AMD CPU speculation... and expert conjecture

[jdwii]

Spoiler

Informations about cache also came out from nowhere (2 times bigger than now):
http://en.wikipedia.org/w/index.php?title=Skylake_(microarchitecture)&diff=580936777&oldid=578769707

http://i.imgur.com/IxWYq.png

 

[de5_Roy]

AMD to launch A68H chipset in September
http://www.cpu-world.com/news_2014/2014082301_AMD_to_launch_A68H_chipset_in_September.html

After the release of Intel's Haswell Refresh processors, demand for AMD's CPUs in the desktop market has been significantly impacted, prompting AMD to take actions to defend its market share, the sources claimed.

AMD's entry-level A58 chipsets, which does not support USB 3.0, target the China market, but demand has been weak since their release, the sources said. The A68, which was not in AMD's originalroadmap, will be priced US$2 higher than the A58, the sources added.

Most motherboard players currently still have high inventories of A58 and A78 chipsets and AMD's decision of releasing the A68 is expected to affect their future product plans, the sources said.

from digitimes. "high inventories of a58 and a78 chipsets" is not good news.

 

[juanrga]

I said that I gave that user only a 10% of credibility, but that I was mentioning what he said because a possible cleanup and improvement of x86 was a notice so good that I was giving him the benefit of doubt in this ocassion. However, my post was not understood and my points ignored (which is the norm here anyways).

What we know (99%) is that Intel will use Skylake launch to introduce a bunch of new ISAs, namely AVX512F, AVX512{VL, DQ, BW}, and CDI, and some others more for the companion Xeon Phi processor: ERI and PFI.

I expect AVX2 support for AMD K12/Zen, but no any of the above.

 

[Cazalan]

I think the extra registers are just to accommodate AVX 3.2/AVX-512 but they could be made available for other extensions.

Anyhow as per usual don't expect a significant performance increase without a recompile.

 

[Cazalan]

Informations about cache also came out from nowhere (2 times bigger than now):
http://en.wikipedia.org/w/index.php?title=Skylake_(microarchitecture)&diff=580936777&oldid=578769707

Intel had to double their cache bandwidth to support AVX2 so makes sense they need significant cache improvements for AVX3.2 as well.

 

[szatkus]

Informations about cache also came out from nowhere (2 times bigger than now):
http://en.wikipedia.org/w/index.php?title=Skylake_(microarchitecture)&diff=580936777&oldid=578769707

Intel had to double their cache bandwidth to support AVX2 so makes sense they need significant cache improvements for AVX3.2 as well.

They don't need to, but they could (it could help, but only Intel can check it). AVX512 has no requirements for cache.

You're right with registers. In fact AVX512 needs more and longer SIMD registers. Wiki says about general registers, I'm not sure if their count should be equal to SIMD's count...

 

[palladin9479]

Informations about cache also came out from nowhere (2 times bigger than now):
http://en.wikipedia.org/w/index.php?title=Skylake_(microarchitecture)&diff=580936777&oldid=578769707

Intel had to double their cache bandwidth to support AVX2 so makes sense they need significant cache improvements for AVX3.2 as well.

They don't need to, but they could (it could help, but only Intel can check it). AVX512 has no requirements for cache.

You're right with registers. In fact AVX512 needs more and longer SIMD registers. Wiki says about general registers, I'm not sure if their count should be equal to SIMD's count...

No x86 doesn't work that way. x86 doesn't have SIMD or FPU support, that's a separate coprocessor inside the CPU that has it's own set of separate registers and language. We tend to lump it all together under the "x86 / Intel / AMD" category but it's really important to distinguish between them when we're discussing something as detailed as the register stack. x86 scalar only knows the typical four (AX/BX/CX/DX) and their 32/64 bit names (E/R) along with the instruction and stack related registers.

http://en.wikipedia.org/wiki/X86_assembly_language

The ones for the FPU, and SIMD unit (MMX/SSE/AVX/ect..) are handled completely separately, you can't mix and match. There are special SIMD instructions used to convert data from one of the SSE registers into an Integer register, you can't just use a direct copy like MOV. So Intel adding more general purpose registers for AVX512 would just be an addition to the SIMD "FPU" unit and not to the core AMD64/EMT64 integer unit. The x86 ISA itself would remain unchanged, there would merely be another extension added onto it.

Here is a map of all the registers inside the current implementation of x86 and it's extensions. Yes I ripped it from wiki cause it's easier that way.

Also of note, since this is a superscalar architecture, none of these registers are hard coded as transistors. Instead the internal pseudo-RISC execution engines have a register rename file that allows them to track FAR more registers then possible with x86 ASM. All your compiled register operands get renamed to something completely different inside and then it's returned as part of the execution environment.

 

[juanrga]

I think the extra registers are just to accommodate AVX 3.2/AVX-512 but they could be made available for other extensions.

Anyhow as per usual don't expect a significant performance increase without a recompile.

Yes, AVX512F will increase to 32 the number of vector registers in 64-bit mode. I think he misunderstood this as X64 receiving an update.

Of course, recompile will be a minimum requirement.

 

[juanrga]

RISC announces new Open ISA; hopes to battle ARM/X86.

http://www.extremetech.com/computing/188405-risc-rides-again-new-risc-v-architecture-hopes-to-battle-arm-and-x86-by-being-totally-open-source

I have been asking about this one and the general consensus seems to be that RISC-V ISA is somewhat outdated when compared against a modern ISA such as A64 in technical terms.

Moreover, the VLE and general extensibility mechanism is considered to be a possible source of inefficiencies and incompatibilities. The 128-bit adress space is considered overkill even for decades to come.

There is also criticism about their comparison Rocket vs A5. First because they compared a simplified 'academic' core to a complex 'commercial' core that supports A32, Thumb/2/EE, Jazelle, VFP/NEON, and TrustZone. Thus part of the claimed area advantage comes from not supporting that, and not from a more efficient ISA.

Second, because the comparison used Dhrystone, and this ancient benchmark wouldn't extress the A5 core enough to show its advantages over Rocket core (e.g. 4-way L1d vs Rocket's 2-way), whereas more realistic workloads would show the advantages of the A5 core.

I am also said that RISC-V offers no fundamental advantage over a pure AArch64 implementation, except by avoidign a license fee.

 

[jdwii]

Who would of thought the Wii U is beating the xbox one for 3 months now. Like i said time and time again no need to have 2 GAME consoles that do the same thing when one does it better and cost the same or less. Nintendo is unique-different and even more so creative something in my opinion the other 2 lack.
Once we see a open world zelda game come out and SSB and maybe a price drop these things are going to sell like crazy i might have to make this claim PS4-Wii U-Xbox one when it comes to sales this lifetime. Many want to claim Microsoft can fix this like sony did with the PS3 but the only issue Sony had with the PS3 was price its product was great as well as its games. Microsoft is trying to turn things around by making games exclusive or adding more FPS into the mix but i'm not to sure, waiting for a price drop we all know its going to happen to the xbox one again before the PS4's first price drop. Its also embarrassing that Microsoft never won in sales ever in its existence its almost like Sony's name is better and well liked over Microsoft.
http://wallstcheatsheet.com/technology/xbox-one-vs-ps4-vs-wii-u-sony-and-nintendo-keep-strong.html/?a=viewall
Either way Amd is really loving it and its amazing they where able to hit such a contract when it needed it more than Nvidia did.

 

[gamerk316]

I am also said that RISC-V offers no fundamental advantage over a pure AArch64 implementation, except by avoidign a license fee.

Translation: Its not much worse, but cheaper.

 

[szatkus]

The ones for the FPU, and SIMD unit (MMX/SSE/AVX/ect..) are handled completely separately, you can't mix and match. There are special SIMD instructions used to convert data from one of the SSE registers into an Integer register, you can't just use a direct copy like MOV. So Intel adding more general purpose registers for AVX512 would just be an addition to the SIMD "FPU" unit and not to the core AMD64/EMT64 integer unit. The x86 ISA itself would remain unchanged, there would merely be another extension added onto it.

Because of these few instructions connecting r and mm registers I'm not sure if there is any reason to double the number of general registers. Like you I don't see any reason to do that, but I don't know every single instruction and these things are sometimes surprising.

BTW. AVX isn't just an addition to FPU. Some of these instructions are executed by ALUs.

 

[logainofhades]

You would have to be crazy to claim the I3 is a better choice for gaming in today's world and its easily better for video encoding-rendering-recording gameplay.

It boils down to what the needs of the user are. If you are playing mainly MMO's, Skyrim, and other low threaded games, and i3 actually is the better choice. For example, my 3.5ghz i5 2400 does better in WoW than my FX 8320 @ 4.0ghz. If you are playing more titles like BF4 and Crysis 3, then I could see an FX 8320 making perfect sense, if you never intend to upgrade the CPU in the system. I agree that FX 8320 would be a better choice for gameplay recording as well. I don't look at anything past the FX 8320.

If you live near a Microcenter, then the AMD bundles will probably hold more appeal for those on a tight budget, and rightly so. FX 6300 with GA-78LMT-USB3 for like $110 before rebates and tax. The FX 8320 is only like $20 more. Those are some killer deals. If you are not an overclocker, or just want an FX 8320 @ FX 8350 speeds, that board is good enough for it. I own 2 myself. If you are an overclocker, and need the 8 threads, then the FX 8320 paired with a 970a-ud3p and hyper 212 is a good budget combo. I often will recommend AMD to those on a budget that, live near a microcenter.

 

[sapperastro]

I didn't look at anything past the 8320 either before, until I heard about the soon to be released 8370E. Unless that thing is majorly gimped, having higher clocks than the 8350 out of the box, along with a TDP of 95w is the sort of improvement I would have hoped for a year ago.

 

[gamerk316]

You would have to be crazy to claim the I3 is a better choice for gaming in today's world and its easily better for video encoding-rendering-recording gameplay.

It boils down to what the needs of the user are. If you are playing mainly MMO's, Skyrim, and other low threaded games, and i3 actually is the better choice. For example, my 3.5ghz i5 2400 does better in WoW than my FX 8320 @ 4.0ghz. If you are playing more titles like BF4 and Crysis 3, then I could see an FX 8320 making perfect sense, if you never intend to upgrade the CPU in the system. I agree that FX 8320 would be a better choice for gameplay recording as well. I don't look at anything past the FX 8320.

In terms of FPS, the i3 matches the FX-6300 line, sometimes the FX-8300 line. Its main problem is latency, where those two cores hurt it. DX12 and the other new APIs should help that a lot, making the i3's a lot more attractive.

 

[Reynod]

Actually, building a 9590 system over the weekend with the Seidon120 that comes with it...will see what it does with the paired water cooler...

I chucked together a Seidon120 and an FX6300 with a 5850 ... runs sweet. Stickin an R280X on it shortly. Put a fan speed controller off an old Corsair cooler and a more efficient fan ... works a treat Build cost was peanuts.

 

[8350rocks]

Actually, building a 9590 system over the weekend with the Seidon120 that comes with it...will see what it does with the paired water cooler...

I chucked together a Seidon120 and an FX6300 with a 5850 ... runs sweet. Stickin an R280X on it shortly. Put a fan speed controller off an old Corsair cooler and a more efficient fan ... works a treat Build cost was peanuts.

Yes, it runs well. The 9590 stressed with the water cooler never breaks 45C, and I picked, though I was not aware when I ordered it (did not pay attention to dimensions), a Mid ATX case that runs on the small side, and run 3 case fans in it.

While I had to get creative mounting the radiator, the system itself with 290X and OCZ SSD, Seagate HDD, all runs really well together. Very smooth. I have not bothered to overclock at this point, I really have not needed to do so yet, but I will do that more when I have some time to sit down and tweak it.

As for performance, I ran the headline benchmark yesterday in a bone stock config just to see how it did. It stacks up well to all but the well overclocked i7-2600, and it got the highest single GPU score for the graphics of the bunch...(while, unsurprisingly, coming in lower than an SLI config).

In Thief benchmarks it maxes everything out on Ultra @ 1080p (not surprised), and was only limited by the refresh rate of my monitor with V-sync on.

Tomb Raider was essentially the same result. BF4 on multiplayer was no issue, and mechwarrior online runs like butter.

Have not tried any other games with it yet because I am still having to download and install but will update after I get a chance to do more of that.

EDIT: As a side note...the OCZ Vector 460 120 GB SSD is a fantastic drive. Seriously...if you are looking for a bang for your buck SSD, that would be a good one for you to pick up. I think that bodes well for the R7 SSD AMD is putting out in a partnership. If it is higher perf than the Vertex 460, it will be a great drive too.

 

[Embra]

Sapphire Reveals Tiny R9 285 Graphics Card:
http://www.tomshardware.com/news/sapphire-r9-285-itx-compact,27527.html
Makes a powerful mini gaming box.

 

[blackkstar]

I noticed after browsing some other forums as well as this one, that FX 8000 series has come a long way. I need like 1.6v on my 8350 to be AVX stable, and I'm noticing a lot of people being AVX stable at my clocks at 1.48v or so (AVX stable for me is high 4.8ghz range, but I run 5.1 in Windows). I bought mine at launch and it's week 37 chip, I believe.

It makes me wonder what I could do with a newer chip. Yet at the same time I don't want to sell this one to someone, I've abused it pretty hard.

It's kind of funny in a way how AMD got a 30% clockspeed increase from 8150 to 9590 and a ~10% IPC increase while Intel spends tons of money on R&D to get 5% IPC increases and no clock speed increases each generation. AMD is stretching their R&D money on HEDT pretty hard.

FX 8370 is probably going to review pretty well. There are enough games out there that will use all those cores now, so it should look good in a review as a gaming chip. If someone whips out a review with Skyrim, Shogun 2, Starcraft 2, etc in this day and age they're going to look a bit foolish. Those games are all awfully old and SC2 is pretty dead.

On a side note, I am about to build a K10 24 core monster from used server parts, so thanks to Murphy's Law, expect some sort of big announcement soon about FX!

 

[8350rocks]

Sapphire Reveals Tiny R9 285 Graphics Card:
http://www.tomshardware.com/news/sapphire-r9-285-itx-compact,27527.html
Makes a powerful mini gaming box.

Saw that, makes for one nasty SFF HTPC for gaming.

 

[logainofhades]

Yea, interesting card indeed, if you are in need of a shorter graphics card.

 

[de5_Roy]

one can use the r9 285-itx with an athlon x4 860k with a78/a88x mini itx motherboard and put together a small all-amd gaming pc. the previous amd-option was a pitcairn-pro based card.

however, i am a bit skeptical about how two of these will fare on a microatx z97 motherboard and in a micro atx case.

 

[etayorius]

one can use the r9 285-itx with an athlon x4 860k with a78/a88x mini itx motherboard and put together a small all-amd gaming pc. the previous amd-option was a pitcairn-pro based card.

however, i am a bit skeptical about how two of these will fare on a microatx z97 motherboard and in a micro atx case.

I think the 860k would Bottleneck the 285x by a good margin.

 

[juanrga]

Translation: Its not much worse, but cheaper.

It is acceptable for basic academic use, but outdated for the industry.

No core based in that ISA will be a relevant competitor for AMD's K12/A57 based products.

 

[juanrga]

The tradittional PC market continue shrinking and AMD is migrating away from it. A company so big as AMD cannot live from niche markets such as HEDT. Below are the previsions for evolution of the target market up to 2015

This is an updated CPU/GPU/APU roadmap