AMD RyZen CPU Architecture for 2017

[sebbbi]

Can't this be resolved in the compiler?

Compiler autovectorization can be configured to target SSE, AVX or AVX2. You can build multiple executables if you want to support all. However autovectorization almost never works with the current vector instruction sets. AVX-512 and ARM SVE are new instruction sets that are designed to be easier to autovectorize than the current SSE/AVX/AVX2. But these instruction sets do not yet exist in consumer products.

Pretty much all games are currently written using vector intrinsics. Each CPU has different intrinsics that are not compatible. For example the PPC based (last gen) consoles had VMX and VMX-128 vector intrinsics. ARM mobile processors have NEON vector intrinsics. x86/x64 processors have MMX, MMX2, SSE, SSE2, SSE3, SSSE3, SSE4, AVX and AVX2 intrinsics. If the programmer uses AVX intrinsics, the binary will not run on older CPU than Sandy Bridge.

Intel has emulator to support newer SSE/AVX instruction sets on older hardware:
https://software.intel.com/en-us/articles/intel-software-development-emulator

Some people have used this emulator to play "No Man's Sky" on AMD Phenom:
https://steamcommunity.com/app/275850/discussions/1/360672304898078434/

This vector instruction emulator is designed to help developers to prototype their code before new hardware is available. It is not designed to provide best possible performance. No Man's Sky seems to use very small amount of SSE4.1 intrinsics, so emulating it on SSE4 hardware seems plausible (as only a handful of instructions are missing). But emulating 256 bit wide AVX on SSE4.1 is going to be much slower as the instruction sets differ so much (3 operand VEX encoding, twice as wide, etc). Every AVX instruction needs to be emulated. I would expect a big performance slow down.

 

[Kaotik]

 

[jacozz]

With the Glofo lpp-process quite mature by now, It seems logical to harvest the faulty 8-core chips and sell them as 6-core chips, to make good use of yield, but what about the 4 core chips?
It wouldn't make much economical sense to turn of up to 50% of the cores just to meet the probably much higher demand for the cheaper Ryzens? Wouldn't it be more logical to have 2 different chips? One with 8 and one with 4 cores, half the size and half the production cost? After all, the 8 core chip looks very much like two clusters of 4 core dies glued together no?

 

[xEx]

From this number we can assume that with oc(higher clocks) a 4 core ryzen chip could be equal than the 7700k?

Enviado desde mi HTC One mediante Tapatalk

 

[Anarchist4000]

It wouldn't make much economical sense to turn of up to 50% of the cores just to meet the probably much higher demand for the cheaper Ryzens?

At that price point you may be looking at APUs for a more price sensitive market. We haven't really seen those yet, but that market stands to be the real winner for AMD. I'd think that offsets a large need for a specific 4 core design. Just salvage 4 cores out of 8 or use an APU with the GPU portion disabled to meet demand.

 

[Laurent06]

Compiler autovectorization can be configured to target SSE, AVX or AVX2. You can build multiple executables if you want to support all.

You can also have multiple paths in your code within the same executable (either generated by the compiler, or by hand).

The problem is that as soon as you have to support multiple code paths, in the same exe or different exe, this increases the validation effort.

Segmenting by removing parts of the instruction set looks incredibly short sighted to me. It slows down adoption.

 

[3dilettante]

With the Glofo lpp-process quite mature by now,

This is the highest performing part AMD has shipped on this node, with a particularly advanced physical implementation versus something like Polaris. It might not be that mature from the standpoint of a leading-edge CPU design.

It seems logical to harvest the faulty 8-core chips and sell them as 6-core chips, to make good use of yield, but what about the 4 core chips?
It wouldn't make much economical sense to turn of up to 50% of the cores just to meet the probably much higher demand for the cheaper Ryzens?

At this point, AMD is probably selling what they can at the highest bins they can. Defects, instability at the clock bins for the more premium products, or excessive consumption for a product in the higher bins can lead to cores being gated.
There would be chips that can clock higher than the quad product, but are down to four due to defects. There would be chips that have 6-8 cores, but cannot clock stably across all of them or require too much voltage to match the parameters of the higher bins.
There might not be enough of any single salvage bucket to give them a unique quad SKU, particularly if it starts to encroach on the more valuable products above or confuses the lineup.

Wouldn't it be more logical to have 2 different chips? One with 8 and one with 4 cores, half the size and half the production cost? After all, the 8 core chip looks very much like two clusters of 4 core dies glued together no?

This is quibbling over a few dollars of silicon at this point, particularly if sticking with dual-channel RAM and keeping the same socket.
Then there'd be dual cores for salvage with no clear demand to sell them to.
The design and up-front costs for a chip with a top price where the salvage quad is at is going to have to sell quite a bit, or AMD can make a little money on 8-cores it would now throw away, and sell the main chip for way more once the top-end and server products launch.


As an aside:
From the marketing picture of Zen, and the disclosure of the 44mm2 CCX size, it looks like the chip is about 183mm2.

 

[Ike Turner]

Ryzen launch event:

 

[Voxilla]

So I wasn't too far off 2 years ago predicting that AMD could beat Intel with true 8 cores vs 4 cores / 8 threads.
(I admit they did even better by adding SMT)
https://forum.beyond3d.com/posts/1866354/

 

[lanek]

So I wasn't too far off 2 years ago predicting that AMD could beat Intel with true 8 cores vs 4 cores / 8 threads.
(I admit they did even better by adding SMT)
https://forum.beyond3d.com/posts/1866354/

In reality they beat / equal 8cores-16thread of Intel ( but i understand what you was mean )... and i cant imagine what the 32cores will do sized against the 24cores of intel ( before new generation ). Now that we have a better idea of what is running with Zen at equal cores and comparable clockspeed. Im curious to see the "Opteron" parts ( i dont know if they will use Opteron naming scheme ).16 cores and up..

 

[Kyyla]

I broke down and preordered a 1700x.

edit: the i5 2500k is dead. Long live the king.

 

[Malo]

I've found one B350 motherboard so far that uses the ALC1220 audio instead of the ALC887, greatly limiting my choices unless I want to spend $200+ on an X370 motherboard instead (which I don't).

 

[Silent_Buddha]

I broke down and preordered a 1700x.

edit: the i5 2500k is dead. Long live the king.

I'm tempted to do the same, but I'm waiting for benchmarks. And at 399 USD, that's a rather large chunk of cash for a CPU. I'm more in the budget range of the i5 range of CPUs (IE 225-250-ish USD). If the performance ends up being mind-blowingly good, I may consider it though.

Regards,
SB

 

[lanek]

I'm tempted to do the same, but I'm waiting for benchmarks. And at 399 USD, that's a rather large chunk of cash for a CPU. I'm more in the budget range of the i5 range of CPUs (IE 225-250-ish USD). If the performance ends up being mind-blowingly good, I may consider it though.

Regards,
SB

Well, like you, i wait complete reviews and i need to see tests on the softwares i use. ( CAD and other ) and motherboard tests. Then for me, it is not really a question of CPU pricing ( i will anyway take the 1700x or 1800x ), but the total amount will increase slighty ( not as high as if i was goes the Intel road )..

- new high end motherboard
- new DDR.
- new waterblock ( could buy only the fixation but as my h2o system is build since some times, i will need clean it, its mean additional pieces as in this case i rarely re use the same, connectors, tubing etc )

But well, complete reviews first.

 

[3dilettante]

The latest die shot does make it clear that are are SRAM arrays external to the CCXs. Possibly some form of filter/directory?

I wonder if there could be any value-add with Vega, since they should be sharing an interconnect.

 

[hoom]

Obligatory die shot via the Anandtech link above

 

[Malo]

It's so sexy it's a three page foldout.

 

[msia2k75]

I wonder what frequency should Ryzen 8 core version reach to be at 45W TDP?

 

[xEx]

2.8?

 

[xEx]

Already best sellers in Amazon, and they aren't available yet

I wish I had some money to buy AMD stock, this is the best time.