AMD Ryzen 5000 Series / Zen 3 Launch

[atomsymbol]

[cut]If it weren't for AMD we'd be spending $500 for a single threaded quad core processor today. If not more.

AMD isn't the only competitor Intel has. For example, there is a real possibility of the notebook market and/or server market being ruled by ARM CPUs by the end of this decade if Intel stops innovating the x86 CPU design and stops lowering the price of a CPU core. There also do exist other competitors than ARM (such as https://www.tachyum.com/assets/img/T...ips%202018.pdf).

[geearf]

Michael isn't that a typo:

Also in the initial Ryzen 5000 series is the 5800X at 8-core / 16-threads with 4.7 / 3.8GHz, 105 Watt TDP, and 35MB cache. There is also the Ryzen 5 3600X at 6 cores / 12 threads.

I'm guessing it should be 5600X not 3600X.

[leipero]

If you don't need it anytime soon, I would skip to Zen 4 on the new socket. Only problem there is that DDR5 prices could be much higher than DDR4, and Zen 2/3 chips will be cheaper. But you get a new and exciting upgrade path.

But that's only when platform is new, if you wait few months DDR5 will go down in cost when speed of the modules goes up same as all previous memory generations. For example DDR4 now is super cheap, you have 3200MT modules for bargain because of availability of faster ones, and considering the fact that there's no CPU that supports modules faster than 3200MT, it's basically useless to waste money on anything above that, and that's talking about current generation, one generation back and youa re capped at 2933MT and so on. Ofc., my assumption here is that you wouldn't overclock RAM since it's pretty much worthless and adds more negatives than positives even for APUs, let alone CPUs.

[jaxa]

I agree that AMD processors made very good progress, but if we want to be realistic then it is worth mentioning that there is a non-negligible probability of Intel being faster throughout the majority of year 2021 (the last 3 of 4 quarters of the year) because of Rocket Lake.

Rocket Lake has to improve by around 15% to tie Zen 3. Rocket Lake is known to use Willow Cove cores backported from 10nm to 14nm. There will be an IPC improvement, but worse than the 25% of Willow Cove on 10nm. Maybe they squeeze out another 100 MHz out of the 14nm process. Add it together and Rocket Lake may claw back the lead, but it won't be a blowout.

AMD could try throwing out another trio of XT CPUs to respond to Rocket Lake. Even a 5% uplift could mean bragging rights.

AMD isn't the only competitor Intel has. For example, there is a real possibility of the notebook market and/or server market being ruled by ARM CPUs by the end of this decade if Intel stops innovating the x86 CPU design and stops lowering the price of a CPU core. There also do exist other competitors than ARM (such as https://www.tachyum.com/assets/img/T...ips%202018.pdf).

AMD's surprise resurgence has definitely hurt ARM in the server and laptop spaces. But ARM stands ready to compete if x86 slows down. From the recent ARM announcement, it looks like the crazy annual IPC increases might be coming to an end.

But that's only when platform is new, if you wait few months DDR5 will go down in cost when speed of the modules goes up same as all previous memory generations. For example DDR4 now is super cheap, you have 3200MT modules for bargain because of availability of faster ones, and considering the fact that there's no CPU that supports modules faster than 3200MT, it's basically useless to waste money on anything above that, and that's talking about current generation, one generation back and youa re capped at 2933MT and so on. Ofc., my assumption here is that you wouldn't overclock RAM since it's pretty much worthless and adds more negatives than positives even for APUs, let alone CPUs.

The zigzagging of DRAM prices since 2012 has effed up all assumptions. We could see another doubling or tripling in prices, DDR4 getting expensive again as production shifts towards DDR5, etc. Maybe we should assume the worst.

[Teggs]

I found it interesting that AMD targeted Piledriver and Excavator owners in one of the slides. 80-90% IPC uplift does seem to make a compelling arguments for a new system.

Also I like that power draw did not (apparently) increase much. I can see a few reasons for it, but it's still refreshing after what Intel and Nvidia have been pulling. It makes upgrading a true one-part process for 400 and 500-series motherboard owners, instead of worrying about power supplies, coolers, and possibly the board itself. Now if they would apply that thinking to the GPU side and let Nvidia be the ones who run hot for once...

Add it together and Rocket Lake may claw back the lead, but it won't be a blowout.

Blow out your power budget.

[leipero]

The zigzagging of DRAM prices since 2012 has effed up all assumptions. We could see another doubling or tripling in prices, DDR4 getting expensive again as production shifts towards DDR5, etc. Maybe we should assume the worst.

Keep on mind that there was serious corruption/conspiracy with DRAM price rigging manufacturers that is solved now, that would explain those prices.

[JS987]

Today I would buy Core i3 10100. Local computer store doesn't sell 4300G. 3200G is unavailable.

[AdrianBc]

This was rather weak IMO. only game and cinebench comparisons, boring. Drink a shot whenever you hear "gaming processor".
it was noted in the end that the ipc comparisons were made with a fixed clock on both 3800xt and 5800x. I'd be interested in
a comparison to tiger lake, the 10900 is old tech.

the other interesting thing was the the big navi benchmarks: they were very close to nvidia 3080 numbers.

There was a result that can be compared to Tiger Lake, the single-thread Cinebench, where Zen 3 is faster by 6% than the best Tiger Lake, both running at the same clock frequency of 4.8 GHz.

So the IPC of Zen 3 is better than of Tiger Lake, so that Rocket Lake (March 2021) will need to reach at least 5.3 GHz in order to match the Zen 3 of 4.9 GHz in single-thread performance.

[user556]

5800X cache figure should be 36 MB (32 MB L3 + 8 x 0.5 MB L2), not 35 MB.

[AdrianBc]

nah the biggest improvement is the new ccx which is comprised of 8 cores and has a unified cache. this will significantly improve latencies pretty much all over the board.

In the presentations they showed in percents the contribution of each area of improvement and each of them contributed just a few percents, but being many, they summed up to 26%.

Both the floating-point and the integer units are said to be wider and the throughput for many operations is rumored to have been improved by e.g. 50%, which is necessary to achieve the advertised speed improvements when averaged over a complete program, which also has many instructions being executed at the same speed as before.