TPU's GPU Database Portal & Updates

[TRINITAS]

Hi T4C Fantasy, how do you find FP32 Compute for CPU?

 

[T4C Fantasy]

Hi T4C Fantasy, how do you find FP32 Compute for CPU?

some engineers have spoke to me and this is what they came up with

Bulldozer Calculation
Module * Clock * 16
Module * Clock * 4 (128/32) FP32 ops per FPU * 2 FPUs * 2 FMA

Zen Calculation
Cores * Clock * 24 (realistically 20) register port sharing issues
Cores * Clock * (2 * 2 FMADs + 2 FADDs) * 4 (128/32)

Zen 2 Calculation
Cores * Clock * 48 (realistically 40) register port sharing issues
Cores * Clock * (2 * 2 FMADs + 2 FADDs) * 8 (256/32)

Skylake Calculation
Cores * Clock * 32
Cores * Clock * 8 (256/32) FP32 ops per FPU * 2 FPUs * 2 FMA

 

[londiste]

How are these calculations reached?
- 128 or 256 divided by 32 is wider units divided to FP32 - is this something you can actually do or is it primarily theoretical? x87 is obviously supported so lower precision works but can it work quite this way?
- Is FMAD considered one or two operations?
- Bulldozer has 4 64-bit FMUL and 4 64-bit FADD units per module. With fuse and combining - 2 128-bit FMAD (or 1 256-bit FMAD) per module.
- Zen/Zen2 seem to have FMUL and FADD units in FP (or FMAD with both used together) based on AMD slides which makes the calculation (2 * 2 FMADs + 2FADDs) seem strange. Zen should effectively have Bulldozer's FP Unit per core - 2 128-bit FMAD per core. Zen2 basically doubled the width on that - 2 256-bit FMAD per core.
- Similarly, what exactly is Skylake capable of? Two ports can do 256-bit FMAD, third (port 5) should be able to do FADD but I am not sure how wide.

 

[T4C Fantasy]

How are these calculations reached?
- 128 or 256 divided by 32 is wider units divided to FP32 - is this something you can actually do or is it primarily theoretical? x87 is obviously supported so lower precision works but can it work quite this way?
- Is FMAD considered one or two operations?
- Bulldozer has 4 64-bit FMUL and 4 64-bit FADD units per module.
- Zen/Zen2 seem to have FMUL and FADD units in FP (or FMAD taken together) based on AMD slides which makes the calculation (2 * 2 FMADs + 2FADDs) a bit strange.
- Similarly, what exactly is Skylake capable of? Two ports can do FMAD, third (port 5) should be able to do FADD.

some calculations were mislabeled

but first to your questions
- yes, thats the point of SIMD
- two
- what exactly is the question? Zen has 2 FMADs and 2 FADDs, from the performance guide you can read that you can do 2 FMADs and 1 FADD at absolute peak due to architectural reasons

and
Skylake Calculation
Core * 8 * 2 * 2 * Clock
8 (256/32) FP32 ops per FPU * 2 FPUs * 2 FMA

 

[londiste]

- yes, thats the point of SIMD

Ouch, do I feel stupid now. Brainfart from my side

- what exactly is the question? Zen has 2 FMADs and 2 FADDs, from the performance guide you can read that you can do 2 FMADs and 1 FADD at absolute peak due to architectural reasons

Zen has 2 FMULs and 2 FADDs.
Do you have a link to the performance guide? I can only find pre-Zen stuff on AMD's page.

Skylake Calculation
Core * 8 * 2 * 2 * Clock
8 (256/32) FP32 ops per FPU * 2 FPUs * 2 FMA

Do I get this correctly?
2 FPUs - two FMAD capable units
2 FMA - two instructions per clock due to FMAD

Edit:
I suppose this is what you were referring to? I am curious about how exactly this result comes to be though.

The FMA instructions use the same pipes as multiplication and they also partially occupy the addition units. The measured throughput of mixed FMA and addition instructions is four 128-bit FMA instructions and four 128-bit additions in three clock cycles.

 

[TRINITAS]

Ok thanks

No information on Haswell, Sandy Bridge, and Broadwell?

 

[londiste]

No information on Haswell, Sandy Bridge, and Broadwell?

Haswell/Broadwell should be the same as Skylake for FP32.
Sandy/Ivy Bridge are IIRC half of that - FMUL and FADD units, not FMA.

 

[T4C Fantasy]

Updated list to include console gpus

 

[T4C Fantasy]

GPU Database now has its own discord, come join and maybe help out with the database

Join the TPU GPU Database Discord Server!

Check out the TPU GPU Database community on Discord - hang out with 156 other members and enjoy free voice and text chat.

discord.gg

 

[T4C Fantasy]

If I only could unlock the fps cap...

oh, wrong thread xD

Post your Final Fantasy XV Benchmark Results

Download Final Fantasy XV Benchmark Benchmarks Cinebench 1080P, Standard, Fullscreen Scores are clickable, leading to the original post. NameCPUGPUScore clean.Core i9 13900KRTX 309023786 igralec84Ryzen 5 7600XRTX 409023771 OthnarkRyzen 7 5800X3DRX 6950 XT23766 HDR247Ryzen 9 5950XRX 6900...

www.techpowerup.com

 

[masterdeejay]

The GT1030 GK107 is a fake or a real nvidia approved card?
https://www.techpowerup.com/gpu-specs/geforce-gt-1030-gk107.c3454

Litte off topik: I have a Zotac GTX650 2gb 128bit GDDR5 card that have the same spec and gpu as that GT1030.
So it is possible to mod my card to a 1030GT (GK107)?
I have an spi eeprom (ch341a) programmer, and i can hex edit the bios to change the id. Or where can i get a 1030GT GK107 bios?
I know there will be no performance difference but i think it will extend the driver support for the card.

 

[T4C Fantasy]

The GT1030 GK107 is a fake or a real nvidia approved card?
https://www.techpowerup.com/gpu-specs/geforce-gt-1030-gk107.c3454

Litte off topik: I have a Zotac GTX650 2gb 128bit GDDR5 card that have the same spec and gpu as that GT1030.
So it is possible to mod my card to a 1030GT (GK107)?
I have an spi eeprom (ch341a) programmer, and i can hex edit the bios to change the id. Or where can i get a 1030GT GK107 bios?
I know there will be no performance difference but i think it will extend the driver support for the card.

Modding this i dont think is possible, it is a real card though.

Edit: this wont extend the life either since its still kepler

 

[masterdeejay]

Modding this i dont think is possible, it is a real card though.

Edit: this wont extend the life either since its still kepler

Not possible, ok i just selected gt1030 device driver. Not fake.

 

[T4C Fantasy]

Updated APIs in gpudb, added WDDM Support in Chip database

 

[T4C Fantasy]

Updated some API information in the chip database, added support for and improved ATI/AMD CLRX, Chip variant, Codename, compute support and various corrections.

 

[eidairaman1]

Updated some API information in the chip database, added support for and improved ATI/AMD CLRX, Chip variant, Codename, compute support and various corrections.

Who handles the bios collection?

I've been over the years made a request to change the search option granularity of the Since: category from 1 year, etc to a Specific number so say bios from 2020, select the number 2020 and bios from 2020 only show up. Same with adding Memory chip Brand as a Search function, Example: Elpida, Hynix, Infineon, Micron, Qimonda, Samsung.

Maybe add a Read Me First Link to @Regeneration, @Solaris17, and my bios flash guides?

The advantage here is it would make it easier for all users to find more specific bios files they need and possibly reduce the amount of failed bios flashes.

I've seen newer users get their requests granted so why not I, being a senior contributor?

 

[eidairaman1]

And yet again my request falls on blind eyes

 

[T4C Fantasy]

And yet again my request falls on blind eyes

not ignoring, just nothing i can do about it xD

 

[eidairaman1]

not ignoring, just nothing i can do about it xD

@W1zzard

 

[W1zzard]

Who handles the bios collection?

I do, and I don't have time to work on new features atm

I want to bring the search engine from the SSD DB to the other DBs in a first step ..

This thread is for the GPU Database btw, not the BIOS Collection.

 

[JohH]

Was there a recent change or recalculation of the 'Relative Performance' scores?
I didn't see it in the change log.

 

[T4C Fantasy]

Was there a recent change or recalculation of the 'Relative Performance' scores?
I didn't see it in the change log.

Yes, changed yesterday a little

 

[Count von Schwalbe]

IGPs now show which CPU they are integrated with and have links to CPUDB
View attachment 157878View attachment 157879

1. It seems the (retail) desktop "Renoir" IGPs are not included. They appear to be identical to "Cezanne", but it would be nice to see them as well. Only "PRO" models so far.
2. Is there any plan for a reciprocal link? It would be nice to be able to jump to the IGP from the CPU page.

 

[T4C Fantasy]

Added picture of power cables in main thread

 

[agent_x007]

Needs 12VHP with 4-pin sense pins added, and possible "v2" later on (with shorter sense pins).