Personal computing discussed

Moderators: renee, Flying Fox, morphine

 
Aranarth
Graphmaster Gerbil
Posts: 1255
Joined: Tue Jan 17, 2006 6:56 am
Location: Big Rapids, Mich. (Est Time Zone)
Contact:

Re: 16 COARZ.... Not announced by AMD

Tue May 28, 2019 11:50 am

wccftech says there is a 16core/32thread 39xx variant in the works.

https://wccftech.com/amd-ryzen-9-16-cor ... enchmarks/

Threadripper is also getting the new zen2 as well.
Main machine: Core I7 -2600K @ 4.0Ghz / 16 gig ram / Radeon RX 580 8gb / 500gb toshiba ssd / 5tb hd
Old machine: Core 2 quad Q6600 @ 3ghz / 8 gig ram / Radeon 7870 / 240 gb PNY ssd / 1tb HD
 
Usacomp2k3
Gerbil God
Posts: 23043
Joined: Thu Apr 01, 2004 4:53 pm
Location: Orlando, FL
Contact:

Re: 16 COARZ.... Not announced by AMD

Tue May 28, 2019 12:02 pm

MileageMayVary wrote:
Dropping this here for comparisons

What is your determiner of same-row status?
It might make more sense visually if all of the *X***'s were in the same row (1700X -> 2700X -> 3700X horizontal).
Also, aren't the 2***g series the '2nd gen" and thus would be under the 2*** non-APU one's?

Edit, after doing some more reading, I’ve answered my own questions. The core count is what reads left to right and AMD confused the naming by making the 2***g APU’s a generation behind but with current generations numbers. I guess that means that the 3000g APU’s will be based on 2000 CPU’s. How confusing.
Last edited by Usacomp2k3 on Tue May 28, 2019 10:03 pm, edited 2 times in total.
 
dragontamer5788
Gerbil Elite
Posts: 522
Joined: Mon May 06, 2013 8:39 am

Re: 16 COARZ.... Not announced by AMD

Tue May 28, 2019 12:06 pm

MileageMayVary wrote:
Dropping this here for comparisons

Image


Thanks. I didn't realize the 3000 series has 2x L3 cache compared to previous generations.

The general rule of thumb is that cache-efficacy grows by sqrt(size). So 2x larger cache means 1.41x performance but the memory controller now has an additional hop. ("Dense" matrix multiplications will benefit from the cache significantly, but "sparse" graph problems will be hit hard by the slower and "further away" memory controller). Its going to be difficult to imagine how most games will be affected: some will clearly get better and some will clearly get worse.

I imagine that the Ryzen 2700x or 2600x could be the "best bang for buck" processor for a while, even in this new generation. 2700x may have less L3 cache and slower clockrate, but its all on one die.
 
blargh4
Gerbil
Posts: 14
Joined: Thu Oct 20, 2016 12:31 pm

Re: 16 COARZ.... Not announced by AMD

Tue May 28, 2019 12:29 pm

Seems like a really solid product. But man, AMD bragging about a 1% per-core edge vs Coffee Lake makes me miss the 10-20% year-to-year leaps in the mobile space lately. Kinda hoped 7nm + new uarch would make for juicier numbers.
 
dragontamer5788
Gerbil Elite
Posts: 522
Joined: Mon May 06, 2013 8:39 am

Re: 16 COARZ.... Not announced by AMD

Tue May 28, 2019 1:03 pm

blargh4 wrote:
Seems like a really solid product. But man, AMD bragging about a 1% per-core edge vs Coffee Lake makes me miss the 10-20% year-to-year leaps in the mobile space lately. Kinda hoped 7nm + new uarch would make for juicier numbers.


AMD is +32% single-threaded speeds from Ryzen 1xxx series, which is +14% per generation. Back when L2 cache was 17 clock cycles (then Zen+ / 2xxx series dropped that to 12 clocks L2).

Intel clearly has some secret-sauce patents going on for its design: AMD and Intel's backends are similarly thick, but somehow Intel executes code more efficiently than AMD's cores. Maybe a cache improvement or some frontend thingy that's hard to figure out... but Intel is consistently clocking higher while delivering better typical IPC.

In any case, AMD is definitely building up their architecture. They started with a lot of cores, and are now tuning the chip to get better and better IPC... catching up to Intel at a respectable rate. When combined with the superior core-count (for less $$), AMD's chips are definitely capturing an important niche.
 
jackbomb
Gerbil XP
Posts: 363
Joined: Tue Aug 12, 2008 10:25 pm

Re: 16 COARZ.... Not announced by AMD

Tue May 28, 2019 1:53 pm

I haven't been this excited about a CPU launch since Core 2 (they're finally putting Pentium M in a desktop and giving it beefy SIMD!) and Athlon 64 (64-bit yet fully backward compatible! Built in memory controller! Brilliant!).
Like a good neighbor jackbomb is there.
 
Waco
Gold subscriber
Grand Gerbil Poohbah
Posts: 3227
Joined: Tue Jan 20, 2009 4:14 pm
Location: Los Alamos, NM

Re: 16 COARZ.... Not announced by AMD

Tue May 28, 2019 2:02 pm

Agreed. Between Zen 2 on desktop and Rome in the datacenter...this is a good year for AMD!
Desktop: X570 Gaming X | 3900X | 32 GB | Alphacool Eisblock Radeon VII | Heatkiller R3 | Samsung 4K 40" | 1 TB SX8200 Pro + 2 TB 660p + 2 TB SATA SSD
NAS: 1950X | Designare EX | 32 GB ECC | 7x8 TB RAIDZ2 | 8x2 TB RAID10 | FreeNAS | ZFS | LSI SAS
 
Aranarth
Graphmaster Gerbil
Posts: 1255
Joined: Tue Jan 17, 2006 6:56 am
Location: Big Rapids, Mich. (Est Time Zone)
Contact:

Re: 16 COARZ.... Not announced by AMD

Tue May 28, 2019 2:10 pm

jackbomb wrote:
I haven't been this excited about a CPU launch since Core 2 (they're finally putting Pentium M in a desktop and giving it beefy SIMD!) and Athlon 64 (64-bit yet fully backward compatible! Built in memory controller! Brilliant!).


My reaction is:
Wow AMD is back in the game, decent price, good performance per watt!

Finally a real upgrade from my 2600k!

Man I hope it overclocks like crazy!
Main machine: Core I7 -2600K @ 4.0Ghz / 16 gig ram / Radeon RX 580 8gb / 500gb toshiba ssd / 5tb hd
Old machine: Core 2 quad Q6600 @ 3ghz / 8 gig ram / Radeon 7870 / 240 gb PNY ssd / 1tb HD
 
Goty
Gerbil
Posts: 31
Joined: Wed Mar 04, 2009 7:41 am

Re: 16 COARZ.... Not announced by AMD

Tue May 28, 2019 3:06 pm

I think I see a 3800X in my future. Now to see if I can find enough junk to sell to raise the money for that and a new motherboard (my PRIME X370-PRO will support all of the new chips according to ASUS, but I want one for the heck of it.)
 
techguy
Gerbil XP
Posts: 404
Joined: Tue Aug 10, 2010 9:12 am

Re: 16 COARZ.... Not announced by AMD

Tue May 28, 2019 3:11 pm

I'm excited about Zen2 also, but I would advise a tiny bit of restraint. We don't have any solid performance numbers yet, let alone from 3rd parties. I'm not saying Ryzen 3000 will be a bust, but what if IPC is still not on-par with Coffee Lake Refresh, let alone Sunny Cove? What if even TSMC's 7nm process can't get these parts up to the same clocks as Intel's 14++?

There are still questions to be answered. Only a few short weeks!
 
freebird
Gerbil
Posts: 89
Joined: Thu Aug 31, 2006 4:03 pm

Re: 16 COARZ.... Not announced by AMD

Tue May 28, 2019 4:52 pm

Krogoth wrote:
The abstinence of a customer-tier 16-core Zen2 desktop SKU can easily be explained that AMD doesn't want to cannibalize their Threadripper 1/2 line-up yet. They'll probably announce them around the same time as Threadripper 3's official debut. It'll also be probably be a challenge to run on older AM4 boards that pre-date 5xx family.


I think maximizing revenue at the beginning of the launch plays a big part also. Not sure what the breadth of the EPYC2 product launch will be, but regardless AMD will be selling working 8-core chiplets for much more in an EPYC2 package and either Ryzen3000 or TR3. In addition, EPYC2 TDP limits are getting bumped up to 240w TDP. So if they can make an 8-core chiplet run at 3.6/4.4 in 65w for Ryzen 3700x, what speeds can they hit with 4 of those in an EPYC2 package? Do you think they could hit 3.2/4.0? Even at 3.0/3.8 will be a killer. Compare that to what is available now and what is being charged.

EPYC 7601 is 32core 2.2/3.2 @ 180w TDP with a $4000 price tag. A 32-core EPYC2 will have much higher clocks, Double FP power at same clocks, IPC improvements, simplified NUMA, increased number of Encrypted VM keys to support more VMs. How much is all of that worth? I'd think a minimum of $1000 a 7nm 8 working core chiplet. Same goes for top 64-core EPYC2 which AMD has compared with Intel 8180 it did double the processing in the molecular DNA demo. So asking $8K for it should not be a problem since the Intel 8180 lists for more than that. Sure it won't double the performance of 8180 in most workloads, but most will be bought because it does offer significant performance benefits over the 8180.

Besides AMD will need that enhanced revenue to try and stay ahead... designing and bringing to market chips below 7nm is going to become EXTREMELY EXPENSIVE.
https://wccftech.com/apple-5nm-3nm-cost-transistors/

Full link:
https://wccftech.com/apple-5nm-3nm-cost-transistors/

I think 16-core Ryzen and TR3 won't show up until around X-mas 2019.
 
Waco
Gold subscriber
Grand Gerbil Poohbah
Posts: 3227
Joined: Tue Jan 20, 2009 4:14 pm
Location: Los Alamos, NM

Re: 16 COARZ.... Not announced by AMD

Tue May 28, 2019 5:09 pm

Don't forget double the IO throughput (128 lanes of PCIe4) on Rome as well as a single domain for them instead of 4 per socket. Easier to extract the performance that's possible for sure.
Desktop: X570 Gaming X | 3900X | 32 GB | Alphacool Eisblock Radeon VII | Heatkiller R3 | Samsung 4K 40" | 1 TB SX8200 Pro + 2 TB 660p + 2 TB SATA SSD
NAS: 1950X | Designare EX | 32 GB ECC | 7x8 TB RAIDZ2 | 8x2 TB RAID10 | FreeNAS | ZFS | LSI SAS
 
techguy
Gerbil XP
Posts: 404
Joined: Tue Aug 10, 2010 9:12 am

Re: 16 COARZ.... Not announced by AMD

Tue May 28, 2019 5:19 pm

freebird wrote:
Krogoth wrote:
The abstinence of a customer-tier 16-core Zen2 desktop SKU can easily be explained that AMD doesn't want to cannibalize their Threadripper 1/2 line-up yet. They'll probably announce them around the same time as Threadripper 3's official debut. It'll also be probably be a challenge to run on older AM4 boards that pre-date 5xx family.


I think maximizing revenue at the beginning of the launch plays a big part also. Not sure what the breadth of the EPYC2 product launch will be, but regardless AMD will be selling working 8-core chiplets for much more in an EPYC2 package and either Ryzen3000 or TR3. In addition, EPYC2 TDP limits are getting bumped up to 240w TDP. So if they can make an 8-core chiplet run at 3.6/4.4 in 65w for Ryzen 3700x, what speeds can they hit with 4 of those in an EPYC2 package? Do you think they could hit 3.2/4.0? Even at 3.0/3.8 will be a killer. Compare that to what is available now and what is being charged.

EPYC 7601 is 32core 2.2/3.2 @ 180w TDP with a $4000 price tag. A 32-core EPYC2 will have much higher clocks, Double FP power at same clocks, IPC improvements, simplified NUMA, increased number of Encrypted VM keys to support more VMs. How much is all of that worth? I'd think a minimum of $1000 a 7nm 8 working core chiplet. Same goes for top 64-core EPYC2 which AMD has compared with Intel 8180 it did double the processing in the molecular DNA demo. So asking $8K for it should not be a problem since the Intel 8180 lists for more than that. Sure it won't double the performance of 8180 in most workloads, but most will be bought because it does offer significant performance benefits over the 8180.

Besides AMD will need that enhanced revenue to try and stay ahead... designing and bringing to market chips below 7nm is going to become EXTREMELY EXPENSIVE.
https://wccftech.com/apple-5nm-3nm-cost-transistors/

Full link:
https://wccftech.com/apple-5nm-3nm-cost-transistors/

I think 16-core Ryzen and TR3 won't show up until around X-mas 2019.


What makes you say that Epyc 2 32-core will have "simplified NUMA" compared to Epyc 1? It's the same number of chiplets and CCXs.
 
Waco
Gold subscriber
Grand Gerbil Poohbah
Posts: 3227
Joined: Tue Jan 20, 2009 4:14 pm
Location: Los Alamos, NM

Re: 16 COARZ.... Not announced by AMD

Tue May 28, 2019 5:50 pm

techguy wrote:
What makes you say that Epyc 2 32-core will have "simplified NUMA" compared to Epyc 1? It's the same number of chiplets and CCXs.

It's one NUMA domain versus 4. Same number of chiplets (for the 32 core version), but all memory/IO will be through the central IO hub.
Desktop: X570 Gaming X | 3900X | 32 GB | Alphacool Eisblock Radeon VII | Heatkiller R3 | Samsung 4K 40" | 1 TB SX8200 Pro + 2 TB 660p + 2 TB SATA SSD
NAS: 1950X | Designare EX | 32 GB ECC | 7x8 TB RAIDZ2 | 8x2 TB RAID10 | FreeNAS | ZFS | LSI SAS
 
MileageMayVary
Gerbil XP
Posts: 370
Joined: Thu Dec 10, 2015 9:18 am
Location: Baltimore

Re: 16 COARZ.... Not announced by AMD

Tue May 28, 2019 6:45 pm

freebird wrote:
I think maximizing revenue at the beginning of the launch plays a big part also. Not sure what the breadth of the EPYC2 product launch will be, but regardless AMD will be selling working 8-core chiplets for much more in an EPYC2 package and either Ryzen3000 or TR3. In addition, EPYC2 TDP limits are getting bumped up to 240w TDP. So if they can make an 8-core chiplet run at 3.6/4.4 in 65w for Ryzen 3700x, what speeds can they hit with 4 of those in an EPYC2 package? Do you think they could hit 3.2/4.0? Even at 3.0/3.8 will be a killer. Compare that to what is available now and what is being charged.


Even server side at work, as much as I like moar cores! I dont need a million of them. Half a million with a speed bump would be better.

How many Xeons or Epycs have a Base Clock of 2.5 GHz or greater and 16 or more cores? There's a few if you can afford Xeon Platinums.
Main rig: Ryzen 3600X, R9 290@1100MHz, 16GB@2933MHz, 1080-1440-1080 Ultrasharps.
 
The Egg
Gold subscriber
Minister of Gerbil Affairs
Posts: 2935
Joined: Sun Apr 06, 2008 4:46 pm

Re: 16 COARZ.... Not announced by AMD

Wed May 29, 2019 8:38 am

I'd be interested to see what would happen if they slapped in a stack or two of HBM2 memory as a chiplet to make a jumbo-sized L4 cache a la Intel's eDRAM. In addition to giving the CPU a solid boost, it would really allow their integrated GPUs to shine.
 
techguy
Gerbil XP
Posts: 404
Joined: Tue Aug 10, 2010 9:12 am

Re: 16 COARZ.... Not announced by AMD

Wed May 29, 2019 10:17 am

Waco wrote:
techguy wrote:
What makes you say that Epyc 2 32-core will have "simplified NUMA" compared to Epyc 1? It's the same number of chiplets and CCXs.

It's one NUMA domain versus 4. Same number of chiplets (for the 32 core version), but all memory/IO will be through the central IO hub.


That's great news! I can actually consider Threadripper for my video production and hosting workloads now!
 
LoneWolf15
Gerbil Elite
Posts: 963
Joined: Tue Feb 17, 2004 8:36 am
Location: SW Meecheegan

Re: 16 COARZ.... Not announced by AMD

Wed May 29, 2019 7:11 pm

dragontamer5788 wrote:
Redocbew wrote:
LoneWolf15 wrote:
Although I'd say that Netburst was a far less efficient architecture, so the analogy isn't near as valid. Lake could scale further IMO if Intel could have gotten their die process ducks in a row, but that killed them. The vulnerabilities in their SMT aren't helping either; it looks like there needs to be a real change in how this is done.

The IPC will be the most interesting thing to see from AMD, IMO. For them to take the crown there will be impressive indeed.


Agreed. I wonder how useful of a metric IPC still is though. It's never been a static figure that always stays the same. Some average value is the best we can do, and now with op-caches, op-fusions, skipping of decoders, and differing numbers of floating point and/or integer instructions accepted for each clock the wiggleroom for this metric has probably increased possibly by quite a lot. It's outside my field, but it always makes me wonder when there's a new announcement like this, and everyone goes bonkers over IPC even though nobody knows what it is.


Anyone who has actually measured IPC using a real profiling tool like AMD uProf or Intel vTune, will know that IPC is something the programmer controls.

I can change the IPC of my code to focus on things that AMD does better, or I can change it to make Intel look better. I can change my code to be memory-limited (0.2 IPC or less), I can change my code to be very CPU heavy (kinda arbitrary, but I can write loops with 2 or 3 IPC. I've heard of programmers hitting 4 to 6 IPC in special cases)

As such, IPC only matters if you somehow keep the fight "fair", and there's a lot of opinion that goes into what makes a fair test. In any case, IPC + Clock rate is a better measurement than just raw clocks, but IPC has "all the complexity" because you have to make a lot of assumptions behind IPC.


I'll agree with you that optimized code certainly plays a part. However, there are still indicators across multiple processors if you use standardized code that works with an instruction-set properly supported by all competitors.

During the P4 Netburst architecture age, you could easily describe the Pentium 4 as what happens when you allow marketing to influence engineering far too much. Intel had a very large hard-on in some circles for "MOAR CLOCK SPEED LOOKS GOOD!" and wanted an architecture that went to 4GHz. Not only didn't they get there, but some high-clocked CPUs like the Pentium 4 Extreme Edition Prescott version at 3.73GHz actually performed more poorly than its previous Northwood version at 3.4GHz. And neither of them were great when compared to an Athlon 64 in the same range. The Athlon 64 X2 was a better choice than the Pentium D, too. In fact, the 2GHz Athlon 64 X2 3800+ kicked the tail of the Pentium D 920 at 2.8GHz, with no real proprietary instructions to give it the advantage of Intel, just a superior design.

What this shows is that some architectures even on the same vendor, in a processor family based on the same architecture (but with changes to allow the clock speed to be increased), with a similar instruction/featureset, instructions per clock can be a real difference with the same code. Also, although a different architecture, one could easily say that Bulldozer was "AMD's Pentium 4". It took considerable clock speed increases and architectural improvements (Steamroller, Piledriver, Excavator) for it to exceed previous K10 architecture CPUs (assuming the same number of cores) in a significant way, and it was still a disappointment. It took radical changes in the form of Ryzen for AMD to become impressive again.

There are apps that take special advantage of features a processor has that may be proprietary. AVX512 or AVX2, for example. Just as SSE4. 4.1, and 4.2 at one point were. Adobe apps took considerable advantage of these, which would definitely skew benchmarks. However, there are applications that are engineered for the x86-64 instruction set without any special features. When a CPU shines in these over another, at the same clock speed (or if the winner is at a lower clock) it isn't just a matter of developers. It's who actually performs more instructions per clock. So it really isn't developer-only.
i9-9900K @4.7GHz, GIGABYTE Z390 Aorus Pro WiFi, 2 x 16GB G.Skill RipJaws V PC3000
Corsair 650D, Seasonic 1Kw Platinum PSU
2x HP EX920 1TB NVMe, Samsung 850 Pro 512GB 2.5", NEC 7200 DVDRW
Gigabyte RTX 2080 Super Gaming OC, Dell S2719DGF 27" LCD
 
LoneWolf15
Gerbil Elite
Posts: 963
Joined: Tue Feb 17, 2004 8:36 am
Location: SW Meecheegan

Re: 16 COARZ.... Not announced by AMD

Wed May 29, 2019 7:24 pm

MileageMayVary wrote:
freebird wrote:
I think maximizing revenue at the beginning of the launch plays a big part also. Not sure what the breadth of the EPYC2 product launch will be, but regardless AMD will be selling working 8-core chiplets for much more in an EPYC2 package and either Ryzen3000 or TR3. In addition, EPYC2 TDP limits are getting bumped up to 240w TDP. So if they can make an 8-core chiplet run at 3.6/4.4 in 65w for Ryzen 3700x, what speeds can they hit with 4 of those in an EPYC2 package? Do you think they could hit 3.2/4.0? Even at 3.0/3.8 will be a killer. Compare that to what is available now and what is being charged.


Even server side at work, as much as I like moar cores! I dont need a million of them. Half a million with a speed bump would be better.

How many Xeons or Epycs have a Base Clock of 2.5 GHz or greater and 16 or more cores? There's a few if you can afford Xeon Platinums.


This is why I didn't go Ryzen 2xxxx even though I think it's a great CPU for the right use, and I also think AMD-chipset mainboards have come a long, long ways since third-parties were making them, or first-gen AMD. Much of the Ryzen 2 line compared about even with my i7-4790K in gaming, and it wasn't because things were GPU-bound. I wanted more cores, but I wouldn't have gotten a performance increase without a clock speed increase and an improvement on Haswell architecture.

That said, I'm really bullish on the Ryzen 3xxxx series CPUs. I really hope they can give Intell a run for their money in IPC and power consumption. As someone who had a K6-233 when it was the fastest desktop CPU around, and also the Duron, Ahtlon, Athlon XP, Athlon 64, and Athlon 64 X2 line, I want my choices to be competitive every time I upgrade. I can't wait to see the benches with them and the X570 mainboards vs Coffee Lake-R.
i9-9900K @4.7GHz, GIGABYTE Z390 Aorus Pro WiFi, 2 x 16GB G.Skill RipJaws V PC3000
Corsair 650D, Seasonic 1Kw Platinum PSU
2x HP EX920 1TB NVMe, Samsung 850 Pro 512GB 2.5", NEC 7200 DVDRW
Gigabyte RTX 2080 Super Gaming OC, Dell S2719DGF 27" LCD
 
Redocbew
Gold subscriber
Minister of Gerbil Affairs
Posts: 2168
Joined: Sat Mar 15, 2014 11:44 am

Re: 16 COARZ.... Not announced by AMD

Wed May 29, 2019 8:05 pm

LoneWolf15 wrote:
When a CPU shines in these over another, at the same clock speed (or if the winner is at a lower clock) it isn't just a matter of developers. It's who actually performs more instructions per clock. So it really isn't developer-only.


Well, yeah. Some chips are faster than others, and you can't always get around that by being clever in software, but what we call IPC is still just an average value. It can change depending on the workload, and/or with bigger caches, more memory bandwidth, and so on. You would hope when attempting to measure IPC that the test workload wouldn't include extensions that don't have widespread support, and that it wouldn't be too pathological towards one thing or another. I was really more curious if having so many tricks and shortcuts in the lifecycle of processing instructions has made it easier for that to happen as one possible reason why nobody publicly discloses IPC for modern CPUs.
Do not meddle in the affairs of archers, for they are subtle and you won't hear them coming.
 
LoneWolf15
Gerbil Elite
Posts: 963
Joined: Tue Feb 17, 2004 8:36 am
Location: SW Meecheegan

Re: 16 COARZ.... Not announced by AMD

Wed May 29, 2019 10:19 pm

Redocbew wrote:
LoneWolf15 wrote:
When a CPU shines in these over another, at the same clock speed (or if the winner is at a lower clock) it isn't just a matter of developers. It's who actually performs more instructions per clock. So it really isn't developer-only.


Well, yeah. Some chips are faster than others, and you can't always get around that by being clever in software, but what we call IPC is still just an average value. It can change depending on the workload, and/or with bigger caches, more memory bandwidth, and so on. You would hope when attempting to measure IPC that the test workload wouldn't include extensions that don't have widespread support, and that it wouldn't be too pathological towards one thing or another. I was really more curious if having so many tricks and shortcuts in the lifecycle of processing instructions has made it easier for that to happen as one possible reason why nobody publicly discloses IPC for modern CPUs.


Tricks only matter if they're developed for. And then they still may only matter if one vendor doesn't implement those tricks as well. An example being that Ryzen 2's AVX implementation was weak compared to Kaby Lake and Coffee Lake, something that is claimed to be rectified in Ryzen 3xxx processors.

Aside from these featuresets, and items like AES-NI encryption (which both do well), you're still talking a standard x86_64 instruction set, developed by AMD and licensed to Intel. Now, the only tricks you might think of here (like out-of-order or speculative execution, branch prediction, etc.) are still executing the same x86_64 instruction set; whether one processor or another does these things better is still a matter of instructions per clock cycle within this instruction set. If one CPU does a better job of branch prediction, there will be a better instructions-per-clock rating, all other things being equal. One of the reasons Intel's Pentium 4 Prescott was rather poor at IPC was its long instruction pipelines (done to allow clock speed increases), combined with a much higher cache latency than the P4 Northwood that preceded it (despite the cache being larger) meant that unless you ramped up a Prescott to high speeds (kind of like revving a Honda V-TEC engine to get its max power at high RPM), the P4 Northwood would beat it, as it had a much shorter instruction pipeline, and a faster (though smaller) cache. Even with enhanced branch prediction in the Prescott, the pipeline was so much longer that the impact of a pipeline stall was much higher, also requiring a high clock speed to make up for it. This resulted in the P4 Northwood having much better IPC (albeit with the exact same x86 instruction set) up until around 3.2GHz or so. Even then, the Athlon 64 had a decided advantage, even in x86 (non-64 bit) instructions, and a lower-clocked Athlon 64 3200+ (2GHz) could compete with a Pentium 4 3GHz quite easily, beating it in most cases.

In short, aside from special-feature instruction sets, everything goes towards x86_64; special tricks still contribute to increasing IPC here as long as they're made to speed processing of the standard instruction set. IPC may be a fairly generic definition these days, not fully quantified without going into technical whitepapers, but it's still very real, and still very dependent on the hardware.
i9-9900K @4.7GHz, GIGABYTE Z390 Aorus Pro WiFi, 2 x 16GB G.Skill RipJaws V PC3000
Corsair 650D, Seasonic 1Kw Platinum PSU
2x HP EX920 1TB NVMe, Samsung 850 Pro 512GB 2.5", NEC 7200 DVDRW
Gigabyte RTX 2080 Super Gaming OC, Dell S2719DGF 27" LCD
 
synthtel2
Gerbil Elite
Posts: 955
Joined: Mon Nov 16, 2015 10:30 am

Re: 16 COARZ.... Not announced by AMD

Thu May 30, 2019 12:33 am

Nobody publicly discloses IPC as-such because the hard numbers they could tout aren't necessarily relevant to much of anything. IIRC the hard maximums are 6 for *Lake and Zen and 7 for Apple's A12, but no mixed/sustained workload comes close to saturating that. Seeing real IPCs in a real workload isn't hard (see perf/VTune/uProf), but as dragontamer was saying, it varies wildly depending on the workload and relatively little depending on the CPU details.

The colloquial use of the term covers things the literal meaning doesn't, and the colloquial meaning can't very well be turned from a relative quantity to an absolute one even by force; if one CPU is faster than another strictly because it supports SIMD twice as wide, it would be colloquially understood to have higher IPC despite the actual number of instructions per clock possibly being the same.
 
dragontamer5788
Gerbil Elite
Posts: 522
Joined: Mon May 06, 2013 8:39 am

Re: 16 COARZ.... Not announced by AMD

Thu May 30, 2019 1:22 am

LoneWolf15 wrote:
Tricks only matter if they're developed for. And then they still may only matter if one vendor doesn't implement those tricks as well. An example being that Ryzen 2's AVX implementation was weak compared to Kaby Lake and Coffee Lake, something that is claimed to be rectified in Ryzen 3xxx processors.


What about how FX-8350 was absurdly good at Stockfish?

https://sedatchess.wordpress.com/benchmarks-2/

AMD FX-8350 @ 4.00 GHz (stock speed) gives a Stockfish score of 7721. In contrast, a i7-2600 Sandy Bridge at stock only scores 5424

There's no AVX, SSE, or any "special" instructions at all in Stockfish. Its a straight forward multithreaded integer-based program. But guess what? Those are the kinds of programs that the AMD Bulldozer line was best at computing.

I pretty much can take any chip, and find you a program (even a "normal" program like Stockfish) that will run better on one CPU than others. The only way to test IPC is to arbitrarily choose some figure that "feels right". IPC is more of an art than a science.
 
Redocbew
Gold subscriber
Minister of Gerbil Affairs
Posts: 2168
Joined: Sat Mar 15, 2014 11:44 am

Re: 16 COARZ.... Not announced by AMD

Thu May 30, 2019 2:59 am

synthtel2 wrote:
The colloquial use of the term covers things the literal meaning doesn't, and the colloquial meaning can't very well be turned from a relative quantity to an absolute one even by force; if one CPU is faster than another strictly because it supports SIMD twice as wide, it would be colloquially understood to have higher IPC despite the actual number of instructions per clock possibly being the same.


I guess that's what I was thinking of, but wasn't using good terminology to describe it. Using the term informally like that and saying "This chip wins! It must have better IPC!" even though that may not strictly be true I guess is just an easier way to think about it.
Do not meddle in the affairs of archers, for they are subtle and you won't hear them coming.
 
Aranarth
Graphmaster Gerbil
Posts: 1255
Joined: Tue Jan 17, 2006 6:56 am
Location: Big Rapids, Mich. (Est Time Zone)
Contact:

Re: 16 COARZ.... Not announced by AMD

Thu May 30, 2019 9:16 am

IPC vs MHZ is similar to talking about power output for cars.
A lot of people look at hp numbers but that does not tell the entire story.
You want HP and Torque numbers. if two cars with the same weight and same HP but one has 30% more torque you will get more performance and better fuel economy.
Usually more torque points to a wider power band and better acceleration and more immediate power without the need to wind the engine up.

Processor performance is related to average IPC x mhz. You need both to determine processor performance.

If I have one processor that gives me 4IPC at 2gz it should perform similarly to a machine doing 2ipc at 4ghz all other things being equal.
The biggest difference will be power usage with the first machine using much less power than the second one on the same manufacturing process.

With that out of the way overall system performance brings everything into play. Memory speed, subsystem latencies, etc. The machine is only as fast as its slowest part.
That is why a much slower machine with an ssd will outperform a faster machine with a Hard drive until all of the requested data is loaded into memory.

Zen2 has the following benefits so far:

PCIE -4
more ipc
more mhz
larger caches
faster memory
more cores

What we don't know is how the changes made will affect latency and what changes have been made to the memory controller.
if memory speed has been increased but controller latency is worse you are not going to see as big of jump in memory bandwidth as you expect.
Main machine: Core I7 -2600K @ 4.0Ghz / 16 gig ram / Radeon RX 580 8gb / 500gb toshiba ssd / 5tb hd
Old machine: Core 2 quad Q6600 @ 3ghz / 8 gig ram / Radeon 7870 / 240 gb PNY ssd / 1tb HD
 
Buub
Maximum Gerbil
Posts: 4867
Joined: Sat Nov 09, 2002 11:59 pm
Location: Seattle, WA
Contact:

Re: 16 COARZ.... Not announced by AMD

Thu May 30, 2019 10:04 am

While I get your gist, that's not actually how horsepower and torque work. Torque is a measure of rotational force. Horsepower is a measure of work. In very simple terms, Horsepower is Torque * RPMs, or in other words, Horsepower is force multiplied by how often (because internal combustion engines pulse their power) that force is applied.

An analog would be in electrical terms, torque is amps, RPMs are volts, and horsepower is watts.
 
Waco
Gold subscriber
Grand Gerbil Poohbah
Posts: 3227
Joined: Tue Jan 20, 2009 4:14 pm
Location: Los Alamos, NM

Re: 16 COARZ.... Not announced by AMD

Thu May 30, 2019 11:51 am

I think the better analogy (if there is one) would be horsepower per pound of weight. Not that it's a great analogy, but it's similar.

500 HP with 4000 pounds is 8 pounds per HP. You can achieve similar performance with 250 HP if your car weighs 2000 pounds, 1000 HP and 8000 pounds, etc.
Desktop: X570 Gaming X | 3900X | 32 GB | Alphacool Eisblock Radeon VII | Heatkiller R3 | Samsung 4K 40" | 1 TB SX8200 Pro + 2 TB 660p + 2 TB SATA SSD
NAS: 1950X | Designare EX | 32 GB ECC | 7x8 TB RAIDZ2 | 8x2 TB RAID10 | FreeNAS | ZFS | LSI SAS
 
defaultluser
Gerbil
Posts: 94
Joined: Tue Feb 14, 2017 11:58 am

Re: 16 COARZ.... Not announced by AMD

Thu May 30, 2019 12:04 pm

dragontamer5788 wrote:
LoneWolf15 wrote:
Tricks only matter if they're developed for. And then they still may only matter if one vendor doesn't implement those tricks as well. An example being that Ryzen 2's AVX implementation was weak compared to Kaby Lake and Coffee Lake, something that is claimed to be rectified in Ryzen 3xxx processors.


What about how FX-8350 was absurdly good at Stockfish?

https://sedatchess.wordpress.com/benchmarks-2/

AMD FX-8350 @ 4.00 GHz (stock speed) gives a Stockfish score of 7721. In contrast, a i7-2600 Sandy Bridge at stock only scores 5424

There's no AVX, SSE, or any "special" instructions at all in Stockfish. Its a straight forward multithreaded integer-based program. But guess what? Those are the kinds of programs that the AMD Bulldozer line was best at computing.

I pretty much can take any chip, and find you a program (even a "normal" program like Stockfish) that will run better on one CPU than others. The only way to test IPC is to arbitrarily choose some figure that "feels right". IPC is more of an art than a science.


No, it's a science called "run more than one benchmark you normally run for new CPU reviews, and average the results together." If you keepm the rest of the systems as close in configuration as possible, you can gt some meaningful IPC results.

This is an established science. It's the reason why benchmarks like SPECINT test have ONE DOZEN DIFFERENT SUBTESTS, that change from one benchmark release to the next (as demands change).

https://en.wikipedia.org/wiki/SPECint

THE ART is determining which benchmarks to add to your tests, but since we've been doing it for thirty years now there's now a method the the madness. The individual test results themselves are NOT ART, only curiosities.
 
Amiga500+
Gerbil
Posts: 76
Joined: Wed Sep 21, 2016 2:10 am

Re: 16 COARZ.... Not announced by AMD

Tue Jun 11, 2019 6:14 am

chuckula wrote:
But you get 12.

AND YOU BETTER FREAKIN' LIKE IT

https://www.anandtech.com/show/14407/am ... -coming-77


That aged well.
 
atari030
Silver subscriber
Gerbil
Posts: 47
Joined: Wed Mar 12, 2014 9:37 pm
Location: MI

Re: 16 COARZ.... Not announced by AMD

Tue Jun 11, 2019 7:48 am

Amiga500+ wrote:
chuckula wrote:
But you get 12.

AND YOU BETTER FREAKIN' LIKE IT

https://www.anandtech.com/show/14407/am ... -coming-77


That aged well.


"Sorry, Charlie."

One day I'll perhaps partially understand his childish obsession with belittling everything Advanced Micro Devices. But it's been years and I still do not.

Who is online

Users browsing this forum: No registered users and 4 guests
GZIP: On