oerk wrote:

Can you post a photo of the case please? I had one like these in the mid-90s. The speaker and mic connectors at the back are meant to be connected to a sound card via 3.5mm cables. If it's the same I had, I'd probably be interested in it.

Case: ftp://retronn.de/pictures/bench_results/NexGe … 90/asi_case.jpg

The plastic piece to push the Turbo and Power button is already missing.

That's the one I had! They were used by ASI and Neckermann (or ASI built them for Neckermann, I don't know).

Thought I'd never see one of those again.

How does it bench compared to an Intel DX4-100? What would you consider its performance equivalent socket 3 CPU to be?

Your speedsys score of 62.69 is unusually high. My Intel 486DX4-100 recieved a speedsys score of 39.24.

What resolution were you testing Duke3D in to drop down to 6fps?

Very cool system by the way 😀

Speedsys CPU scores are not accurate. F.e. a 386 without FPU scores significantly higher than a 386 with a FPU installed. The same occurs here: The NexGen has no FPU, hence the unusual high CPU score.
I would use other programs to measure the ALU performance, NSSI for example.

The DOS gaming performance of the NexGen is crippled because of the Vesa Local waitstates. The Vesa Bus clocks at 42MHz, so waitstates apply and restrict the video bandwidth quite a lot. In the consequence, the Benchmarks are slower as one would expect from a "Pentium 90" class CPU. Even my Pentium 66 is faster in every DOS game related department.

FGB wrote:

Speedsys CPU scores are not accurate. F.e. a 386 without FPU scores significantly higher than a 386 with a FPU installed. The same occurs here: The NexGen has no FPU, hence the unusual high CPU score.
I would use other programs to measure the ALU performance, NSSI for example.

About the Waitstates. It is not as critical as thought. Primarily it is important that the graphics card can keep up at all, so e.g. the ET4000 shows that the 32 bit Read Transfer Rate is 2x the 16 bit one, while e.g. the S3 or Trident don't. This means the graphics card can keep up at 32 bit. However the ET4000 does not work stable with 1 WS and at 2 WS it is even with the Avance Logic ALG2228. Mainboard VLB is 0 WS.
So with each additional WS the performance in 3DBench goes down by 2.2 fps (at 65.3 fps max.).
So in this scenario the S3 86C805 is faster at 1 WS in 8 bit transfers than the ET4000 with 2 WS with 8 bit transfers. However if the software uses 32 bit transfers then the S3 can't keep up.
(BTW thats also the main behaviour in phils vga benchmark, as 3DBench2 shows 8 bit speed, while pcplayer bench uses 32 bit transfers).

Numbers: (System is the NexGen at 42 MHz VLB)

ET4000W32i (VESA 1.0)

3DBench2
1 WS 65.3
2 WS 63.1
3 WS 60.8
4 WS 58.7

1 WS:
PCP VGA 17.5
PCP 101 ? (no VESA 1.2)

VIDSPEED 320x200x256
width write read
8 3719 1810
16 7440 3588
32 14881 7034

VIDSPEED 640x480x256 VESA banked
width write read
8 3719 1809
16 7438 3586
32 14882 7023

2 WS:
PCP VGA 17.4

VIDSPEED 320x200x256
width write read
8 3409 1810
16 6818 3587
32 13641 7031

VIDSPEED 640x480x256 VESA banked
width write read
8 3409 1809
16 6818 3584
32 13636 7020

ALG2228

3DBench2
50 MHz 60.8 -> 3 WS
33/40 MHz 63.1 -> 2 WS

2 WS:

PCP VGA 17.4
PCP 101 6.5

VIDSPEED 320x200x256
width write read
8 3409 1554
16 6818 3108
32 13641 6217

VIDSPEED 640x480x256 VESA banked
width write read
8 3409 1524
16 6818 3049
32 13636 6101

S3 86C805

3DBench2
65.6 -> 1 WS?

PCP VGA 16.6
PCP 101 5.9

VIDSPEED 320x200x256
width write read
8 3719 1364
16 7440 2728
32 8614 3614

VIDSPEED 640x480x256 VESA banked
width write read
8 2044 1406
16 4087 2812
32 8174 5624

SPEA VEGA CL-GD5424

60.8 -> 3 WS
PCP VGA 16.5
PCP 101 5.9

VIDSPEED 320x200x256
width write read
8 3147 1521
16 6294 3043
32 8183 4051

VIDSPEED 640x480x256 VESA banked
width write read
8 3146 1509
16 6293 3018
32 8182 4019

Trident TGUI9400CXI

O/C
?/? NI/I ?/+APS ?/FAST CLK ?/?
J1 J8 J7 J10 JX7
1-2 C O O C

3DBench2 52.8
PCP VGA 15.8
PCP 101 6.1

VIDSPEED 320x200x256
width write read
8 2383 1051
16 4768 2103
32 6059 2797

VIDSPEED 640x480x256 VESA banked
width write read
8 2406 1498
16 4812 2997
32 9628 5994

This card works with 1 WS:

Stealth 64 VRAM 86C968

3DBench2
65.6 -> 1 WS

PCP VGA 17.5
PCP 101 6.6

VIDSPEED 320x200x256
width write read
8 3719 1308
16 7440 2616
32 14881 3721

VIDSPEED 640x480x256 VESA banked
width write read
8 3719 1406
16 7438 2068
32 14882 2716

DOOM: 2134 Gametics, 2087 Realtics

Why is there such a large gap in the Vidspeed results between the ET4000W32i and the Stealth 64 VRAM? For example, read at 640x480x256 on ET4000W32i is 7023; on the Stealth64 it is 2716.

From your DOOM benchmak results, you get approx. 36 fps. From my tests using a G200, an Intel DX4-100-WB scores 41 fps, an AMD DX4-100-WT scores 38 fps, a Cyrix DX4-100 (non-5x86) scores 34 fps, and a POD83 scores 44 fps. Assuming your CPU is running at 84 MHz, NexGen results seem OK on a 486-per-clock basis.

feipoa wrote:

Why is there such a large gap in the Vidspeed results between the ET4000W32i and the Stealth 64 VRAM? For example, read at 640x480x256 on ET4000W32i is 7023; on the Stealth64 it is 2716.

From your DOOM benchmak results, you get approx. 36 fps. From my tests using a G200, an Intel DX4-100-WB scores 41 fps, an AMD DX4-100-WT scores 38 fps, and a Cyrix DX4-100 (non-5x86) scores 34 fps. Assuming your CPU is running at 84 MHz, NexGen results seem OK on a per-clock basis.

I don't know where the cause of the read performance gap is. For typical applications it's not such an issue as usually not many application read back from video ram.

feipoa wrote:

Why is there such a large gap in the Vidspeed results between the ET4000W32i and the Stealth 64 VRAM? For example, read at 640x480x256 on ET4000W32i is 7023; on the Stealth64 it is 2716.

From your DOOM benchmak results, you get approx. 36 fps. From my tests using a G200, an Intel DX4-100-WB scores 41 fps, an AMD DX4-100-WT scores 38 fps, and a Cyrix DX4-100 (non-5x86) scores 34 fps. Assuming your CPU is running at 84 MHz, NexGen results seem OK on a per-clock basis.

The NexGen aimed to compete with the Pentium, not the 486. Therefor the scores are all really low. Even my Pentium 66 is faster in every DOS game benchmark despite scoring aprox. 2K less in the NSSI Alu bench.

Compared to the Pentium this CPU was hyped. However the architecture seems to have convinced AMD as option to leave the K5 behind.
However performance is indeed varying a lot. I mean 14 fps in Duke3D, I think my 486DX2-66 shows more.
The speedsys cache anomaly is most likely due to the fact that it puts also the Tag information into the L2 Cache memory, there is no extra Tag RAM.

rgart wrote:

Your speedsys score of 62.69 is unusually high. My Intel 486DX4-100 recieved a speedsys score of 39.24.

Not sure why you would say it's unusually high.
The 586 is a superscalar CPU, capable of executing two instructions per cycle, like a Pentium. The 486 can only execute one instruction per cycle at most.
The 586 is known to be slightly more efficient than the Pentium in some cases. So it is perfectly possible to score much better than a 486 at roughly the same clockspeed. Twice as fast, or even better is a possibility.

Given the relatively low performance in games, it may well be that the CPU is being held back by an underperforming chipset. The cache, memory and ISA/VLB buses may be very slow compared to best-in-class 486 and Pentium boards.

What was the target audience of the NexGen systems? And how was the cost compared to an Intel DX4-100 at the time? Given the low benchmark results, I am trying to establish where the market was with NexGen.

elianda: Are the RAM and cache speeds optimised in the BIOS?

feipoa wrote:

What was the target audience of the NexGen systems? And how was the cost compared to an Intel DX4-100 at the time? Given the low benchmark results, I am trying to establish where the market was with NexGen.

Judging from these reviews, it seems they were marketed as Pentium-class systems:
https://books.google.nl/books?id=aTgEAAAAMBAJ … 0review&f=false
https://books.google.nl/books?id=xDoEAAAAMBAJ … 0review&f=false

feipoa wrote:

elianda: Are the RAM and cache speeds optimised in the BIOS?

It is an AMI WinBIOS. There are very few options and none regarding RAM and Cache. I can set VLB WS, ISA Clock Speed, ISA DMA Cycle Waits, 32 bit IDE transfers....
That are basically all performance relevant options.

The L2 cache is soldered on board, it is 256 kB at half speed in WB mode and holds the Tag information as well. PS/2 SIMMs are 60ns FPM.

I will do some cachechk screens that give at least some latency numbers in microseconds.

Also the book links are noteworthy as well as the stuff I put here: ftp://retronn.de/driver/CPU/nexgen/docs/
specifically ftp://retronn.de/driver/CPU/nexgen/docs/2_Har … rchitecture.pdf
On page 43 the internal architecture is shown with 1 decode and 2 integer units

Also page 48 shows that L1 runs at 2x CPU CLK and L2 at 1/2 CPU CLK. (So the statement in the linked book about full speed cache might be wrong.) L2 chips are 12ns.
At page 49 then it shows that the CPU supports three L2 cache variants, either async. at 1/2 CPU CLK, sync. at 1/2 CPU CLK and sync. at CPU CLK. As the board has async. L2 cache it uses the first (and slowest) option. (so the book is not completely wrong, however it was not implemented on the board)

Comparison to pentium: ftp://retronn.de/docs/Pentium/1000px-Intel_Pe … um_arch.svg.png
It has also two Integer units, however the linked book source states, that the Nexgen can execute two RISC instructions in parallel while the Pentium can execute two x86 instructions in parallel.

For completeness also 486DX2 architecture: ftp://retronn.de/docs/486/1000px-80486DX2_arch.svg.png

So someone has a P60 with async. Cache on a VLB board?

elianda wrote:

It has also two Integer units, however the linked book source states, that the Nexgen can execute two RISC instructions in parallel while the Pentium can execute two x86 instructions in parallel.

It's impossible to compare these without knowing how x86 instructions map to RISC instructions on the NexGen.
Another missing bit of information is how quickly the instructions execute.

A Pentium may be able to execute two x86 instructions in parallel, but the throughput is always the lowest common denominator, because the pipeline is in-order.
Best-case, you can execute two x86-instructions in a single cycle.
But many instructions take more than one cycle. Then there's the added complication that some instructions can only be executed in the U-pipe, so if they accidentally line up for the V-pipe, only one instruction is executed, and the other is delayed until the next round, so it lines up with the U-pipe.

Yes can not be directly compared, it is a different core after all.

I think the Duke3D is a particular bad score with 14 fps. My 486DX2-66 VLB shows 20 fps:

Now about memory speed:
CacheCheck V7 results from the Nexgen:

It shows 9 / 9 / 21 us/kB for Read (L1 / L2 / Main) and 6 / 14 / 32 us/kB (L1, L2, Main) for writing.
With 117.7 MB/s ; 71.2 MB/s ; 53.3 MB/s down to main memory.
One has to be careful with the L2 number here, L1 ends at 16 kB and L2 ends before 256 kB due to reservation for Tag RAM. So the L1 numbers actually refer to L2.

Cachechk V7 on the 486DX2-66 VLB with 2-1-1-1 0WS Cache timing (256 KB async.):

It shows 16 / 24 / 34 us/kB for Read (L1 / L2 / Main) and 24 / 24 / 24 us/kB (L1, L2, Main) for writing due to Write Through.
With 68.9 MB/s ; 46.5 MB/s ; 36.6 MB/s down to main memory.

I don't have a Pentium with Async. Cache here to compare.

Addition: I looked up the CacheCheck scores of my 486DX2-66 on a PCI based board and it allows to tweak the memory timings more. There it goes down to 16 us / kB for writing. This is twice as fast as the main memory time for the NexGen. So it seems to be rather slow for memory writes est. 32 MB/s.

Are these release days approximately correct?

NexGen Nx586 - Sept. 1994
Intel Pentium P54C 75-100 - Oct 1994
Intel DX4-100 WT - Mar 1994
Intel DX4-100 WB - Oct 1994

If so, would it be fairer to compare the Nx586 to the P54C and DX4-100WB?

feipoa wrote:

If so, would it be fairer to compare the Nx586 to the P54C and DX4-100WB?

The article I posted, compared against P54C machines. Apparently at that time the FPU was not yet available, so we can't tell how well it performs with it.
Without it, it didn't stand much of a chance against the P54C.
The most interesting part in the article is that the cheapest P54C models were actually CHEAPER than the NexGen was, because prices had come down quickly as Pentiums were adopted by the masses.
I suppose that is the biggest reason why the NexGen wasn't a big success. It was interesting technology, but they couldn't deliver on price/peformance.
The FPU not being available at launch and no PCI support didn't really help either.

At some stage it did get PCI and later the FPU was integrated into the CPU, though this equipment is even harder to find than the non-FPU containing chips and VLB boards.

I find it amusing how I and others have owned one of these for quite some time, yet you wait until now to ask someone about the chip and motherboard.

I suspect performance will more closely resemble a Socket 4 machine as opposed to a Socket 5, though there will be inconsistencies. My last thoughts are that ECC RAM might skew tests a little and it looks to me that you have such RAM installed in the first two slots. I am unsure if under this configuration it will have any effect but thought I'd point it out, mostly because I was bored.