I'm happy to see non-Intel solutions in the cloud and I'm looking forward to seeing more - AMD currently has a superior product at any price point (except for AVX2 workloads).
So? Checked lots of retailers/vendors in China, they all told me the same story - no Epyc in stock, but Intel's latest servers/processors can be shipped in hours. In fact, I ordered several Xeon processors last Sunday and they showed up on my doorstep yesterday. For Epyc, I need to pay the full price and wait for 4 weeks to get anything with the Epyc logo on it.
when searching the keyword Xeon on taobao.com, I got over 100 pages of results, from $130 second hand E5-2670 to stupidly expensive Platinum 8180. Searching on Epyc gave me a blank page with 1 and only 1 seller promising to ship in 4 weeks.
AMD is not ready for such a battle with Intel and I am not going to waste my time on such half baked products which is not even really available to consumers months after its paper release.
That's because Epyc is a new name, while Xeon is up to what, E7? I bet you're not getting so many hits for the latest Kaby Lake Xeons. And you'd get better results looking for comparable Opterons for older systems.
No, if I search for "Platinum 8180", the top model from the latest Xeon families released roughly at the same time as Epyc, I have a full page of results covering dozens of different vendors. Most of them can ship within hours, e.g. the one I used 2 days ago shipped within 30 minutes.
As comparison, if I search for Epyc, I can see 1 and only 1 vendor who is asking for the full price to be paid up front for a 4 weeks delivery.
Okay, you're getting downvotes for no reason. So Imma +1 you. I don't think you're very wrong. With that said, you're still "somewhat wrong", but not in a way that deserves downvotes IMO.
> No, if I search for "Platinum 8180", the top model from the latest Xeon families released roughly at the same time as Epyc
Xeon Platinum 8180 is Skylake, which is -2 architectures behind. There's Kaby Lake and Coffee Lake ahead of the Skylake architectures. Granted, Intel releases servers at -1 architecture compared to consumers, but... yeah.
In any case, I'm not aware of any Kaby Lake Xeons available, let alone Coffee Lake.
These are effectively i3/i5s there is no Kaby Lak SP/X (well there is technicall X but these are 2-4 core CPUs which isn’t a traditional
X class) and it’s not coming the 2018 refresh is Cascade Lake which is annoying since you’ll have Coffe, Canon and Cascade Lake all the the same time so I’m guessing it will be CFL, CNL and CSL not confusing at all.
Platinum 8180 is THE latest/fastest Xeon from Intel.
I got down-voted because many people consider AMD as an alternative to Intel, betting AMD is going to deliver real competition and innovation. My comments probably broke their heart. Not an Intel fan, I just believe that real competition will eventually be pushed by a completely different architecture.
This is great news, I hope they make it in the cloud! It will be interesting to see how this plays out... There are many optimizations for Intel CPUs in kernels and VM hosts, hopefully this will not be too big obstacle for AMD.
ROCm looks good, but so did OpenCL and I'm not very convinced that OpenCL is a good platform to build code in. ROCm addresses a lot of my concerns on paper, but there isn't much buzz on ROCm... specifically HCC.
It looks like ROCm builds off of Microsoft's C++ AMP specification (which also looks like a great idea, but that project unfortunately looks dead). So it provides a "unified" C++ source code that can compile down into x86 OR GPU Assembly (GCN in the case of AMD).
Developing OpenCL requires a bit of "copy/paste" going on whenever you start processing data on the GPU vs CPU. Because the OpenCL "world" is completely different from the host language.
It looks all good, but it doesn't seem like anyone has really tested ROCm out. There are some benchmarks over at Phoronix testing out their OpenCL implementation on ROCm (and apparently, the ROCm OpenCL is a different compiler from the older AMD stuff).
--------
Its kind of hard for me to see where AMD's "focus" is. If I had to bet, it seems like ROCm HCC has the most effort right now. In the past two years, ROCm HCC has seen multiple releases from 1.0 through 1.7, and their Github page indicates a lot of activity. Way more activity than AMD's OpenCL forums for example.
But there's no community forums of ROCm / HCC yet! I think most discussions happen in GitHub issues, which is... a subpar community solution.
AMD has been working on a lot of technologies that seem redundant: Vulkan Compute Shaders, OpenCL Compute, and now ROCm / HCC Compute. In addition, there's Microsoft C++ AMP and Microsoft DirectCompute.
NVidia has done a good job at communicating "CUDA" as the #1 platform of choice on NVidia systems. But I'm not entirely sure which technology AMD is focusing on right now. If I had to guess, they're all in on ROCm / HCC. But clarifications would be nice.
I do realize that AMD is also offering OpenCL support on ROCm. But things like Packed FP16 are better supported on ROCm HCC, and I don't think packed FP16 is supported on ROCm OpenCL yet. This suggests that AMD's #1 effort is HCC, despite the relative silence on the HCC platform.
What sort of virtualization performance and features does Epyc have? I haven't seen much good information.
EDIT: Actually, yes I have. I know that Epyc allows encryption of every VM with a key the host can't access, which is wicked. But I haven't seen benchmarks or in-depth analysis.
No.
"We asked several questions about the deployment, such as the Azure locations that will be EPYC enabled as well as which EPYC-specific security features are in use on the Azure platforms. Deployments in specific datacenters are not being discussed at this time, and the SME features of EPYC are not being used in Lv2."
https://www.anandtech.com/show/12116/amd-and-microsoft-annou...
Hooray for AMD! Breaking into cloud hosting is really what they need to do. I'm seeing them show up on bare metal box providers like Hetzner and OVH too.
30 comments
[ 1.8 ms ] story [ 70.2 ms ] threadwhen searching the keyword Xeon on taobao.com, I got over 100 pages of results, from $130 second hand E5-2670 to stupidly expensive Platinum 8180. Searching on Epyc gave me a blank page with 1 and only 1 seller promising to ship in 4 weeks.
AMD is not ready for such a battle with Intel and I am not going to waste my time on such half baked products which is not even really available to consumers months after its paper release.
posted from my new dual Xeon based workstation.
As comparison, if I search for Epyc, I can see 1 and only 1 vendor who is asking for the full price to be paid up front for a 4 weeks delivery.
> No, if I search for "Platinum 8180", the top model from the latest Xeon families released roughly at the same time as Epyc
Xeon Platinum 8180 is Skylake, which is -2 architectures behind. There's Kaby Lake and Coffee Lake ahead of the Skylake architectures. Granted, Intel releases servers at -1 architecture compared to consumers, but... yeah.
In any case, I'm not aware of any Kaby Lake Xeons available, let alone Coffee Lake.
Kaby Lake will most likely won’t appear in Xeons and the next refresh will be either Coffelake or Canonlake.
This time around there is also a big difference between Skylake S/SP and Skylake this isn’t Haswell where the only difference was the corecount.
Skylake S/SP has a different FPU structure AVX512 and SIMD CD support as well as a new core to core fabric interconnect and a few other changes.
Intel's marketing, naturally, makes knowing what their current product lines are extremely confusing.
https://en.wikichip.org/wiki/intel/microarchitectures/cascad...
I got down-voted because many people consider AMD as an alternative to Intel, betting AMD is going to deliver real competition and innovation. My comments probably broke their heart. Not an Intel fan, I just believe that real competition will eventually be pushed by a completely different architecture.
The latest Xeons naming scheme for Skylake or newer is Xeon Platinum / Xeon Gold / Xeon Silver.
Means there is finally some competition in data centers.
E5-2600v4 vs LGA3647 based processors
betting on AMD is nothing different from waiting for an open sourced version of Windows from Microsoft.
It looks like ROCm builds off of Microsoft's C++ AMP specification (which also looks like a great idea, but that project unfortunately looks dead). So it provides a "unified" C++ source code that can compile down into x86 OR GPU Assembly (GCN in the case of AMD).
Developing OpenCL requires a bit of "copy/paste" going on whenever you start processing data on the GPU vs CPU. Because the OpenCL "world" is completely different from the host language.
It looks all good, but it doesn't seem like anyone has really tested ROCm out. There are some benchmarks over at Phoronix testing out their OpenCL implementation on ROCm (and apparently, the ROCm OpenCL is a different compiler from the older AMD stuff).
--------
Its kind of hard for me to see where AMD's "focus" is. If I had to bet, it seems like ROCm HCC has the most effort right now. In the past two years, ROCm HCC has seen multiple releases from 1.0 through 1.7, and their Github page indicates a lot of activity. Way more activity than AMD's OpenCL forums for example.
But there's no community forums of ROCm / HCC yet! I think most discussions happen in GitHub issues, which is... a subpar community solution.
AMD has been working on a lot of technologies that seem redundant: Vulkan Compute Shaders, OpenCL Compute, and now ROCm / HCC Compute. In addition, there's Microsoft C++ AMP and Microsoft DirectCompute.
NVidia has done a good job at communicating "CUDA" as the #1 platform of choice on NVidia systems. But I'm not entirely sure which technology AMD is focusing on right now. If I had to guess, they're all in on ROCm / HCC. But clarifications would be nice.
I do realize that AMD is also offering OpenCL support on ROCm. But things like Packed FP16 are better supported on ROCm HCC, and I don't think packed FP16 is supported on ROCm OpenCL yet. This suggests that AMD's #1 effort is HCC, despite the relative silence on the HCC platform.
See this ROCm issue for example:
https://github.com/RadeonOpenCompute/ROCm/issues/219
EDIT: Actually, yes I have. I know that Epyc allows encryption of every VM with a key the host can't access, which is wicked. But I haven't seen benchmarks or in-depth analysis.
https://semiaccurate.com/2017/06/22/amds-epyc-major-advance-...