TR Forums

Well I suppose this question has been answered...

I don't think it was ever really in question.


..really though, the results weren't to shabby considering the number of cores. More cores = more heat = lower freq. clock = lower "enthusiast" performance. (..it's not like Intel's 8 core offerings outperform their 4 core versions for enthusiasts.)

The problems with simultaneous multi-threading however is a real disappointment (..edit: and the lack of AVX is as well).


Edit: also sources are mentioning poor memory freq.. but that's probably a chipset problem.

CScottG wrote:

I don't think it was ever really in question.


..really though, the results weren't to shabby considering the number of cores. More cores = more heat = lower freq. clock = lower "enthusiast" performance. (..it's not like Intel's 8 core offerings outperform their 4 core versions for enthusiasts.)

The problems with simultaneous multi-threading however is a real disappointment (..and the lack of AVX is as well).

I don't know where people keep getting this "it doesn't have AVX" thing from. It HAS AVX. So did Bulldozer. AMD's AVX implementation just isn't as efficient as Intel's.

just brew it! wrote:

CScottG wrote:

I don't think it was ever really in question.


..really though, the results weren't to shabby considering the number of cores. More cores = more heat = lower freq. clock = lower "enthusiast" performance. (..it's not like Intel's 8 core offerings outperform their 4 core versions for enthusiasts.)

The problems with simultaneous multi-threading however is a real disappointment (..and the lack of AVX is as well).

I don't know where people keep getting this "it doesn't have AVX" thing from. It HAS AVX. So did Bulldozer. AMD's AVX implementation just isn't as efficient as Intel's.

My bad.

..the lack of *PERFORMING* AVX?

just brew it! wrote:

CScottG wrote:

I don't think it was ever really in question.


..really though, the results weren't to shabby considering the number of cores. More cores = more heat = lower freq. clock = lower "enthusiast" performance. (..it's not like Intel's 8 core offerings outperform their 4 core versions for enthusiasts.)

The problems with simultaneous multi-threading however is a real disappointment (..and the lack of AVX is as well).

I don't know where people keep getting this "it doesn't have AVX" thing from. It HAS AVX. So did Bulldozer. AMD's AVX implementation just isn't as efficient as Intel's.

I don't know that we know about AMD's effective efficiency because AMD decided, for whatever reason, to give Zen literally half the AVX execution capability of modern Intel designs. That 50% theoretical performance deficit trumps whatever relative design capability we'd try to suss out.

Vhalidictes wrote:

just brew it! wrote:

I don't know where people keep getting this "it doesn't have AVX" thing from. It HAS AVX. So did Bulldozer. AMD's AVX implementation just isn't as efficient as Intel's.

I don't know that we know about AMD's effective efficiency because AMD decided, for whatever reason, to give Zen literally half the AVX execution capability of modern Intel designs. That 50% theoretical performance deficit trumps whatever relative design capability we'd try to suss out.

Yeah, I should've said "perfomant" not "efficient". My bad.

Vhalidictes wrote:

just brew it! wrote:

CScottG wrote:

The problems with simultaneous multi-threading however is a real disappointment (..and the lack of AVX is as well).

I don't know where people keep getting this "it doesn't have AVX" thing from. It HAS AVX. So did Bulldozer. AMD's AVX implementation just isn't as efficient as Intel's.

I don't know that we know about AMD's effective efficiency because AMD decided, for whatever reason, to give Zen literally half the AVX execution capability of modern Intel designs. That 50% theoretical performance deficit trumps whatever relative design capability we'd try to suss out.

Yes, because one is AVX, and the other is AVX2. The distinction is AVX is 128bit operations, and AVX2 is 256bit operations. Intel and Ryzen both support AVX gen1, but only Intel supports AVX2.

At least for the time being, now that Intel has been updating open source project code for AVX2 support (as claimed by Ian from Anandtech) AMD has more incentive to support it in future chips.

Kougar, my understanding from the various reviews is that Ryzen can run AVX2 code just fine, it just runs it at half IPC because of retiring instructions as 128mbitx2 instead of 256mbit.

I could be misunderstanding that though.

Vhalidictes wrote:

Kougar, my understanding from the various reviews is that Ryzen can run AVX2 code just fine, it just runs it at half IPC because of retiring instructions as 128mbitx2 instead of 256mbit.

I could be misunderstanding that though.

That's the way I understand it. I don't know if "half IPC" is the proper measurement, but indeed the AVX units in Zen are only 128 bit wide.

DancinJack wrote:

Vhalidictes wrote:

Kougar, my understanding from the various reviews is that Ryzen can run AVX2 code just fine, it just runs it at half IPC because of retiring instructions as 128mbitx2 instead of 256mbit.

I could be misunderstanding that though.

That's the way I understand it. I don't know if "half IPC" is the proper measurement, but indeed the AVX units in Zen are only 128 bit wide.

Right. Ryzen's FPUs are only 128 bits wide, so 256-bit AVX2 instructions can't be executed as efficiently as they can on a CPU with a true 256-bit AVX2 data path. AVX2 code will still run, but there will be a performance penalty since Ryzen will need to perform multiple operations internally to implement the 256-bit operations, whereas Intel will be able to execute them in one go.

As far as how much of a performance penalty that is, we'll have to wait for someone to do some detailed benchmarks of 128- vs. 256-bit AVX on Ryzen to find out. The narrower FPU is only part of the equation; cache and DRAM bandwidth will also factor into it (since twice as much data needs to be shoveled around). For that reason, I expect Intel probably takes a throughput hit on 256-bit AVX2 too; it's just likely to be a much smaller hit than Ryzen.

Looks like my 4790K will see it's third year, and the way I see it the Ryzen-X SKUs are ripoffs versus the non-X with when all of them struggle to breach 4GHz. The IVB level IPC and the lack of OC headroom killed any desire for me to switch.

It's a mixed bag: A steal for MT heavy users, great for the IVB/SB/earlier holdouts versus current Intel offerings, meh for everyone else including ironically the server guys where CPUs are a tiny part of the overall cost. AMD will definitely be forced to cut prices when Coffee Lake 6C mainstream appears later this year.

strangerguy wrote:

Looks like my 4790K will see it's third year

My 4790K is in its (see, no apostrophe) third year and, if the electromigration deities accept my offerings once I find a need to OC, it'll last 7 years just like the C2D E6600 and AXP-2800 (Barton) before it.

I build high-end at the time of build and ride it for many years.

Captain Ned wrote:

I build high-end at the time of build and ride it for many years.

I tend to build mid-range and ride it for fewer years. Which has inevitably led to a surplus of spare parts of varying vintages (many of which have been sold or even given away at TRBBQs). The build I stuck with the longest was probably my dual Athlon MP system, which was probably the only build I've done which could've been considered close to high-end at the time.

Now that the oldest daughter has a decent job (so can afford her own "nice things"), and the parental units have moved to a tiny apartment in a retirement community (only computing devices they have space for are a laptop and smartphones), two of the places I have historically sent my "hand me downs" no longer need them.

just brew it! wrote:

I tend to build mid-range and ride it for fewer years.

That is starting to become more expensive.. CPU pricing is similar (maybe a bit broader range over the last 5 years.. and I'm limiting myself here to a less than $400 part), but m-board pricing seems to be escalating. (..I don't know, maybe it's just me wanting more from my m-board.)

-on the plus-side, at least for Intel, these days unless you purchase a K version - that cheap-ass stock cooler seems to work rather well.

CScottG wrote:

just brew it! wrote:

I tend to build mid-range and ride it for fewer years.

That is starting to become more expensive.. CPU pricing is similar (maybe a bit broader range over the last 5 years.. and I'm limiting myself here to a less than $400 part), but m-board pricing seems to be escalating. (..I don't know, maybe it's just me wanting more from my m-board.)

-on the plus-side, at least for Intel, these days unless you purchase a K version - that cheap-ass stock cooler seems to work rather well.

I think it's you.

With the exception of the above-mentioned dual Athlon MP system I don't think I've ever spent more than $150 for a motherboard or $200 for a CPU. (Current mobo/CPU: Asus M5A99FX/FX-8350)

One of the better systems I've owned was even one of the cheaper ones, based around a sub-$100 Asus M3A78-CM micro-ATX motherboard. Asus was very good about updating the BIOS on that board to support newer CPUs; by the time I finally retired it, it had a hex-core Phenom 1090T in it. The motherboard even still lives on, nearly a decade later, in my home file server; I'd say I really got my money's worth on that M3A78-CM.

Well, it seems that the rumors were true, for the most part, not a knock out punch but competitive if you pick and choose your battles.

I don't see them as the great bargain that some seem to think Ryzen 8 core cpu's are, if you factor in price of cpu, motherboard and ram you still end up with a system that will cost you in excess of $500 for the cheapest octocore setup but you will have half the memory channels, half the AVX2 throughput and no ECC support compared to say a cheap Xeon setup.

The real deal for enthusiast will come later on when AMD releases their 6C/12T chip with a rumored price of about $210, coupled with a lower end board, once they're released, that will probably be the processor that gets many SB and IVB holdouts to upgrade.

The reports of the death of ECC support on AMD's consumer CPU line may have been premature. Anandtech is claiming that the Ryzen CPUs support ECC. We just need the motherboard makers to follow suit.

AMD said the Zen arch HAS ECC support, they just don't validate it on the consumer parts. So, as long as you have a board that supports it and a Ryzen CPU, it should work (for the most part).

DancinJack wrote:

AMD said the Zen arch HAS ECC support, they just don't validate it on the consumer parts.

Where did you see this? Would be nice to have firsthand confirmation.

Edit: I know Zen arch has it, since the arch has to support server/datacenter use cases. What I haven't seen definitively confirmed is that AM4 socket actually has the pins for it. Enterprise platform will be a different (LGA) socket.

just brew it! wrote:

CScottG wrote:

just brew it! wrote:

I tend to build mid-range and ride it for fewer years.

That is starting to become more expensive.. CPU pricing is similar (maybe a bit broader range over the last 5 years.. and I'm limiting myself here to a less than $400 part), but m-board pricing seems to be escalating. (..I don't know, maybe it's just me wanting more from my m-board.)

-on the plus-side, at least for Intel, these days unless you purchase a K version - that cheap-ass stock cooler seems to work rather well.

I think it's you.

With the exception of the above-mentioned dual Athlon MP system I don't think I've ever spent more than $150 for a motherboard or $200 for a CPU. (Current mobo/CPU: Asus M5A99FX/FX-8350)

One of the better systems I've owned was even one of the cheaper ones, based around a sub-$100 Asus M3A78-CM micro-ATX motherboard. Asus was very good about updating the BIOS on that board to support newer CPUs; by the time I finally retired it, it had a hex-core Phenom 1090T in it. The motherboard even still lives on, nearly a decade later, in my home file server; I'd say I really got my money's worth on that M3A78-CM.

..probably.

Still, I rarely spent more than $110-$115, but in the near past you'd be hard-pressed to find the most current USB input on your m-board for less than that. (..my recent purchase for server boards is a pricing exception, but one geared towards ECC memory - so at least I'm expecting to get fleeced.)

Sub $100 Asus boards are pretty rare these days as well (..when compared to the number of m-boards Asus now produces, and its not just an inflationary thing).

Also, it wasn't THAT long ago that the "top" of the enthusiast range of m-boards was around $180-$190, where these days top-of-the-line seems to start close to $300 - beyond that of many server boards.

DancinJack posted a link over on the front page that does the best job of explaining the state of ECC I've yet seen. Thanks Jack!

Edit: Whoops, DancinJack is already in this convo. Sorry if you were about to post that yourself.

Yeah, if the m-board manufacturer's start producing ECC memory bios updates - then look-out: a whole 'nother market will open-up.

The board needs to have the memory bus traces for the extra 8 ECC bits too. It's not just a firmware issue.

..bummer.

Anyone can correct me if I'm wrong, but running AVX2 256bit ops at half speed is the same thing as running AVX 128bit ops at full speed. Or at least that'd be my thought process. So one can still say Ryzen just runs AVX2 code at AVX speed... maybe I'm nitpicking the language too much.

just brew it! wrote:

The reports of the death of ECC support on AMD's consumer CPU line may have been premature. Anandtech is claiming that the Ryzen CPUs support ECC. We just need the motherboard makers to follow suit.

Ryzen's ECC support has been confirmed in the Reddit AMA https://www.reddit.com/r/Amd/comments/5 ... r/def6vs2/ Can't remember if it was mentioned but ASRock does claim ECC support on its X370 boards.

You guys think there will be a Windows scheduling patch for Ryzen? Given how many games see a ~10 FPS boost when SMT is disabled it really seems like it could be another case of Windows scheduling cores unevenly again.

I suppose it's possible, but I would think there'd be less to do at the OS level concerning scheduling for Ryzen than there was with Bulldozer. I'm not sure if that's something which can be handled from the OS this time or not.

Redocbew wrote:

I suppose it's possible, but I would think there'd be less to do at the OS level concerning scheduling for Ryzen than there was with Bulldozer. I'm not sure if that's something which can be handled from the OS this time or not.

If the problem is AMD's SMT implementation, then it could be a *lot* like scheduling for Bulldozer, i.e. "don't schedule two threads where they will need to share execution resources unless you've got more than C/2 threads to schedule" (where C is the number of physical cores). That may be an over-simplification, but then again it might not.

SMT is weird.

But yeah, that makes total sense now that I think about it. Do we know if this penalty in Ryzen associated with SMT is unique to gaming, or does it happen across the board?

A virtual SMT core will always be slower than a "real" core if its partner virtual core is busy. How much slower is going to be a complex function of the workload of both threads, and details of the architecture and SMT implementation.

SMT increases overall throughput and efficiency of a CPU by using otherwise idle execution resources to do useful work. But it makes performance harder to predict because threads running on the same physical core can affect each other in difficult to predict ways.

Kougar wrote:

Vhalidictes wrote:

just brew it! wrote:

I don't know where people keep getting this "it doesn't have AVX" thing from. It HAS AVX. So did Bulldozer. AMD's AVX implementation just isn't as efficient as Intel's.

I don't know that we know about AMD's effective efficiency because AMD decided, for whatever reason, to give Zen literally half the AVX execution capability of modern Intel designs. That 50% theoretical performance deficit trumps whatever relative design capability we'd try to suss out.

Yes, because one is AVX, and the other is AVX2. The distinction is AVX is 128bit operations, and AVX2 is 256bit operations. Intel and Ryzen both support AVX gen1, but only Intel supports AVX2.

At least for the time being, now that Intel has been updating open source project code for AVX2 support (as claimed by Ian from Anandtech) AMD has more incentive to support it in future chips.

AVX essentially updates the SSE instruction sets to a three operand format and also permits 256-bit floating point operations. AVX2 adds 256-bit support for integer operations. While Ryzen lags behind with AVX(2) code it has twice the floating point resources if the code is still using SSE.