FX-9590 at 5.5GHz-6GHz

Threadripper maybe based on a server platform for all practical purposes, but it most certainly is not a server part. I won't disagree about the likelyhood of a 5ghz all core part being very low, However as 32 Core parts can hit 4.2Ghz(precision boost 2) on 14nm, I don't see 5ghz being unreasonable for a 7nm product. Especially as this will also be a major revision for ryzen and will likely include other fixes to help it clock better. Its fairly safe to assume their are some critical path issues effecting max frequency on ryzen 1 parts given the odd way almost every chip hits a wall in about the same place(4.1/4.2ghz). And ryzen+ only improves on this slightly, I am guessing those same issues are their, most of the benefit seems to come from the process improvement and cache tweaks.
 
luminousone said:
Threadripper maybe based on a server platform for all practical purposes, but it most certainly is not a server part. I won't disagree about the likelyhood of a 5ghz all core part being very low, However as 32 Core parts can hit 4.2Ghz(precision boost 2) on 14nm, I don't see 5ghz being unreasonable for a 7nm product. Especially as this will also be a major revision for ryzen and will likely include other fixes to help it clock better. Its fairly safe to assume their are some critical path issues effecting max frequency on ryzen 1 parts given the odd way almost every chip hits a wall in about the same place(4.1/4.2ghz). And ryzen+ only improves on this slightly, I am guessing those same issues are their, most of the benefit seems to come from the process improvement and cache tweaks.
Click to expand...

Well the part is better then what you can buy on the desktop. And that is still a server part. If the 7nm process is really good should mean that you get a part that is capable of getting there. If the 5 ghz capable part is a greater percentage of the yield it should be in the Ryzen series.
 
luminousone said:
Threadripper maybe based on a server platform for all practical purposes, but it most certainly is not a server part. I won't disagree about the likelyhood of a 5ghz all core part being very low, However as 32 Core parts can hit 4.2Ghz(precision boost 2) on 14nm, I don't see 5ghz being unreasonable for a 7nm product. Especially as this will also be a major revision for ryzen and will likely include other fixes to help it clock better. Its fairly safe to assume their are some critical path issues effecting max frequency on ryzen 1 parts given the odd way almost every chip hits a wall in about the same place(4.1/4.2ghz). And ryzen+ only improves on this slightly, I am guessing those same issues are their, most of the benefit seems to come from the process improvement and cache tweaks.
Click to expand...

One thing to remember is that whenever a process shrinks higher levels of leakage occur when you push them via voltage out of their operating envelope. Smaller processes can tolerate less and less voltage increases before the leakage becomes dangerously high. That means while stock clocks can increase, overclocking headroom starts decreasing.
 
The way the chip hits a wall is much more indicative of a critical path issue, a wire that is slightly to long, or to many gates in series without a buffer causing circuits to fail if the frequency goes passed a certain point. Ryzen is fairly sippy power wise, and even with extensive cooling to adjust for extra heat etc, the wall doesn't really move much.

Pieter3dnow said:
Well the part is better then what you can buy on the desktop. And that is still a server part. If the 7nm process is really good should mean that you get a part that is capable of getting there. If the 5 ghz capable part is a greater percentage of the yield it should be in the Ryzen series.
Click to expand...

Server part/desktop part is really a mostly moot argument for ryzen, they all use the same die, only real difference is microcode, and socket configuration. And Threadripper uses desktop microcode absent of several feature enablements for the Epyc chips.
 
luminousone said:
Server part/desktop part is really a mostly moot argument for ryzen, they all use the same die, only real difference is microcode, and socket configuration. And Threadripper uses desktop microcode absent of several feature enablements for the Epyc chips.
Click to expand...

But binning is completely different. Better yields and maturing process will allow some of the numbers to change remember the test that was done with Intel chips and the cherry picked cpu the reviewers got that is in the top 5% of best parts. The same goes for the server parts they get the top percentage (don't forget that these can sometimes push a chip further then the worst binned ones with lower voltage and produce less heat).
 
Its like 8 pseudo cores, the biggest issue is the latency of the cache system and the narrow front end that can't feed all the cores at the same time, recurring on interleaving.
 
thebufenator said:
sweet necro.

And if it wasn't a real 8 core, it would scale 4 -> 8 thread performance like a 4c 8t I7 does. But it doesn't. It just has slow cores, and 8 of them.
Click to expand...
it does exactly that when running 64 bit workloads. The FX has 2 shared 128 bit floating points. Each cluster shares one floating point. Each thread has a 32 bit bandwidth. When running a 64 bit application the module uses both threads. That is why you will see in 64-bit applications like Cinebench or Unigine benchmarks the software recognizes it as a 4C/8T processor
 
Last edited:
luminousone said:
No its 8 cores, its just an odd setup, each set of 2 integer cores share either thread of a single SMT FPU. They had some really neat idea's planned for bulldozer but most of them where never finished or implemented. Among these was a reverse hyper-threading/SMT mode that could expand the width of the integer pipeline of each core, by executing instructions for the same thread across 2 cores, this was possible due to the shared decoding, and instruction scheduling hardware. It would have required a deep L0 or micro-op cache, a much expanded cache access width. Pretty much a list of things outside the transistor budget of a 32/28nm process. If they would have fixed the cache, doubled(at a minimum) the fpu width, implemented the reverse smt, micro-op cache, etc, it would have been competitive. Was kinda of a shame that kinda completely abandoned the dozer other then a few minor tweaks for the apu lines.

But it has been implied that the cache system was fundamentally fubar and perhaps not fixable without major redesign of the whole chip.
Click to expand...
It acts as an 8 core in 32 bit applications, I will give you that. But as I said earlier it has two shared 128 Bit fpus, each cluster has one shared 128 bit fpu dedicated to that cluster, one cluster has two modules, each module has two 32 bit threads. If you run a 64 bit workload, the module will use both threads in the module to process it like SMT. That is why the FX handles 64 bit workloads much faster than it does 32 Bit workloads. This is why when using CPU benchmarking software like Cinebench or Unigine, it will recognize the processor as a quad core with eight threads or as it displays 4C/8T. If you noticed on the newer FX boxes they still call them an eight core processor but now on the bottom right corner of the box they say 4M/8T.
 
Overshock81 said:
it does exactly that when running 64 bit workloads. The FX has 2 shared 128 bit floating points. Each cluster shares one floating point. Each thread has a 32 bit bandwidth. When running a 64 bit application the module uses both threads. That is why you will see in 64-bit applications like Cinebench or Unigine benchmarks the software recognizes it as a 4C/8T processor
Click to expand...

Its not a 4 core CPU or it would be as slow as an I7 when it comes to loads that can use 8 threads.

Always the Intel apologists trying to explain an AMD 8 core as no different than an I7 with hyperthreading......

Yes, bulldozer sucked, but it has 8 cores. Yes, they are limited, but it handles on 8 core load a shit ton better than a 4 core i7. Examples being a virtual machine host or crypto mining. Did you know a shitty fx8320e is a good bit faster at XMR mining than an I7? Its becuase is scales past 4 threads, unlike the i7 on the same load.
 
thebufenator said:
Yes, bulldozer sucked, but it has 8 cores. Yes, they are limited, but it handles on 8 core load a shit ton better than a 4 core i7. Examples being a virtual machine host or crypto mining. Did you know a shitty fx8320e is a good bit faster at XMR mining than an I7? Its becuase is scales past 4 threads, unlike the i7 on the same load.
Click to expand...

The only types of loads that Bulldozer is good at is integer loads and heavily threaded loads. Single threaded loads suffer, especially if they're stuffed on a single module with another thread.

Of course the FX8320e would outgun an i7 in XMR, it has eight integer units and the XMR algorithm looks to be primarily integer math. The second you start loading it down with floating point math, it will crash and burn, especially if two threads on the same module are hitting the floating point unit at the same time. This is why both MS in later Windows 7 and Linux in some 2.6.x kernel had to re-write the thread dispatcher to avoid this situation, it will separate resource intensive threads to different modules as long as possible.

You can watch top in a bash console in Linux when you load up multiple heavy CPU time threads, it will distribute them to one per module as long as it can and backfill the second integer unit with low priority threads to avoid perf degradation for single threaded tasks.
 
luminousone said:
Threadripper maybe based on a server platform for all practical purposes, but it most certainly is not a server part. I won't disagree about the likelyhood of a 5ghz all core part being very low, However as 32 Core parts can hit 4.2Ghz(precision boost 2) on 14nm, I don't see 5ghz being unreasonable for a 7nm product. Especially as this will also be a major revision for ryzen and will likely include other fixes to help it clock better. Its fairly safe to assume their are some critical path issues effecting max frequency on ryzen 1 parts given the odd way almost every chip hits a wall in about the same place(4.1/4.2ghz). And ryzen+ only improves on this slightly, I am guessing those same issues are their, most of the benefit seems to come from the process improvement and cache tweaks.
Click to expand...

Are you aware of the near 1kw that a 2990wx pulls when aggressively overclocked?

64 cores at 5ghz ? You would need to hook up a liquid helium coolant system just to keep it chilled lmao

What I want to see is more efficient processor usage where IPC is improved to run at 5ghz performance but can do so at 3ghz. x86 is just an inefficient system. I remember when RISC was huge in 90s and far far far more instruction per clock efficient.
 
tangoseal said:
Are you aware of the near 1kw that a 2990wx pulls when aggressively overclocked?

64 cores at 5ghz ? You would need to hook up a liquid helium coolant system just to keep it chilled lmao

What I want to see is more efficient processor usage where IPC is improved to run at 5ghz performance but can do so at 3ghz. x86 is just an inefficient system. I remember when RISC was huge in 90s and far far far more instruction per clock efficient.
Click to expand...

I got 400 watts out of my Intel 8 core 5960x at 4.8Ghz. Not hard to do if you push the OC to the limit on anything... and you’re in diminishing returns. 4.6Ghz was 250 watts.

One positive of the 2990wx is so large so it’s way easier to cool since the heat density is way lower. As long as the work output makes sense per the watt spent I don’t see a problem.
 
Dayaks said:
I got 400 watts out of my Intel 8 core 5960x at 4.8Ghz. Not hard to do if you push the OC to the limit on anything... and you’re in diminishing returns. 4.6Ghz was 250 watts.

One positive of the 2990wx is so large so it’s way easier to cool since the heat density is way lower. As long as the work output makes sense per the watt spent I don’t see a problem.
Click to expand...

I have a 2950x and at stock clocks on an xspc waterblock I can hover around 65c full load. I cant imagine 64 cores at 5ghz...
 
tangoseal said:
I have a 2950x and at stock clocks on an xspc waterblock I can hover around 65c full load. I cant imagine 64 cores at 5ghz...
Click to expand...

I didn’t say it was easy... but way easier than a single die.
 
tangoseal said:
Are you aware of the near 1kw that a 2990wx pulls when aggressively overclocked?

64 cores at 5ghz ? You would need to hook up a liquid helium coolant system just to keep it chilled lmao

What I want to see is more efficient processor usage where IPC is improved to run at 5ghz performance but can do so at 3ghz. x86 is just an inefficient system. I remember when RISC was huge in 90s and far far far more instruction per clock efficient.
Click to expand...

I never said all core 5 ghz, as long as PB2/PB3!?, can get their on some number of cores that would meet my quote.

RISC more efficient IPC?!? are you delusional?!?, RISC requires more instruction ops todo the same work, RISC abandons many many opportunities for parallelism because of its LOAD/STORE arch rather then having real address modes, RISC abandons common and easy stack optimizations by having no built in hardware stack, RISC bloats the cache by having always large fixed length instructions(4 bytes), RISC bloats the bytecode stream by requiring many instructions to load intermediate values.

RISC being "better" only fits in the 90's where very limited transistor budgets which made working with fixed length instructions easier, and often where cost was less of an issue on expensive large iron machines of the time.

Today, the largest platform is x86 CISC, the most durable platform is IBM Z series which is CISC, the "largest" "RISC" platform is becoming more CISC by the year as ARM adds features like address modes on load/store and a dedicated hardware stack.

SPARC only continues to exist because of existing customers and the amazing BIXBY cache coherency fabric developed at sun that scales to 128+ sockets on a single system image platform(aka not a cluster, but rather a single box with epic ton of processor sockets). PowerPC continues to hit its application space in the high throughput small/medium database space, DB2 performs better on these then intel rigs usually. OpenMIPS, OpenRisc, Alpha, are either dead or relegated to nitch/embedded platforms.

RISC does have a place, its simplicity lends itself well to embedded platforms, and other applications that don't require high speed processing.

Even old desktop platforms knew CISC made more sense, the 6502 being the most common processor of its day, being replaced by the Z80, and then the Motorola 68000.

RISC never enjoyed wide success on any desktop platform save for the short stint of PowerPC based Apple computers, and maybe Acorn computers built under the BBC brand(and this toke the British government flushing an epic amount of money down the tubes to support).
 
luminousone said:
I never said all core 5 ghz, as long as PB2/PB3!?, can get their on some number of cores that would meet my quote.

RISC more efficient IPC?!? are you delusional?!?, RISC requires more instruction ops todo the same work, RISC abandons many many opportunities for parallelism because of its LOAD/STORE arch rather then having real address modes, RISC abandons common and easy stack optimizations by having no built in hardware stack, RISC bloats the cache by having always large fixed length instructions(4 bytes), RISC bloats the bytecode stream by requiring many instructions to load intermediate values.

RISC being "better" only fits in the 90's where very limited transistor budgets which made working with fixed length instructions easier, and often where cost was less of an issue on expensive large iron machines of the time.

Today, the largest platform is x86 CISC, the most durable platform is IBM Z series which is CISC, the "largest" "RISC" platform is becoming more CISC by the year as ARM adds features like address modes on load/store and a dedicated hardware stack.

SPARC only continues to exist because of existing customers and the amazing BIXBY cache coherency fabric developed at sun that scales to 128+ sockets on a single system image platform(aka not a cluster, but rather a single box with epic ton of processor sockets). PowerPC continues to hit its application space in the high throughput small/medium database space, DB2 performs better on these then intel rigs usually. OpenMIPS, OpenRisc, Alpha, are either dead or relegated to nitch/embedded platforms.

RISC does have a place, its simplicity lends itself well to embedded platforms, and other applications that don't require high speed processing.

Even old desktop platforms knew CISC made more sense, the 6502 being the most common processor of its day, being replaced by the Z80, and then the Motorola 68000.

RISC never enjoyed wide success on any desktop platform save for the short stint of PowerPC based Apple computers, and maybe Acorn computers built under the BBC brand(and this toke the British government flushing an epic amount of money down the tubes to support).
Click to expand...

Not delusional sigh.... why ya got to be so harsh.

It's possible I didnt fully understand the tech like you do. I always thought it was supposed to be far more efficient. But your right. It was back in the 90s.
 
luminousone said:
..

Even old desktop platforms knew CISC made more sense, the 6502 being the most common processor of its day, being replaced by the Z80, and then the Motorola 68000.
Click to expand...

I used to love coding assembler for those platforms. Particularly the 68000 (on an Amiga 1000). Things were so much simpler then... *sigh!*
 
For folks who have threaded workloads and who own the 8 core AMD FX chips: These chips are hard to justify replacing.

AMD may have got it wrong for pure gamers. But for gamers who have to do something threaded with their computers they really got it right.

Of course it's no match for Ryzen. On any measure. But it was ahead of it's time.

Computer enthusiasts have been screaming "single thread performance" for a long time. In some ways we are like crusty old men hanging on to single thread benchmarks while the rest of the world is moving on.
 
cyberguyz said:
I used to love coding assembler for those platforms. Particularly the 68000 (on an Amiga 1000). Things were so much simpler then... *sigh!*
Click to expand...

I did assembler on x86(16/32bit), was always nice to see how a few well optimized lines of code could completely destroy the speed of code generated by the compiler. Now days don't get to do any assembler.

Tho, work in threads/concurrency/parallelism can occasionally bring the same sense of satisfaction. But yes, the good ole, bad ole days will never return.
 
thebufenator said:
Its not a 4 core CPU or it would be as slow as an I7 when it comes to loads that can use 8 threads.

Always the Intel apologists trying to explain an AMD 8 core as no different than an I7 with hyperthreading......

Yes, bulldozer sucked, but it has 8 cores. Yes, they are limited, but it handles on 8 core load a shit ton better than a 4 core i7. Examples being a virtual machine host or crypto mining. Did you know a shitty fx8320e is a good bit faster at XMR mining than an I7? Its becuase is scales past 4 threads, unlike the i7 on the same load.
Click to expand...
I have an FX 9590 and I've had it since November 2013. It is a Quad Module/ 8 Thread CPU and this was also settled out of court in 2016 by AMD for falsely advertising it as an 8 core cpu.
 
Overshock81 said:
I have an FX 9590 and I've had it since November 2013. It is a Quad Module/ 8 Thread CPU and this was also settled out of court in 2016 by AMD for falsely advertising it as an 8 core cpu.
Click to expand...

4 module, 8 core cpu. Shared resources does not mean that they are not 8 cores. On a different note, are you still using that 9590 and how well does it run?
 
ManofGod said:
4 module, 8 core cpu. Shared resources does not mean that they are not 8 cores. On a different note, are you still using that 9590 and how well does it run?
Click to expand...
I never had any instability issues like many have had but most of those people didn't buy a board that was compatible the FX 9000 series.
It does very well in gaming and light duty workloads. I don't overclock because in all honesty it's still pretty quick at 4.7. The only gripe I have with it is it's memory performance but that applies to all of the FX series.

specs

Coolermaster HAF 932
Asus Crosshair V Formula-Z
AMD FX 9590 4.7 GHz @1.45V Turbo Disabled
Thermaltake Ultimate 3.0 360mm AIO Liquid Cooler
Gskill Ares DDR3 2400 32 GB kit (4x8GB) 11-13-13-31 2T
XFX R9 Fury X 1123/500
Samsung 840 Pro 128GB SSD
Seagate Barracuda 7200 RPM 1 TB HDD x2 in Raid-0
LG Super Multi BluRay optical drive
Rocketfish 1000W peak/ 900W continuous Power Supply
ePower EP450CD 450W Modular Juice Box (Sole GPU PSU)
 
You must log in or register to reply here.
Back
Top