The + is Dead - Long Live The 3! - Zen 3 information suggests big IPC and cache improvements

[Dabombinable]

It's like we're back in the early 2000's, with Intel pushing for the highest clock speeds...while being behind AMD in the IPC department.

[leadeater]

25 minutes ago, Dabombinable said:

It's like we're back in the early 2000's, with Intel pushing for the highest clock speeds...while being behind AMD in the IPC department.

Athlon 3400+, anyone want to guess the frequency? Hint: Not 3400Mhz lol

[Arika]

25 minutes ago, Dabombinable said:

It's like we're back in the early 2000's, with Intel pushing for the highest clock speeds...while being behind AMD in the IPC department.

I find it amusing that people think this is the first time AMD has been on top. They both trade spots being the top, it's nothing new and it's what you want. Intel has had a long stint at the top, so it's good that AMD is up there, It makes both companies push for better products

But people thinking this is going to be the downfall of Intel annoy me.

That's not how business works, nor should it be what people root for

[mr moose]

4 minutes ago, leadeater said:

Athlon 3400+, anyone want to guess the frequency? Hint: Not 3400Mhz lol

was it 2.8?  I seem to remember the Athlon 3200 was like 2.2 out of the box. 

 

EDIT: I'm googling now.

[Dabombinable]

20 minutes ago, Arika S said:

I find it amusing that people think this is the first time AMD has been on top. They both trade spots being the top, it's nothing new and it's what you want. Intel has had a long stint at the top, so it's good that AMD is up there, It makes both companies push for better products

But people thinking this is going to be the downfall of Intel annoy me.

That's not how business works, nor should it be what people root for

Going back further, and we have the 386 DX-40. Better than anything Intel had, and rock solid.

[cj09beira]

15 hours ago, 5x5 said:

Intel is in the position AMD were in 2011 with the FX CPUs where they had clockspeeds, high heat output and lower IPC than Intel. Basically, the mighty have fallen.

or the same position they were in in the P4 days

15 hours ago, Dash Lambda said:

Okay, I'm gonna go back a bit farther than might be necessary, but anyway: Originally AMD was actually a second-supplier for Intel chips, they made their business reverse-engineering and copying them for early personal computers. But Intel started to get finicky about licensing and patents so AMD had to do something else. Intel owns the x86 instruction set, so AMD couldn't use it without an agreement with Intel. In 2000, AMD released the AMD64 extension to the x86 instruction set, which added a 64-bit mode. Intel was forced to enter a cross-licensing agreement so AMD could use x86 and Intel could use AMD64.

 

From that point, AMD had a few great years of competition. They were the first to push dual-cores, they bought ATI and started doing GPUs, and they they generally beat Intel on all fronts. But then, for reasons that still aren't clear, they just kind of stopped working on their CPUs. Around 2008 they introduced K10, which was pretty good, and tweaked that until 2011 when they launched Bulldozer, which was underwhelming from the start, and over the next 6 years they just kept bumping the clocks on Bulldozer. They didn't even do it with process optimizations or anything, they just kept raising the TDP. They were actually the first to make a 5Ghz CPU, and it was hot, power-hungry, expensive, and rather slow. They even stuck with PCIe 2.0, never bothered with 3.0 until Zen. Intel kept improving features, efficiency, and IPC and eventually gained such a lead that they stopped innovating too, keeping the desktop at 4 cores and rationing out token 5% performance boosts every now and then while quietly raising prices.

 

Now, after Ryzen, AMD's absolutely crushing it with IPC and efficiency, while Intel's stuck in the mud revising Skylake over and over again trying to compensate with clock speed. Hell, Intel's even still on PCIe 3.0, aren't they?

 

I just hope mid-2020s AMD doesn't turn into early/mid-2010s Intel.

intel had been quite a bully, first they didn't follow the ibm contract just a year into it, then latter when amd was making good products they bribed every oem into not selling amd products, in japan they even managed to prevent almost all sales.

8 hours ago, mr moose said:

And contrary to what people want to believe, 50% is a guess.  It is exactly the same probability as a guess, which means for any given prediction,  anyone on this forum would have exactly the same accuracy.  I don't even understand why people watch his videos let alone take him seriously. 

 

On topic.  It's nice to see AMD getting back on top of things over the last few years and I am genuinely excited to see what the next generation of products bring.  But where they land performance wise still remains to be seen.  I mean IPC is becoming such an ambiguous metric when different workloads perform so differently.    There is no point in jumping up and down about IPC in multi core process when your program only uses one thread and different CPU's clock very differently.

btw the ipc numbers were specifically about single core

48 minutes ago, Arika S said:

I find it amusing that people think this is the first time AMD has been on top. They both trade spots being the top, it's nothing new and it's what you want. Intel has had a long stint at the top, so it's good that AMD is up there, It makes both companies push for better products

But people thinking this is going to be the downfall of Intel annoy me.

That's not how business works, nor should it be what people root for

right now people should root for a weak intel in the next 2-5 years or this competition will be short lived

[Arika]

1 minute ago, cj09beira said:

right now people should root for a weak intel in the next 2-5 years or this competition will be short lived

people should root for both companies, root for competition.

[Curufinwe_wins]

43 minutes ago, Dabombinable said:

Going back further, and we have the 386 DX-40. Better than anything Intel had, and rock solid.

Ehhhhhh, the 486 was still better (at 25 MHz or higher), but was also ludicrously more expensive. Not that this is relevant. Plus it isn't like AMD made any improvements to the 386 design other than straight cloning (well the SX had idle power reduction, but that's really it).

 

Anyways, I would definitely agree that this is much more comparable to the P4 situation, but thankfully (for Intel) Intel doesn't have an arch that is trash, just process nodes that are stagnant really. It is unreasonable for anyone to expect that Intel doesn't have plans for returns to competitive leads, but ofc execution is the key here. At least they aren't in a P4/Bulldozer situation where their UArch is dead end and additionally regression clock for clock from previous variants.

[Blademaster91]

7 minutes ago, cj09beira said:

btw the ipc numbers were specifically about single core

It's still a guess, adordedtv and wccftech make up so much crap I have a difficult time believing any numbers they give.

7 minutes ago, cj09beira said:

right now people should root for a weak intel in the next 2-5 years or this competition will be short lived

People should be rooting for both, so we actually have competition. AMD is still a company, don't think they wouldn't slow down on improvements if Intel failed to have a competing product.

[mr moose]

8 minutes ago, cj09beira said:

btw the ipc numbers were specifically about single core

 

Absolutely.  IPC  remained somewhat relevant while many programs were polarized as either single threaded or multiple threaded.  but now there is so much crossover (not just between different workloads but even across similar workloads like games) the whole IPC = better is no longer a straight up advantage/disadvantage.  

[Chricri]

I need to understand the leak: "32MB L3 shared cache in ONE CCX"
the current Zen 2 configuration is as it 4c+16MB L3/CCX and 2CCX/CCD


so the Zen 3 CCX will be Zen 2 CCD but with a L3 unified (chat about inifinity frabric beween CCX/ CCD?) or something else (8c/CCX and 2CCX/CCD and 2CCD/CPU (->32c/64t)) ?

Planning to upgrade my Ryzen 3000 to a high Ryzen 4000, not interested in Ryzen 5000, DDR5 and all those costs

[leadeater]

14 minutes ago, oeris said:

so the Zen 3 CCX will be Zen 2 CCD but with a L3 unified (chat about inifinity frabric beween CCX/ CCD?) or something else (8c/CCX and 2CCX/CCD and 2CCD/CPU (->32c/64t)) ?

The two most probable configurations are:

• 1x 8 core CCX per CCD offering the same core count configurations we have today
• 2x 8 core CCXs per CCD offering double what we have today

The first one seems like the most likely to me.

[5x5]

1 minute ago, leadeater said:

The two most probable configurations are:

• 1x 8 core CCX per CCD offering the same core count configurations we have today
• 2x 8 core CCXs per CCD offering double what we have today

The first one seems like the most likely to me.

Agreed - I doubt they'll release 32-core AM4 CPUs and risk someone putting it on B350 or A320

[leadeater]

6 minutes ago, 5x5 said:

Agreed - I doubt they'll release 32-core AM4 CPUs and risk someone putting it on B350 or A320

I pressed the on button and the motherboard released the magical blue smoke, how do I get it back in?

[Master Disaster]

20 hours ago, Bombastinator said:

My memory is that there are physics barriers preventing 6gjz from being possible on non exotic cooling.  In any case AMD is beating intel with IPC not clock.  Their chips are clocking lower than intel at least in the gaming space.

Yeah, its called the law of heat = bad for electronics

[greenhorn]

31 minutes ago, 5x5 said:

Agreed - I doubt they'll release 32-core AM4 CPUs and risk someone putting it on B350 or A320

I assume manufacturers will never issue an updated BIOS for those boards that would allow those power-hungry CPUs to be used in the first place.

[Chricri]

1 hour ago, leadeater said:

The two most probable configurations are:

• 1x 8 core CCX per CCD offering the same core count configurations we have today
• 2x 8 core CCXs per CCD offering double what we have today

The first one seems like the most likely to me.

if "1x 8 core CCX per CCD "  CCX will be equal to CCD ? and 2 CCD for 49x0X


Maybe i've difficulties to understand lol



 

1 hour ago, 5x5 said:

Agreed - I doubt they'll release 32-core AM4 CPUs and risk someone putting it on B350 or A320

It could be funny seen YTers testing 32c CPU on A320 but lock at 1GHz because it already use too much power

[leadeater]

5 minutes ago, oeris said:

if "1x 8 core CCX per CCD "  CCX will be equal to CCD

Yep, sort of makes CCXs a bit redundant but in reality even cores in the Intel architecture are compartmentalized too. A CCD has one or more CCXs and an Infinity Fabric PHY interface, that interface could also be doubled in width or doubled in connection bus count. Basically there is more in a CCD than just CCXs and I wonder if more/different things could be added. 

[Bombastinator]

56 minutes ago, Master Disaster said:

Yeah, its called the law of heat = bad for electronics

I don’t know how the determination was reached.  It’s quite old and may be out of date.  At the time iirc it assumed processes getting down to 1nm

[Chricri]

I'm very impatient to see the presentation, have technicals details, benchmarks, price and date to order. AMD, TAKE ALL MY MONEY !!!

[porina]

6 hours ago, LAwLz said:

Again, not factoring in AVX-512. If you take that into account it's somewhere between 100%* and 150%* faster than Whiskey Lake.

We have to be ever more careful about AVX-512 support. It isn't a single set of instructions, but includes multiple sets,and is still dependant on hardware support.

 

You may hear of one unit and two unit AVX-512 implementations on Intel. The two unit ones have double the floating point resources compared to one unit ones. A one unit CPU may be no faster than AVX2 since it has about the same hardware resource. With a two unit implementation you can do "up to" double the work of AVX2 if not limited elsewhere. Reason for mentioning this, is we've seen similar for AVX2 between Zen and Zen 2. Zen and Zen+ only had half the FP resource compared to Intel mainstream implementations. Zen 2 finally caught up (and maybe went slightly ahead) when they finally put in that extra FP resource. If AMD go AVX-512, it will be interesting to see how they go. Indication so far is that Intel future mainstream implementations will be one unit, leaving 2 unit to more premium solutions. AMD might eventually follow on that, since historically they have lagged Intel in this area downplaying its importance.

 

Also note AVX-512 includes new instruction sets that might become more interesting in future. Most recently was VNNI extension to accelerate neural networks.

 

5 hours ago, leadeater said:

The way I read is Zen3 isn't getting AVX-512 too, reads to me as coming in Zen4.

Agreed.

 

 

 

Anyway, back on the report, nothing we haven't heard already. The bigger unified CCX will help with more demanding multi-thread workloads. More IPC will be interesting to see how or where that is delivered in practice. I'd suspect it would lean towards SMT friendly work more so than single thread, but let's see.

[cj09beira]

40 minutes ago, porina said:

We have to be ever more careful about AVX-512 support. It isn't a single set of instructions, but includes multiple sets,and is still dependant on hardware support.

 

You may hear of one unit and two unit AVX-512 implementations on Intel. The two unit ones have double the floating point resources compared to one unit ones. A one unit CPU may be no faster than AVX2 since it has about the same hardware resource. With a two unit implementation you can do "up to" double the work of AVX2 if not limited elsewhere. Reason for mentioning this, is we've seen similar for AVX2 between Zen and Zen 2. Zen and Zen+ only had half the FP resource compared to Intel mainstream implementations. Zen 2 finally caught up (and maybe went slightly ahead) when they finally put in that extra FP resource. If AMD go AVX-512, it will be interesting to see how they go. Indication so far is that Intel future mainstream implementations will be one unit, leaving 2 unit to more premium solutions. AMD might eventually follow on that, since historically they have lagged Intel in this area downplaying its importance.

 

Also note AVX-512 includes new instruction sets that might become more interesting in future. Most recently was VNNI extension to accelerate neural networks.

 

Agreed.

 

 

 

Anyway, back on the report, nothing we haven't heard already. The bigger unified CCX will help with more demanding multi-thread workloads. More IPC will be interesting to see how or where that is delivered in practice. I'd suspect it would lean towards SMT friendly work more so than single thread, but let's see.

amd says its waiting for it to make sense, due to the massive memory bottlenecks it has, so it being planned for zen 4 makes perfect sense as that is when they also plan to have ddr5 support 

[leadeater]

47 minutes ago, porina said:

We have to be ever more careful about AVX-512 support. It isn't a single set of instructions, but includes multiple sets,and is still dependant on hardware support.

 

You may hear of one unit and two unit AVX-512 implementations on Intel. The two unit ones have double the floating point resources compared to one unit ones. A one unit CPU may be no faster than AVX2 since it has about the same hardware resource. With a two unit implementation you can do "up to" double the work of AVX2 if not limited elsewhere. Reason for mentioning this, is we've seen similar for AVX2 between Zen and Zen 2. Zen and Zen+ only had half the FP resource compared to Intel mainstream implementations. Zen 2 finally caught up (and maybe went slightly ahead) when they finally put in that extra FP resource. If AMD go AVX-512, it will be interesting to see how they go. Indication so far is that Intel future mainstream implementations will be one unit, leaving 2 unit to more premium solutions. AMD might eventually follow on that, since historically they have lagged Intel in this area downplaying its importance.

 

Also note AVX-512 includes new instruction sets that might become more interesting in future. Most recently was VNNI extension to accelerate neural networks.

Here's a nice graphic, while a bit old now, of just how not easy it is to talk about AVX-512 as a single thing

 

Skylake supports AVX-512 right? 

[porina]

34 minutes ago, cj09beira said:

amd says its waiting for it to make sense, due to the massive memory bottlenecks it has, so it being planned for zen 4 makes perfect sense as that is when they also plan to have ddr5 support 

It is quite a struggle for sure on Intel side, although the generous caches on Zen 2 help, likely even more so when they get more unified in Zen 3. It's even better if you don't have to touch the ram in the first place.

 

25 minutes ago, leadeater said:

Skylake supports AVX-512 right? 

Skylake-X (and Cascade Lake-X) at least have the two unit implementations so can have a nice throughput, if you can tame the power.

[RejZoR]

9 hours ago, mr moose said:

So a metric that is independent of total clock speed only matters due to clock speed?

 

IPC is instruction per clock,  so it doesn't matter if the clock is 3Ghz or 33Ghz the frequency of the clock does not change IPC.  What changes IPC is how well the the HT/SMT is working, how well the code is refined and how many cores nd combinations of the three.   Clocks itself do not change IPC.

Erm, no. Can any AMD do 5GHz at any point? No. Intel can. So, how are you going to compare IPC clock to clock when AMD can't do 5GHz at all? By downclocking Intel to 4 point something? Yeah, no. Sure you get raw IPC metric that way, but you don't get realistic performance numbers. And that's the point I was making. Intel might not have class leading IPC, but they make up for that with clock. For now. They can't just increase clocks to infinity though as 5GHz is already stupid high for a clock...