Intel Internal Memo Reveals that even Intel is Impressed by AMD's Progress

[Drak3]

4 hours ago, leadeater said:

Xeons for gaming have been terrible for ages, perform much worse and cost much more. Even HEDT has been questionable for gaming since X79 and it still was then anyway.

LGA 1150 and 1151 (more specifically Skylake) Xeons could easily be had for roughly 60% the cost of their Core i counterparts for a long time.

 

And HEDT, especially X99, was only really questionable on the value aspect. Now that Intel and AMD have good Hexacores on the mainstream platform, it really boils down to PCIe lane count and if quad channel RAM is beneficial for your setup, software included.

[GloriousPain]

4 hours ago, leadeater said:

And the terrible CPU clocks that make them so much worse than more appropriate options.

I'd suggest you just stop arguing with Neftex, he's incredibly stubborn. 

[CarlBar]

How your keeping you patience leadenter i don't know.

 

Look when considering performance of server CPU's lie the Xeon's forget anything related to FP64, INT64, or any smaller bit width stepping of that, (i.e. FP32, 16, 8, and Int 32, 16, 8, e.t.c.). For that kind of general purpose math it's all done on GPU's fed data by low clock, low core count CPU's, many of which have less computing power than the CPU your using to type with. But a single high end datacenter GPU will destroy multi-socket CPU configurations. CPU's get used for computation only when using very specific instructions, almost allways those working with large bit width data-sets and/or data sets too large to fit into any GPU configurations onboard memory. But given how high a bar that is it is itself a highly specialised use case.

 

Most of the rest of the use cases have the CPU interacting primarily with data storage and ram with the actual data processing again being negligible. It may sound stupid at first glance but these days the main use of a CPU in servers is to act as a communications hub between other bits of the system. Typically through PCI-E.

 

Basically you can pull up al the specific benchmarks you like, it dosen;t mean a lot because they test things that no one uses a server CPU for in the first place.

 

At the end of the day when you buy a 32 core EPYC or 28 coe Xeon you do it not for the multi-purpose computing represented in benchmarks, but it's performance using a very specific and small group of instructions. And weather the AMD or Intel offering is the best will depend entirely on the specific instruction combination because contrary to what you might think, (this took me a while to get, (and it still tends to slip my mind when talking desktop and HEDT because it's so rarely relevant)), the performance of a given CPU in one instruction is not related to it's performance in another by any kind of linear fixed factor. nd that means that not only do different architectures have different strengths in different area's compared to each other, but they also see big performance swings depending on what other instructions are being used alongside that one.

[leadeater]

17 minutes ago, Drak3 said:

LGA 1150 and 1151 (more specifically Skylake) Xeons could easily be had for roughly 60% the cost of their Core i counterparts for a long time.

Those were all still locked multipliers though weren't they? E3 Xeons were cheaper as the older generations were produced and sold for longer plus there were way more SKUs, not exactly the high end Xeons that were more being referred to though.

 

E3v6 I believe is even cheaper, MSRP is a bit lower across the range and the upper end is nearly half of v5. I've never actually purchased an E3 Xeon though.

[Drak3]

3 minutes ago, leadeater said:

Those were all still locked multipliers though weren't they?

At worst, close to lock sku Core i processors whilst being 60% the price.

 

Basically, if you weren't a ultra high FPS gamer (like most of us), a Xeon was a good option.

[leadeater]

28 minutes ago, CarlBar said:

Basically you can pull up al the specific benchmarks you like, it dosen;t mean a lot because they test things that no one uses a server CPU for in the first place.

The average CPU load on our VMware cluster is around 30%, it goes up to around 50% overnight when backups are running and we're pulling data out of databases and backing up the VMs etc but what most people don't realize is the majority of server CPUs are idle almost always. You simply do not want a CPU hitting 100% load when you are sharing the CPU resources across 10-30 VMs and need to maintain service and performance reliability. A CPU at 100% in that situation means someone is likely having service degradation and could quite legitimately complain about it.

 

The other factor is the requirement of HA so you survive hardware failures, when a server fails it's load is distributed across the cluster and that load can't cause the rest of the servers to be over utilized for the same reason as above. So generally speaking servers are only capacity utilized to around 60% maximum in terms of CPU and RAM resources. Specifically how everyone handles this and what their set limits are for utilization varies but it's pretty safe to use that 60% figure.

 

This only applies to VM hosting though, but that's like most of everything. The use cases where you actually desire 100% utilization are much more rare and are not hardware sharing platforms, like rendering or data computation (which could be scientific or non-scientific). 

[mr moose]

1 hour ago, leadeater said:

Those were all still locked multipliers though weren't they? E3 Xeons were cheaper as the older generations were produced and sold for longer plus there were way more SKUs, not exactly the high end Xeons that were more being referred to though.

 

E3v6 I believe is even cheaper, MSRP is a bit lower across the range and the upper end is nearly half of v5. I've never actually purchased an E3 Xeon though.

 

From what I remember yes they were all locked,  their price was the only reason anyone would consider them.  Getting stock i7 performance for the cost of an i5 at the expense of overclocking.  I  recommended them to friends who wanted more cores, didn't overclock yet had i5 budgets.  

[mrthuvi]

7 hours ago, mr moose said:

 

From what I remember yes they were all locked,  their price was the only reason anyone would consider them.  Getting stock i7 performance for the cost of an i5 at the expense of overclocking.  I  recommended them to friends who wanted more cores, didn't overclock yet had i5 budgets.  

Sort of, there was the exception of skylake xeon which could be OCed decently with BCLK, just like the non-k oc loophole with that generation.  

 

I don't know what happened but the last 1-2 generations the xeon was still decently cheaper BUT the board price for those has skyrocketed which makes the xeon thing not worth it. 

[spartaman64]

On 6/27/2019 at 12:40 AM, Neftex said:

youre trying to argue your point of ryzen making money by saying it was top end product and it wasnt...

 

if amd gpus had similar price/performance difference from competition like they have in cpus, their mid range gpus would sell just as well, meaning they dont need "crowns" to make money.

 

it was in terms of multicore performance. what intel cpu can compete with the 32 core threadripper?

[Neftex]

56 minutes ago, spartaman64 said:

it was in terms of multicore performance. what intel cpu can compete with the 32 core threadripper?

nice necro dude

read the rest of the thread, im not getting back to this

[spartaman64]

3 hours ago, Neftex said:

nice necro dude

read the rest of the thread, im not getting back to this

its on the 3rd page but ok

[The King of the Undead]

20 minutes ago, Schnoz said:

Honestly, I believed that Intel would be like "but we have a few more fps in these several games, amd is still slow and hot, now moving on to our 28-core 280-watt furnace"

on A side note they probably do not need a furnace.