Zen2. Is Intel now the gamers choice & price/perf king? (devils advocate).

[GoldenLag]

58 minutes ago, porina said:

Intel HEDT already offers AVX-512. I'm most interested in the subset of instructions which double the FPU performance again over existing Intel consumer CPUs

Isnt the issue with those that the bandwidth is starved?

 

So they have fullsize 512 AVX units. Welp. I guess AMD has some way to go there.

[porina]

9 minutes ago, GoldenLag said:

Isnt the issue with those that the bandwidth is starved?

As always it depends on the use case. On 7800X, for my workloads comparable to Prime95 small FFT, it can show its potential (and burn all the power in the universe while doing so). For large data sets it seems to hit some core scaling limitations more so than bandwidth which I don't entirely understand. I think that more a limitation of mesh and the small L3 cache.

 

9 minutes ago, GoldenLag said:

So they have fullsize 512 AVX units. Welp. I guess AMD has some way to go there.

To be clear, only Skylake-X HEDT and higher end Xeons have the two unit AVX-512 implementation. The lesser Xeons, and future consumer CPUs are expected to have single unit implementations. So they wont see the same uplift in FP performance, but still gain other AVX-512 family instructions I don't care about personally (deep learning and other stuff).

 

I don't understand all the features of AVX-512, but I suspect AMD might be hesitant to follow as they could offer a good part of this in GPU instead. On that note, it will be interesting to see how future Intel GPU will play into this.

[GoldenLag]

4 minutes ago, porina said:

don't understand all the features of AVX-512, but I suspect AMD might be hesitant to follow as they could offer a good part of this in GPU instead. On that note, it will be interesting to see how future Intel GPU will play into this.

They could very much. They could also decide to wait untill ddr5 to actually feed all of it.

 

5 minutes ago, porina said:

As always it depends on the use case. On 7800X, for my workloads comparable to Prime95 small FFT, it can show its potential (and burn all the power in the universe while doing so). For large data sets it seems to hit some core scaling limitations more so than bandwidth which I don't entirely understand. I think that more a limitation of mesh and the small L3 cache.

Could possibly be intercore bandwidth but im not that wellhearsed in that field.

[jerubedo]

10 hours ago, Jurrunio said:

PCPP shows $485

It doesn't take into account all sources. For example, the 9900K is $450 at Microcenter (when you buy a motherboard with it). If you also bring with you the $5 off $30 or more coupon from Retailmenot (which I use every time I go), it's $445. It's also $474.99 at Newegg with the coupon code EMCTBTC22. 

[asand1]

4 hours ago, jerubedo said:

It doesn't take into account all sources. For example, the 9900K is $450 at Microcenter (when you buy a motherboard with it). If you also bring with you the $5 off $30 or more coupon from Retailmenot (which I use every time I go), it's $445. It's also $474.99 at Newegg with the coupon code EMCTBTC22. 

Most people don't have a microcenter so that would be a very fringe case.

[jerubedo]

51 minutes ago, asand1 said:

Most people don't have a microcenter so that would be a very fringe case.

The point is that PCPP is not all-inclusive. It doesn't take into account Newegg coupons (otherwise it would show the current price of $474.99), it doesn't take into account any Brick and Mortar stores like Microcenter, Frys, and TigerDirect, etc. Frys had it on sale the other week for $460. So yeah, maybe Microcenter is a "fringe case" but between all of the brick and mortar stores, there's plenty of accessibility from stores not on PCPP. 

[Jurrunio]

On 6/12/2019 at 9:53 PM, Ebony Falcon said:

as fclk is decoupled entirely with own gate 

FCLK, as in Asus's name for PCIe, is not decoupled at all. 102MHz BCLK can still brick NVMe SSDs. Fabric clock is still tied to memory frequency, just simple ratios say 1:2.