CPU prices starting to fall ( for some processors )

[tServo]

Checking Amazon and Newegg ( yes, I know about Newegg's legal problems ), I have noticed that the prices on some processors I am interested in have been going down:
AMD's Ryzen 1700x & 1800x look like they have been reduced by 100 US dollars and the Threadripper has gone down by about 120 -> 130 dollars.
Intel's have gone down just a little:
The I7-7820x and I9-7900x have only been reduced by 20-30 dollars.
I'm planning a build with a I7-7820x. After I get it going and the build debugged,
I'm going to run some tests to examine thermal throttling.
Then I'm going to send it to a guy who'll delid the processor
and replace the TIM with solder-type material.
Then, I'll run the tests again. I plan to use a 280mm AIO cooler.
My next build will be a Threadripper.

[ewmayer]

Originally Posted by tServo

Then I'm going to send it to a guy who'll delid the processor and replace the TIM with solder-type material.

If that is a by-mail-for-fee service any PC owner can take advantage of, could you post a link? Have you used said service before?

[mackerel]

On CPU pricing, I have seen many more offers on Ryzen products since Coffee Lake was "released". Never mind you can hardly buy CL except at significantly elevated pricing. Intel felt they had to respond, and AMD likewise responded to that. I'm kinda hoping for but not expecting a drop on 7700k too. It'll make a minor upgrade to my main system and I can shuffle out an i3 from another.

On Skylake-X delid, if you don't mind doing it yourself, the tools are not that expensive. In the US I'd suggest the Rockit 99 for $40 (+ whatever liquid metal TIM you prefer) https://rockitcool.myshopify.com/pro...akex-kabylakex

I bought their Rockit 88 for 115x CPUs in the past and it has served well, but the international shipping + custom fees meant it cost a LOT more by the time it got to me in UK.

This time around, I'm getting De8auer's delid die mate X, which is more expensive and over-engineered, but at least it is "local" so would be faster shipping, without customs.

BTW my Skylake-X 7800X seems P95 stable at stock voltage and all cores at 4.3 GHz, and that'll probably remain my standard overclock. The delid will allow me to explore what sort of clocks I can get for non-AVX workloads with elevated voltage. I have benchmarked at 4.9 GHz but that was into the thermal danger zone.

[tServo]

Originally Posted by ewmayer

If that is a by-mail-for-fee service any PC owner can take advantage of, could you post a link? Have you used said service before?

It's on Ebay, so anyone can apply:https://www.ebay.com/itm/Intel-CPU-D...oAAOSwhqhZ5kzD

In case the link is bad, the seller's name is ac_gaming and the auction number is 222681839495.

No, I haven't used his services yet but I checked his feedback. He has had a socket 1151 delid service for a while and I saw about 25 VERY satisfied customers. 3 of them mentioned specific drops of 20-25 degrees !
The most recent 3 of his feedback pages have their reviews, in case you want to check.
He says he tests the processor before shipping it back and he guarantees his work.

I'm still in the process of getting all my parts together. My research has revealed that Intel 2066 MBs and AMD tr4 MBs are "a work in progress" and early at that.

BTW, I have no relationship with this guy.
If you do this, post it here and I'll do the same as soon as I can.
Darn work gets in the way of the fun stuff, you know?

[M344587487]

I've been keeping a close eye on Ryzen pricing since Coffeelake. Threadripper has come down by about 10% in the UK, the 1700 has been getting cheaper as time goes on, and the 1700X and 1800X have just come down after a long time staying steady (but not as much as TR). 1600 has still not budged. Not surprising that prices haven't changed much as Coffeelake is still basically vaporware. I'm hoping for some decent discounts say a month from now, when intel (hopefully) has sorted out its supply issues.

When looking at bang for buck at stock clocks, the 1700 currently beats all other Ryzen offerings including Threadripper when you take the entire system cost into account. This is true even when you factor in that XFR boosts are 50 mhz for non-x, 100 mhz for X, and 200 mhz for TR. Using the XFR speeds and todays UK prices (ebuyer), my naive bang for buck rankings are:
Best
1700 (15)
1700X (15.7)
1600X (17.4)
1600 (17.63)
1800X (17.65)
1300X (19.4) (this does not take into account lack of SMT)
1950X (20.1)
Worst

1700 is also competitive when it comes to power consumption, the 1950X at 4.0Ghz stock is well above the optimum frequency on the power curve. If power efficiency is very important, you'd probably want to underclock and undervolt a 1700 or 1700X (slightly less bang for buck than 1700, but as a better binned chip it should be more power efficient than 1700 at the same clocks). No matter how you look at it, if you're not overclocking the best bang for buck is currently 1700/1700X. For TR to match these in bang for buck, the combined cost of 1950X + mobo needs to come down by around £300. As the high end parts this is unlikely for a while.

Originally Posted by mackerel

...
On Skylake-X delid, if you don't mind doing it yourself, the tools are not that expensive. In the US I'd suggest the Rockit 99 for $40 (+ whatever liquid metal TIM you prefer) https://rockitcool.myshopify.com/pro...akex-kabylakex
...

I've seen youtube videos of these tools being used, it looks easy. Definitely safer than the old razorblade technique ;)

[mackerel]

Originally Posted by M344587487

I've been keeping a close eye on Ryzen pricing since Coffeelake.

Just wanted to check in case I'm missing something here, within the context of prime number finding (excluding sieves), are there many use cases where Ryzen is more desirable over Intel offerings? I can only think in FMA performance, and my 1700 is kinda ok with the doubling in cores somewhat offsetting the halving of IPC, but I regret getting a 1600 to backfill a spare mobo. With Coffee Lake i3-8100, once the lower end mobos come out, I think that will be a killer value combo (within limits of ram bandwidth).

[M344587487]

Originally Posted by mackerel

Just wanted to check in case I'm missing something here, within the context of prime number finding (excluding sieves), are there many use cases where Ryzen is more desirable over Intel offerings? I can only think in FMA performance, and my 1700 is kinda ok with the doubling in cores somewhat offsetting the halving of IPC, but I regret getting a 1600 to backfill a spare mobo. With Coffee Lake i3-8100, once the lower end mobos come out, I think that will be a killer value combo (within limits of ram bandwidth).

Not that I know of. Ryzen is competitive in the general sense, but my understanding is that it's second fiddle to intel for prime searching mainly due to a less performant SIMD implementation. Someone more knowledgable will have to confirm. For LL, high end intel definitely wins with AVX512. But for desktop parts without AVX512, I think you're right that the lower end of the intel lineup is best. Maybe the i3-8350K beats out the i3-8100 due to its increased L3 cache (2 MiB vs 1.5 MiB, theoretically that's less swapping to/from memory but I don't know how much that really applies to tuned programs like Prime95/Mlucas)?

Speaking of cache, some of the lower end Epyc parts appear to have 4 MiB of L3 cache per core. It'll be interesting to see how that shakes out: https://www.scan.co.uk/products/amd-...70w-cpu-retail

[VictordeHolland]

AVX512 is nice, but without sufficient memory bandwidth it has little value over AVX2.

[Mark Rose]

Originally Posted by M344587487

But for desktop parts without AVX512, I think you're right that the lower end of the intel lineup is best. Maybe the i3-8350K beats out the i3-8100 due to its increased L3 cache (2 MiB vs 1.5 MiB, theoretically that's less swapping to/from memory but I don't know how much that really applies to tuned programs like Prime95/Mlucas)?

I suspect it won't make much difference. In this post George says L3 size won't affect FFT selection, and with FFT selection taking into account cache sizes, my guess is L3 is largely irrelevant unless the data fits entirely within the L3 cache.

I highly doubt the extra 50% of a i3-8350k over a i3-8100 will really give 50% more performance per dollar.

[ewmayer]

Originally Posted by M344587487

Not that I know of. Ryzen is competitive in the general sense, but my understanding is that it's second fiddle to intel for prime searching mainly due to a less performant SIMD implementation.

Sticking to AVX2-capable Intel chips, I get around half the per-cycle throughput on Ryzen vs Intel but Ryzen has double the core count, so the question boils down to total cost of ownership for the hardware + electricity. How does Ryzen's total power draw compare to a current Intel quad?

[mackerel]

Originally Posted by M344587487

Speaking of cache, some of the lower end Epyc parts appear to have 4 MiB of L3 cache per core. It'll be interesting to see how that shakes out: https://www.scan.co.uk/products/amd-...70w-cpu-retail

I haven't looked enough to understand it, but it does look like the performance isn't as good as expected for the exclusive cache in Ryzen, similarly for the non-inclusive cache of Skylake-X. They don't seem to work as well in many use cases compared to the inclusive cache for Intel mainstream CPUs.

Originally Posted by VictordeHolland

AVX512 is nice, but without sufficient memory bandwidth it has little value over AVX2.

It might still be useful for smaller FFT sizes where the data still fits in CPU cache. Or maybe the lower end Intel HEDT CPUs since they have fewer cores but still have quad channel ram.

The bigger problem is lack of software that uses AVX-512... there is talk that future generation of Intel mainstream CPUs will also get it which might give general software developers more of a reason to look at it. Even then it is complicated as there seems to be different implementations of AVX-512. The lower end Skylake-X CPUs were originally reported to only have one unit activated, giving no more performance than AVX2. There have been reports this seems not to be the case and they have both units active, but I haven't looked further into this without something to try on it...

As the total cache generally increases with core counts, we could get ever bigger FFTs to fit in it (assuming multi-thread on single task).

Originally Posted by Mark Rose

I suspect it won't make much difference. In this post George says L3 size won't affect FFT selection, and with FFT selection taking into account cache sizes, my guess is L3 is largely irrelevant unless the data fits entirely within the L3 cache.

I highly doubt the extra 50% of a i3-8350k over a i3-8100 will really give 50% more performance per dollar.

From my observations, you either have enough cache to be practically unlimited, or you have not enough and you're affected by ram bandwidth. There is a transition zone between them where it could make a difference, but unless you're working just in that zone it isn't significant. Once you go well into ram limited zone, I haven't seen a significant difference from L3 cache quantity all else being as equal as possible. Also the difference between 8350k and 8100 is 8MB vs 6MB, just like earlier i7 vs. i5 CPUs.

Originally Posted by ewmayer

Sticking to AVX2-capable Intel chips, I get around half the per-cycle throughput on Ryzen vs Intel but Ryzen has double the core count, so the question boils down to total cost of ownership for the hardware + electricity. How does Ryzen's total power draw compare to a current Intel quad?

Tricky question, and you have to compare individual CPUs. The cheapest 8 core is the 1700, and mine seems to turbo to 3.2 GHz all cores active. The 8100 has 4 cores at 3.6 GHz, although I think we're going to be somewhat ram limited.