- Mar 3, 2017
- 1,747
- 6,598
- 136
The real shocker for server is that Zen 5 is an actual slam dunk.
I'm sure it would sell a couple of units. People here maybe. But this is such a niche of a niche of a niche that's it's clearly not worth the engineers and packaging effort.You'd be surprised how many people buy based on some irrelevant benchmark.
It's the main reason Strix is 4+8 instead of something more rational. If the competition is spamming cores you better too.
I don't think it would be a good part for many people but it might sell anyway.
Yeah, but prebuilt desktop PCs use same CPUs as DIY.DiY desktop is a small slice of client. Mostly that's laptops or prebuilt desktop.
Not anymore because AMD has bumped the price so much for that platform and it’s CPUs that it’s no longer interesting for most people.If you really want higher core counts you get threadripper.
Hold those horses.I think people are beginning to understand how amazing Strix Halo will be for developers of all sorts of programs.
Just give it day 1 ROCm on both Windows and all major Linux distros and we are good to go. You can pretty much do anything you would usually do on a desktop workstation outside of the really extreme stuff, in which case you just load up a cloud instance or remote desktop.
If AMD marketing has some braincells left, they might steal some customers from the very entrenched Mac dev crowd.
Halo has a 256-bit memory interface unlike STX1 with 128-bit. It also supports 8533 MT/s memory speed compared to STX's 7500, with 32 MB of MALL on top.Hold those horses.
Membw is going to be a MASSIVE issue. I can't even imagine how they intend to do it, unless they're going to go for something like LPDDR5X + some GDDR6 locked in somewhere for the GPU. Something like 16Go GDDR6 for 64Go total RAM may just be a product. Otherwise I really can't fathom how Halo is going to handle 40 CUs when Point is already seeing some games jump 40% from PHX to STX-P and some others barely nudge a 15% because membw is the problem, not compute.
Halo has all the enhancements to uncore, fabrics, membw, mem capacity and even good old MALL.Hold those horses.
Membw is going to be a MASSIVE issue. I can't even imagine how they intend to do it, unless they're going to go for something like LPDDR5X + some GDDR6 locked in somewhere for the GPU. Something like 16Go GDDR6 for 64Go total RAM may just be a product. Otherwise I really can't fathom how Halo is going to handle 40 CUs when Point is already seeing some games jump 40% from PHX to STX-P and some others barely nudge a 15% because membw is the problem, not compute.
Depends on what you mean by slam dunk. Is it ~20% better than Zen4 in those specific use cases, looking at the perf tests that were linked to (comparing 7700X vs 9700X)?The real shocker for server is that Zen 5 is an actual slam dunk.
- DBs
- in memory DBs
- web server
- python/JS/php
STX-P's 16CUs are already feeling the pressure.Halo has a 256-bit memory interface unlike STX1 with 128-bit. It also supports 8533 MT/s memory speed compared to STX's 7500.
Ok the MALL may make a more serious difference. But I'm still skeptical. When I'll see direct bench comparisons between 12CU STX and 16CU STX, I think we'll see just how bad the bottlenecks are.Halo has all the enhancements to uncore, fabrics, membw, mem capacity and even good old MALL.
The CPU is the selling point more than the GPU, at least until GPGPU software gets more competitive.
Can I see some numbers/tests?@CouncilorIrissa beat me to it so props there. Though it isn't just "pretty good" at js, it's flipping the script there and blowing everything the competition has out of the water.
Zen (4)5%Depends on what you mean by slam dunk. Is it ~20% better than Zen4 in those specific use cases, looking at the perf tests that were linked to (comparing 7700X vs 9700X)?
If they were aiming for 16% general IPC increase with Zen5, then I’d expect more than 20% in some cherry-picked server/DC use cases.
Fixed it for you.You'd be surprised how many people buy without knowing anything about the CPU in their laptop.
Been saying this since day1, also for Zen6 (Halo successor) move to on-package memory and that would enable a 512-bit bus on a laptop. If AMD is serious about AI then it’s something they have to do to fully convince that crowd.I think people are beginning to understand how amazing Strix Halo will be for developers of all sorts of programs.
Just give it day 1 ROCm on both Windows and all major Linux distros and we are good to go. You can pretty much do anything you would usually do on a desktop workstation outside of the really extreme stuff, in which case you just load up a cloud instance or remote desktop.
If AMD marketing has some braincells left, they might steal some customers from the very entrenched Mac dev crowd.
Except it's much more than 20% in some of those.Depends on what you mean by slam dunk. Is it ~20% better than Zen4 in those specific use cases, looking at the perf tests that were linked to (comparing 7700X vs 9700X)?
If they were aiming for 16% general IPC increase with Zen5, then I’d expect more than 20% in some cherry-picked server/DC use cases.
Ok, seems like Zen5 is primarily for server/DC and not client then.Except it's much more than 20% in some of those.
View attachment 105080
+25.5% perf @ iso-power is nothing to scoff at. And that's GNR, whereas cloud will buy Turin-D on N3E with even bigger iso-power gains and a non-cope memory subsystem.
And I didn't have to take something like OpenSSL which uses the AVX-512.
Besides, for AMD server/DC is not a "cherry-picked" use case, it's literally what makes them the most money.
Probably? There are some rumours that Zen 6 DT client will be different from server.Ok, seems like Zen5 is primarily for server/DC and not client then.
Would be nice if AMD went with separate client vs server dies/cores going forward, so client could make other prioritizations. Zen6 perhaps?
An 8x full core (optimized for client) with 3d vcache & 16x c core (optimized for server) for total of 24 cores could help AMD tick all boxesOk, seems like Zen5 is primarily for server/DC and not client then.
Would be nice if AMD went with separate client vs server dies/cores going forward, so client could make other prioritizations. Zen6 perhaps?
Yeah so people can complain about how bad the inter-CCX latency is, scheduling issues, and we can hear endlessly “Why didn’t AMD just make a full fat 16 core CPU?”An 8x full core (optimized for client) with 3d vcache & 16x c core (optimized for server) for total of 24 cores could help AMD tick all boxes
Not sure if it makes sense for AMD
Because a 16 stop ringbus will have a pretty harsh clock ceiling, or you end up with melting fabrics due to voltage spikes like on a certain companies 12 stop ringbuses.Why didn’t AMD just make a full fat 16 core CPU?
Ok, seems like Zen5 is primarily for server/DC and not client then.
Would be nice if AMD went with separate client vs server dies/cores going forward, so client could make other prioritizations. Zen6 perhaps?
MLID said that would happen with Zen7 (and he wasn't even sure it would be called Zen7- "post Zen core").Would be nice if AMD went with separate client vs server dies/cores going forward, so client could make other prioritizations. Zen6 perhaps?
The real shocker for server is that Zen 5 is an actual slam dunk.
- DBs
- in memory DBs
- web server
- python/JS/php
All that's left is for someone to test a load balancer and we're set. It's basically THE chip to buy for anything cloud/general web services. Freaking 8 core Zen 5 is nibbling at 13900ks/14700ks, with the odd "destroy everything including 14900Ks" benchmark.
If I were a CP I'd just buy EPYC Zen 5 and nothing else. It may be a bust in games and in so many apps, but for web services it's obliterating any competition.
MLID said that would happen with Zen7 (and he wasn't even sure it would be called Zen7- "post Zen core").