I know very few cases that fit comfortably in cpu cache. And those are better handled by gpus nowdays anyway.

[itoffshore]

Would it be possible to setup Ryzen CPUs (Threadripper probably) with multiple GPUs and virtualize gaming desktops there for an entire family? I would put the machine in a rack, provide each family member with 16GB RAM + 4 cores and dedicated GPU. Is it a viable solution? How the video from the VM would be streamed to cut on the latency?

Have a look at the Debian Wiki for KVM VGA passthrough

& also see the Level 1 Tech video guides.

[Wielkie G]

I don't know if it's just because the frequencies are different, but I find it odd how Epyc seems to be so much better than Xeon, whereas Ryzen is underwhelming compared to the Core i series.

Michael: Might be interesting to see an IPC test of Ryzen vs Epyc, with the 8700K and the Xeon Silver included in the mix too.

Two possibilities:
1. With the Xeon line, Intel doesn't have the huge clockspeed lead over AMD. Most of the reason why 7700k or 8700k seem to be "so great" is because they are clocked right near their silicon limit, 15-20% over their Ryzen competition.
2. SkyLake-X vs SkyLake. AVX-512 notwithstanding, having bigger slower L2 cache and mighty slower, smaller L3 cache might actually cause performance regressions in certain workloads.

[Spooktra]

I beg to differ with the "Great Value" claim; both Xeons mentioned, the Silver 4108 and the Silver 4112 have AVX-512 support, Epyc cpu's do not, in many floating point heavy applications, such as those in HPC workloads like various simulation type workloads, the Xeons will simply walk away from Epyc's.

So long as AMD continues to make processors based on a competitors ISA technology, they will continue to be in second place in market share, revenue and performance.

Perhaps AMD should have focused on their ARM based Opteron line, grow that market, as it seems that ARM is the future, hell it's even rumored that Apple is developing their own ARM based cpu to replace Intel's processors in the Apple desktops.

[Brane215]

Meh.Using Epyc for CPU muscle alone doesn't make much sense. It's low frequency part that doesn't have that much to offer if anyone doesn't need extra PCIe lanes, ECC, 2-socket suport and quad-channel RAM. And those extra RAM channels count for something only if one needs ECC and register, otherwise they might not be such a bandwidth upgrade over two smokin fast DDR4 channels.

These things are for apps that absolutely must have an access to big amounts of RAM and need two sockets to do that, possibly through several boards, interconnected with Infiniband or similar glue.

For one onbard/one socket solution, Threadripper would smoke it. 1950x is twice the price of 7251, but that will diminish once you factor in price of board + RAM.
And at that ptice, 1950x offers almost four times of CPU muscle -ie 2x cores at almost 2x speed...

[Michael]

"The EPYC 7251 is a 2P-capable part priced ...."

Is it supposed to be obvious what '2p-capable' means .. ????

Can be used with Epyc dual socket motherboards if wanting to run two of them; not all EPYC parts are able to be used in a dual socket configuration (the 1P parts).

[orzel]

"The EPYC 7251 is a 2P-capable part priced ...."

Is it supposed to be obvious what '2p-capable' means .. ????

[thelongdivider]

Wow. EPYC is such a robust platform. 8 channel memory, 128 PCI-E 3 memory lanes, and really excellent processing power to boot. If I had the 32 core variant I'd be set for 10 years.

[schmidtbag]

Would it be possible to setup Ryzen CPUs (Threadripper probably) with multiple GPUs and virtualize gaming desktops there for an entire family? I would put the machine in a rack, provide each family member with 16GB RAM + 4 cores and dedicated GPU. Is it a viable solution?

Yes, this is doable and if done properly it would work pretty well. Multi-seat setups have been around for a while and there are dozens of guides for them, but they're a real PITA to set up. If this is supposed to act as a family PC though, I think 16GB per person is seriously overkill. Even 8GB would be plenty, unless you intend to do a lot of modern gaming. If you don't use a VM (it is possible to do multi-seat in Linux without VMs), you could definitely get by with 8GB per user, since you'll have a lot of shared memory.

How the video from the VM would be streamed to cut on the latency?

The best route is to do GPU-passthrough (meaning IOMMU needs to be enabled from the BIOS), where you're detaching the GPU from the host OS and making it only visible/usable to the guest. That implies you will either need a GPU for the host, or, the host will need to be accessed via SSH in order to maintain it while guests are still running. I don't think most multiseat setups run a seat on the host. Anyway, since each user gets their own direct access to a GPU, latency is hardly a problem. But if you care about things like VR or playing games like CS:GO, I'd advise against multi-seat.

I know Xen has proven to work in the way you're imagining. Might want to look into that.